Pattern Recognition and Image Analysis

Image enhancement is very important for increasing the sensitivity of screening luggage performance at airports. On the basis of 11 statistical measures of image viewability we propose a novel approach to optimizing the choice of image enhancement tools. We propose a neural network predictor that can be used for predicting, on a given test image, the best image enhancement algorithm for it. The network is trained using a number of image examples. The input to the neural network is a set of viewability measures and its output is the choice of enhancement algorithm for that image. On a number of test images we show that such a predictive system is highly capable in forecasting the correct choice of enhancement algorithms (as judged by human experts). We compare our predictive system against a baseline approach that uses a fixed enhancement algorithm for all batch test images, and find the proposed model to be substantially superior.

Image segmentation is one of the most fundamental steps of image analysis. Almost all image segmentation algorithms have their parameters that need to be optimally set for a good segmentation. The problem of automatically setting algorithm parameters on a per image basis has been largely ignored in the vision community. In this paper we present a novel solution to this problem based on classification complexity and image edge analysis.

Contrary to popular belief, despite decades of research in fingerprints, reliable fingerprint recognition is still an open problem. Extracting features out of poor quality prints is the most challenging problem faced in this area. This paper introduces a new approach for fingerprint enhancement based on Short Time Fourier Transform(STFT) Analysis. STFT is a well known technique in signal processing to analyze non-stationary signals. Here we extend its application to 2D fingerprint images.The algorithm simultaneously estimates all the intrinsic properties of the fingerprints such as the foreground region mask, local ridge orientation and local frequency orientation. We have evaluated the algorithm over a set of 800 images from FVC2002 DB3 database and obtained a 17% relative improvement in the recognition rate.

The possibility that a biometric database is compromised is one of the main concerns in implementing biometric identification systems. The compromise of a biometric renders it permanently useless. In this paper we present a method of hashing fingerprint minutia information and performing fingerprint identification in a new space. Only hashed data is transmitted and stored in the server database, and it is not possible to restore fingerprint minutia locations using hashed data. We also present a performance analysis of the proposed algorithm.

This paper presents a digital rights management approach for gray-level images. The ownership of the original image is identified with an ownership statement, which is a gray-level image as well. The proposed scheme utilizes block truncation coding (BTC) to create a master share, which is then used to produce an ownership share against the ownership statement. When in doubt about the property of an image, the author should address his/her ownership share to reveal the ownership statement to claim the ownership. Since our method does not embed the ownership statement into the host image, we can register more than one ownership statements for a single image without destroying the former ownership statements. Besides, the original image does not need to involve in the process of identifying the ownership. Finally, experimental results will show the robustness of our scheme against several common attacks.

The ATRIUM project aims to the automatic detection of threats hidden under clothes using millimetre-wave imaging. We describe a simulator of realistic millimetre-wave images and a system for detecting metallic weapons automatically. The latter employs two stages, detection and tracking. We present a detector for metallic objects based on mixture models, and a target tracker based on particle filtering. We show convincing, simulated millimetre-wave images of the human body with and without hidden threats, including a comparison with real images, and very good detection and tracking performance with eight real sequences. (International Workshop on Pattern Recognition for Crime Prevention, Security and Surveillance)

Many biometrics, such as face, fingerprint and iris images, have been studied extensively for personal verification purposes in the past few decades. However, verification using vein patterns is less developed compared to other human traits. A new personal verification system using the thermal-imaged vein pattern in the back of the hand is presented in the paper. The system consists of five individual steps:

Data Acquisition, Image Enhancement, Vein Pattern Segmentation, Skeletonization and Matching

. Unlike most biometric systems that carry out comparisons based on a pre-selected feature set, this system directly recognizes the shapes of the vein pattern by measuring their Line-Segment Hausdorff Distance. Preliminary testing on a database containing 108 different images has been carried out and all the images are correctly recognized.

This paper presents a robust method for recognizing human faces under varying illuminations. Unlike conventional approaches for recognizing faces in the spatial domain, we model the phase information of face images in the frequency domain and use them as features to represent faces. Then, Support Vector Machines (SVM) are applied to claim an identity using different kernel methods. Due to large variations of the face images, algorithms which perform in the space domain need more training images to achieve reasonable performance. On the other hand, the SVM combined with the phase-only representation of faces performs well even with small number of training images. Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and 3D Linear Subspace (3DLS) are included in the experiment changing the size of images and the number of training images in order to find the best parameters associated with each method. The illumination subset of the CMU-PIE database is used for the performance evaluation.

Within this paper a new data clustering algorithm is proposed based on classical clustering algorithms. Here

k

-means neurons are used as substitute for the original data points. These neurons are online adaptable extending the standard

k

-means clustering algorithm. They are equipped with perceptive fields to identify if a presented data pattern fits within its area it is responsible for.

In order to find clusters within the input data an extension of the

ε

-nearest neighbouring algorithm is used to find connected groups within the set of

k

-means neurons.

Most of the information the clustering algorithm needs are taken directly from the input data. Thus only a small number of parameters have to be adjusted.

The clustering abilities of the presented algorithm are shown using data sets from two different kind of applications.

The reliable identification of human activities in video, for example whether a person is walking, clapping, waving, etc. is extremely important for video interpretations. Since different people could perform the same action across different number of frames, matching two different sequences of the same actions is not a trivial task. In this paper we discuss a new technique for video sequence matching where the matched sequences are of different sizes. The proposed technique is based on frequency domain analysis of feature data. The experiments are shown to achieve high recognition accuracy of 95.4% on recognizing 8 different human actions, and out-perform two baseline methods of comparison.

Pattern Recognition (PR) techniques have proven their ability for detecting malicious activities within network traffic. Systems based on multiple classifiers can further enforce detection capabilities by combining and correlating the results obtained by different sources.

An aspect often disregarded in PR approaches dealing with the intrusion detection problem is the use of temporal information. Indeed, an attack is typically carried out along a set of consecutive network packets; therefore, a PR system could improve its reliability by examining sequences of network connections before expressing a decision.

In this paper we present a system that uses a multiple classifier approach together with temporal information about the network packets to be classified. In order to improve classification reliability, we introduce the concept of rejection: instead of emitting an unreliable verdict, an ambiguously classified packet can be logged for further analysis.

The proposed system has been tested on a wide database made up of real network traffic traces.

This paper presents a view independent video-based face recognition method using posterior probability in Kernel Fisher Discriminant (KFD) space. In practical environment, the view of faces changes dynamically. The robustness to view changes is required for video-based face recognition in practical environment. Since the view changes induces large non-linear variation, kernel-based methods are appropriate. We use KFD analysis to cope with non-linear variation. To classify image sequence, the posterior probability in KFD space is used. KFD analysis assumes that the distribution of each class in high dimensional space is Gaussian. This makes the computation of posterior probability in KFD space easy. The effectiveness of the proposed method is shown by the comparison with the other feature spaces and classification methods.

Symmetry is a fundamental structure that is found to some extent in all images. It is thought to be an important factor in the human visual system for obtaining understanding and extracting semantics from visual material. This paper describes a method of detecting axes of reflective symmetry in faces that does not require prior assumptions about the image being analysed. The approach is derived from earlier work on visual attention that identifies salient regions and translational symmetries.

In this paper, an efficient iris segmentation method for recognition is described. The method is based on crossed chord theorem and zigzag collarette area. We select the zigzag collarette region as personal identification pattern, which can remove unnecessary areas and get good recognition rate. Zigzag collarette area is one of the most important parts of iris complex pattern. It is insensitive to the pupil dilation and not affected by the eyelid or eyelash since it is closed with the pupil. In our algorithm, we could avoid procedure for eyelid detection and searching the radius and the center position of the outer boundary between the iris and the sclera, which is difficult to locate when there is little contrast between iris and sclera regions. The method was implemented and tested using two iris database sets, i.e CASIA and SJTU-IDB, with different contrast quality. The experimental results show that the performance of the proposed method is encouraging and comparable to the traditional method.

This paper describes a novel method for palmprint recognition based on registration information and 2D Gabor features. After preprocessing, a unified coordinate system is constructed for each palmprint image and used to guide ROI extraction. A multi-scale matching strategy is employed to match registration information and 2D Gabor features. In the first two levels, registration information is extracted and used to measure the global similarity between two palmprint patterns. In the third level, two palmprints are aligned with their registration information and then are matched using their corresponding Gabor features. The experimental results demonstrate the effectiveness of the method.

In this paper we analyse the effects that JPEG and JPEG2000 compression have on subspace appearance-based face recognition algorithms. This is the first comprehensive study of standard JPEG2000 compression effects on face recognition, as well as an extension of existing experiments for JPEG compression. A wide range of bitrates (compression ratios) was used on probe images and results are reported for 12 different subspace face recognition algorithms. Effects of image compression on recognition performance are of interest in applications where image storage space and image transmission time are of critical importance. It will be shown that not only that compression does not deteriorate performance but it, in some cases, even improves it slightly. Some unexpected effects will be presented (like the ability of JPEG2000 to capture the information essential for recognizing changes caused by images taken later in time) and lines of further research suggested.

In this paper we present a technique for predicting the 2D human body joints and limbs position in monocular image sequences, and reconstructing its corresponding 3D postures using information provided by a 3D action model. This method is used in a framework based on particle filtering, for the automatic tracking and reconstruction of the 3D human body postures. A set of the reconstructed postures up to time

t

are projected on the action space defined in this work, which is learnt from Motion Capture data, and provides us a principled way to establish similarity between body postures, natural occlusion handling, invariance to viewpoint, robustness, and is able to handle different people and different speeds while performing an action. Results on manually selected joint positions on real image sequences are shown in order to prove the correctness of this approach.

We present a technique for dimension reduction. The technique uses a gradient descent approach to attempt to sequentially find orthogonal vectors such that when the data is projected onto each vector the classification error is minimised. We make no assumptions about the structure of the data and the technique is independent of the classifier model used. Our approach has advantages over other dimensionality reduction techniques, such as Linear Discriminant Analysis (LDA), which assumes unimodal gaussian distributions, and Principal Component Analysis (PCA) which is ignorant of class labels. In this paper we present the results of a comparison of our technique with PCA and LDA when applied to various 2-dimensional distributions and the two class cancer diagnosis task from theWisconsin Diagnostic Breast Cancer Database, which contains 30 features.

Retinal vessel extraction has become an important task of medical image processing applications in order to diagnose ocular diseases. In this paper, a novel methodology is proposed to extract vessels automatically from retinal angiographies. The proposed methodology has been implemented by means of Cellular Neural Networks techniques to take advantage of their capabilities of massively parallel processing reducing computation time required.

It has been established that information distinguishing one human face from another is contained to a large extent in the Fourier domain phase component of the facial image. However, to date, formal statistical models for this component have not been deployed in face recognition tasks. In this paper we introduce a model-based approach using Gaussian mixture models (GMM) for the phase component for performing human identification. Classification and verification are performed using a MAP estimate and we show that we are able to achieve identification error rates as low as 2% and verification error rates as low as 0.3% on a database with 65 individuals with extreme illumination variations. The proposed method is easily able to deal with other distortions such as expressions and poses, and hence this establishes its robustness to intra-personal variations. A potential use of the method in illumination normalization is also discussed.

A novel and efficient approach to facial expression classification based on the belief theory and data fusion is presented and discussed. The considered expressions correspond to three (

joy, surprise, disgust

) of the six universal emotions as well as the

neutral

expression. A robust contour segmentation technique is used to generate an expression skeleton with facial permanent features (mouth, eyes and eyebrows). This skeleton is used to determine the facial features deformations occurring when an expression is present on the face defining a set of characteristic distances. In order to be able to recognize “pure” as well as “mixtures” of facial expressions, a belief-theory based fusion process is proposed. The performances and the limits of the proposed recognition method are highlighted thanks to the analysis of a great number of results on three different test databases: the Hammal-Caplier database, the Cohn-Kanade database and the Cottrel database. Preliminary results demonstrate the interest of the proposed approach, as well as its ability to recognize non separable facial expressions.

In this paper, we propose an uncorrelated, weighted LDA (UWLDA) technique for face recognition. The UWLDA extends the uncorrelated LDA (ULDA) technique by integrating the weighted pairwise Fisher criterion and nullspace LDA (NLDA), while retaining all merits of ULDA. Experiments compare the proposed algorithm to other face recognition methods that employ linear dimensionality reduction such as Eigenfaces, Fisherfaces, DLDA and NLDA on the AR face database. The results demonstrate the efficiency and superiority of our method.

In this paper, we propose a novel heteroscedastic weighted kernel discriminant analysis (HW-KDA) method that extends the linear discriminant analysis (LDA) to deal explicitly with heteroscedasticity and nonlinearity of the face pattern’s distribution by integrating the weighted pairwise Chernoff criterion and Kernel trick. The proposed algorithm has been tested, in terms of classification rate performance, on the multiview UMIST face database. Results indicate that the HW-KDA methodology is able to achieve excellent performance with only a very small set of features and outperforms other two popular kernel face recognition methods, the kernel PCA (KPCA) and generalized discriminant analysis (GDA).

In this paper, a supervised feature extraction method having both non-negative bases and weights is proposed. The idea is to extend the

Non-negative Matrix Factorization

(NMF) algorithm in order to extract features that enforce not only the spatial locality, but also the separability between classes in a discriminant manner. The proposed method incorporates discriminant constraints inside the NMF decomposition in a class specific manner. Thus, a decomposition of a face to its discriminant parts is obtained and new update rules for both the weights and the basis images are derived. The introduced methods have been applied to the problem of frontal face verification using the well known XM2VTS database. The proposed algorithm greatly enhance the performance of NMF for frontal face verification.

For databases of facial images, where each subject has only a few images, the query precision of interactive retrieval suffers from the problem of extremely small class sizes. A novel method is proposed to relieve this problem by applying partial relevance to the interactive retrieval. This work extends an existing content-based image retrieval system, PicSOM, by relaxing the relevance criterion in the early rounds of the retrieval. Moreover, we apply linear discriminant analysis as a preprocessing step before training the Self-Organizing Maps (SOMs) so that the resulting SOMs have stronger discriminative power. The results of simulated retrieval experiments suggest that for semantic classes such as “black persons” or “bearded persons” the first image which depicts the target subject can be obtained three to six times faster than by retrieval without the partial relevance.

This paper presents a new approach to efficient and effective personal identification for the security of network access by combining techniques in biometrics and mobile computing. To overcome the limitations of the existing password-based authentication services on the Internet, we propose a dynamic feature selection scheme to extract multiple personal features and integrate them in a hierarchical structure for fast and reliable identity authentication. To increase the speed and flexibility of the process, we use mobile agents as a navigational tool for parallel implementation in a distributed environment, which includes hierarchical biometric feature extraction, multiple feature integration, dynamic biometric data indexing and guided search.

An efficient algorithm to iris segmentation and its application to automatic and non-intrusive gaze tracking and vigilance estimation is presented and discussed. A luminance gradient technique is used to fit the irises from face images. A robust preprocessing which mimics the human retina is used in such a way that a robust system to luminance variations is obtained and contrast enhancement is achieved. The validation of the proposed algorithm is experimentally demonstrated by using three well-known test databases: the FERET database, the Yale database and the Cohn-Kanade database. Experimental results confirm the effectiveness and the robustness of the proposed approach to be applied successfully in gaze direction and vigilance estimation.

In this paper, we present a new statistical projection-based face recognition method, called Bilinear Discriminant Analysis (BDA). The proposed technique effectively combines two complementary versions of Two-Dimensional-Oriented Linear Discriminant Analysis (2DoLDA), namely Column-Oriented Linear Discriminant Analysis (CoLDA) and Row-Oriented Linear Discriminant Analysis (RoLDA). BDA relies on the maximization of a generalized bilinear projection-based Fisher criterion. A series of experiments was performed on various international face image databases in order to evaluate and compare the effectiveness of BDA to RoLDA and CoLDA. The experimental results indicate that BDA outperforms RoLDA, CoLDA and 2DPCA for face recognition, while leading to a significant dimensionality reduction.

Adaptation to dynamically changing environment is very important since advanced applications become pervasive and ubiquitous. This paper addresses a novel method of adaptive object recognition using environmental context-awareness and genetic algorithm and t-test. The proposed method tries to distinguish the category of input environment and decides an optimal classifier combination structure accordingly by GA and t-test. It stores its experiences in terms of the data context categories and the evolved artificial chromosomes so that the evolutionary knowledge can be used later. The proposed method has been evaluated in the area of face recognition. Most previous face recognition schemes define their system structures at the design phases, and the structures are not adaptive during operation. Such approaches usually show vulnerability under varying illumination environment. The context-awareness, modeling and identification of input data as context categories, is carried out by Fuzzy ART. The face data context is described based on the image attributes of light direction and brightness. The superiority of the proposed system is shown using four data sets: Inha, FERET and Yale database.

In this paper, the framework and implementation of a real time multi-scale face detection system using appearance-based learning method and multi-pose hybrid learning approach. Multiple scale and pose based object detection is attractive since it could accumulate the face models by autonomous learning process. Face image, however, can be approximated even though it is represented with many scales. A real time face detection determines the location and size of each human face(if any) in an input image. Detecting varying human face in video frames is an important task in many computer vision applications such as human-computer interface. The face detection proposed in this paper employs hybrid learning approach and statistical method. We employ FuzzyART and RBF Network and Mahalanobis distance. We achieve a very encouraging experimental results.

Statistical dependence of classifiers has recently been shown to improve accuracy over statistically independent classifiers. In this paper, we focus on the verification application and theoretically analyze the AND fusion rule to find the favorable conditional dependence that improves the fusion accuracy over conditionally independent classifiers. Based on this analysis, we come with a method to design such classifiers by training the classifiers on different partitions of the training data. The AR face database is used for performance evaluation and the proposed method has a false rejection rate (FRR) of 2.4% and a false acceptance rate of 3.3% on AND fusion, which is better than an FRR of 3.8% and FAR of 4.3% when classifiers are designed without taking account the AND fusion rule.

In this paper, we propose a method to evaluate the possible recognition degree of a face, called face recognizability, before face recognition. If we can measure the recognizability, we can increase the system efficiency by avoiding recognizing the faces with poor recognizabilities. Based on the features of the orientation distribution on the face regions, we found the facial components. Then we collected lines on the face with major orientations. Last, we used the triangle formed by two eyes and mouth, the degree of the face shape symmetry and intensity symmetry to define the measurement of face recognizability. Experimental results show that recognizability can be used as a measurement to determine whether we need to perform face recognition or not.

With the advancement of mobile technology, mobile phones can store significant amount of sensitive and private information. The security issue of mobile phones becomes an important field to investigate. This paper proposes a prototype of fingerprint authentication mobile phone based on sweep sensor MBF310. The prototype is composed of the front-end fingerprint capture sub-system and the back-end fingerprint recognition system. A sweep fingerprint sensor MBF310 is used to fit the request of the mobile phone in the field of the size, cost, and power consumption. The performance of the proposed prototype is evaluated on the database built by the sweep fingerprint sensor. The EER is 4.23%, and the average match time of the prototype is about 4.5 seconds.

This paper presents a robust and efficient eye detection algorithm for gray intensity images. The idea of our method is to combine the respective advantages of two existing techniques, feature based method and template based method, and to overcome their shortcomings. Firstly, after the location of face region is detected, a feature based method will be used to detect two rough regions of both eyes on the face. Then an accurate detection of iris centers will be continued by applying a template based method in these two rough regions. Results of experiments to the faces without spectacles show that the proposed approach is not only robust but also quite efficient.

Gait has received substantial attention from researchers. Different from other biometrics, gait can be captured in a distance and it is difficult to disguise. In this paper, we propose a feature template: Silhouette Spatio-temporal (SStS). It generates by concatenating silhouette projection vectors (SPV) which is formulated by projection of silhouette in vertical direction. We applied the Principle Component Analysis (PCA) for dimension reduction of the input feature space for recognition. The proposed algorithm has a promising performance, the identification rate is 95% in SOTON dataset and 90% CASIA dataset. Experiments showed that SStS has a high discriminative power and it is suitable for real-time gait recognition system.

To detect a human body and recognize its posture, a component-based approach is less susceptible to changes in posture and lighting conditions. This paper proposes a component-based human-body model that comprises ten components and their flexible links. Each component contains geometrical information, appearance information, and information on the links with other components. The proposed method in this paper uses hierarchical links between components of human body, so that it allows to make coarse-to-fine searches and makes human-body matching more time-efficient. To adaptively estimate the posture in change of posture and illumination, we update the component online every time a new human body is incoming.

We proposed a novel approach for face recognition to address the challenging task of recognition using a fusion of nonlinear dimensional reduction; Locally Linear Embedding (LLE) and Principal Component Analysis (PCA) .LLE computes a compact representation of high dimensional data combining the major advantages of linear methods, With the advantages of non-linear approaches which is flexible to learn a broad of class on nonlinear manifolds. The application of LLE, however, is limited due to its lack of a parametric mapping between the observation and the low-dimensional output. In addition, the revealed underlying manifold can only be observed subjectively. To overcome these limitations, we propose our method for recognition by fusion of LLE and Principal Component Analysis (FLLEPCA) and validate their efficiency. Experiments on CMU AMP Face EXpression Database and JAFFE databases show the advantages of our proposed novel approach.

This paper presents novel features for face detection in the paradigm of AdaBoost algorithm. Features are multi-dimensional histograms computed from a set of rectangles in the filtered images, and they represent marginal distributions of these rectangles. The filter banks consist of intensity, Laplacian of Gaussian (Difference of Gaussians), and Gabor filters, aiming at capturing spatial and frequency properties of human faces at different scales and different orientations. The best features selected by AdaBoost, pairs of filter and rectangle, can thus be interpreted as boosted marginal distributions of human faces. The result of preliminary experiments demonstrate that the selected features are much more powerful to describe the face pattern than the simple features of Viola and Jones and some variants which can only capture several moments of ONE dimensional histogram in intensity images.

The feature selection is very important for improving classifier’s accuracy and reducing classifier’s running time. In this paper, a novel feature selection method based on KPCA, SVM and GSFS is proposed for face recognition. The proposed method can be described as follows, first KPCA is used for extracting initial face features, secondly, the extracted features are divided into some single feature sets, and then the single feature sets are trained separately by SVM to obtain the best feature set through GSFS. In this way, the dimensionality of the initial features can be reduced and also the best features can be obtained. Experimental results on ORL, IITL and UMIST face databases indicate the effectiveness of the proposed method.

In this paper we propose a novel method for performing pose-tolerant face recognition. We propose to use Fourier Magnitude Spectra of face images as signatures and then perform principal component analysis (PCA) and Fisher-faces (LDA) leading to new representations that we call Eigen and Fisher-Fourier Magnitudes. We show that performing PCA and Fisherfaces on the Fourier magnitude spectra provides significant improvement over traditional PCA and Fisherfaces on original spatial-domain image data. Furthermore, we show analytically and experimentally that our proposed approach is shift-invariant, i.e., we obtain the same Fourier-Magnitude Spectra regardless of the shift of the input image. We report recognition results on the ORL face database showing the significant improvement of our method under many different experimental configurations including the presence of noise.

In this article we propose simple descriptor for the purposes of 3D objects recognition and classification. Princeton Shape Benchmark 2004 is used for testing the proposed descriptor. Small size (512b) of the proposed descriptor and short generation and comparison times combine with relatively high recognition abilities. Surprisingly, we found that despite its simplicity and the small size the proposed descriptor took the first place in “coarser” classification test, where all 3D models were divided into 6 large classes: buildings, household, plants, animals, furniture, vehicles and a miscellaneous class not included in averaged retrieval results.

In this paper, we present a new application of the mathematical morphology: a single-image approach for the automatic detection and elimination of highlights in colour images. We use a 2D-histogram that allows us to relate the achromatic and saturation signals of a colour image and to identify interior brightness. To eliminate the highlights detected, we use an image-inpainting method, by means of connected vectorial filters of the mathematical morphology. This new filter operates exclusively on bright zones, reducing the high cost of processing the connected filtersand avoiding over-simplification. The new method proposed here achieves good results, which are similar to those obtained from other multimedia techniques, yet does not require either costly multiple-view systems or stereo images.

This paper evaluates whether shape features can be used for clustering objects in Sidec

TM

Electron Tomography (SET) reconstructions. SET reconstructions contain a large number of objects, and only a few of them are of interest. It is desired to limit the analysis to contain as few uninteresting objects as possible. Unsupervised hierarchical clustering is used to group objects into classes. Experiments are done on one synthetic data set and two data sets from a SET reconstruction of a human growth hormone (1hwg) in solution. The experiments indicate that clustering of objects in SET reconstructions based on shape features is useful for finding structural classes.

Keeping track of a target by successive detections may not be feasible, whereas it can be accomplished by using tracking techniques. Tracking can be addressed by means of particle filtering. We have developed a new algorithm which aims to deal with some particle-filter related problems while coping with expected difficulties. In this paper, we present a novel approach to handling complete occlusions. We focus also on the target-model update conditions, ensuring proper tracking. The proposal has been successfully tested in sequences involving multiple targets, whose dynamics are highly non-linear, moving over clutter.

As an orthogonal moment, Zernike moment (ZM) is an attractive image feature in a number of application scenarios due to its distinguishing properties. However, we find that for digital images, the commonly used Cartesian method for ZM computation has compromised the advantages of ZMs because of their non-ideal accuracy stemming from two inherent sources of errors, i.e., the geometric error and the integral error. There exists considerable errors in image reconstruction using ZMs calculated with the Cartesian method. In this paper, we propose a polar coordinate based algorithm for the computation of ZMs, which avoids the two kinds of errors and greatly improves the accuracy of ZM computation. We present solutions to the key issues in ZM computation under polar coordinate system, including the derivation of computation formulas, the polar pixel arrangement scheme, and the interpolation-based image conversion etc. As a result, ZM-based image reconstruction can be performed much more accurately.

We present a new approach to detecting defects in random textures which requires only very few defect free samples for unsupervised training. Each product image is divided into overlapping patches of various sizes. Then, density mixture models are applied to reduce groupings of patches to a number of textural exemplars, referred to here as texems, characterising the means and covariances of whole sets of image patches. The texems can be viewed as implicit representations of textural primitives. A multiscale approach is used to save computational costs. Finally, we perform novelty detection by applying the lower bound of normal samples likelihoods on the multiscale defect map of an image to localise defects.

In this paper, we propose a novel method for cross-media semantic-based information retrieval, which combines classical text- based and content-based image retrieval techniques. This semantic-based approach aims at determining the strong relationships between keywords (in the caption) and types of visual features associated with its typical images. These relationships are then used to retrieve images from a textual query. In particular, the association

keyword/visual feature

may allow us to retrieve non-annotated but similar images to those retrieved by a classical textual query. It can also be used for automatic images annotation. Our experiments on two different databases show that this approach is promising for cross-media retrieval.

This paper presents an effective texture descriptor invariant to translation, scaling, and rotation for texture-based image retrieval applications. The proposed texture descriptor is built taking the Fourier space of the image. In order to find the best texture descriptor, a quantization scheme based on Lloyd’s technique is proposed. As frequency descriptors are not invariant to all geometrical transformations as scaling and rotation, the modal analysis is applied to overcome these problems. Our image database is extracted from Brodatz album as well other sources. The proposed method is also compared with other content-based techniques and their performance is evaluated through several experiments. The effectiveness of both methods is measured by the commonly used retrieval performance measurement – Precision and Recall.

In this paper, we propose a method for reconstructing the surfaces of objects from stereo data. The proposed method quantitatively defines not only the fitness of the stereo data to surfaces but also the connectivity and smoothness of the surfaces in the framework of a three-dimensional (3-D) Markov Random Field (MRF) model. The surface reconstruction is accomplished by searching for the most possible MRF’s state. Experimental results are shown for artificial and actual stereo data.

This contribution shows how unsupervised Markovian segmentation techniques can be accelerated when implemented on graphics hardware equipped with a Graphics Processing Unit (GPU). Our strategy exploits the intrinsic properties of local interactions between sites of a Markov Random Field model with the parallel computation ability of a GPU. This paper explains how classical iterative site-wise-update algorithms commonly used to optimize global Markovian cost functions can be efficiently implemented in parallel by

fragment shaders

driven by a

fragment processor

. This parallel programming strategy significantly accelerates optimization algorithms such as ICM and simulated annealing. Good acceleration are also achieved for parameter estimation procedures such as

K

-means and ICE. The experiments reported in this paper have been obtained with a mid-end, affordable graphics card available on the market.

Many applications such as image compression, pre-processing or segmentation require some information from the regions composing an image. The main objective of this paper is to define a methodology to extract some local information from an image. Each region is characterized in terms of homogeneity (region composed with the same grey-level or a single texture) and its type (textured or uniform). The decision criterion is based on the use of classical texture attributes (cooccurrence matrix and grey-levels moments) and a support vector machine in order to realize the fusion of the different attributes. We then characterize each region considering its type by appropriate features.

The goal of this work is to propose a criterion for the evaluation of ultrasound image compression. We want to measure the image quality as easily as with a statistical criterion, and with the same reliability as the medical assessment. An initial psychovisual experiment is proposed to medical experts, and represents our reference value for the comparison of the evaluation criteria. Several statistical criteria are selected from the literature. We define a cumulative absolute similarity measure as a distance between the criterion to evaluate and the reference value. A fusion method by a genetic algorithm is proposed to improve the results obtained by each criterion separately. We show the benefit of fusion through some experimental results.

In this paper, we propose a method for 3D model indexing based on 2D views, named AVC (Adaptive Views Clustering). The goal of this method is to provide an optimal selection of 2D views from a 3D model, and a probabilistic Bayesian method for 3D model retrieval from these views. The characteristic views selection algorithm is based on an adaptive clustering algorithm and using statistical model distribution scores to select the optimal number of views. Starting from the fact that all views do not contain the same amount of information, we also introduce a novel Bayesian approach to improve the retrieval. We finally present our results and compare our method to some state of the art 3D retrieval descriptors on the

Princeton 3D Shape Benchmark

database.

A fast and robust type of unsupervised multispectral texture segmentation method with unknown number of classes is presented. Single decorrelated monospectral texture factors are represented by four local autoregressive random field models recursively evaluated for each pixel and for each spectral band. The segmentation algorithm is based on the underlying Gaussian mixture model and starts with an over segmented initial estimation which is adaptively modified until the optimal number of homogeneous texture segments is reached. The performance of the presented method is extensively tested on the Prague segmentation benchmark using nineteen most frequented segmentation criteria.

A new object-based image retrieval (OBIR) scheme is introduced. The images are analyzed using the recently developed, human-based 11 colors quantization scheme and the color correlogram. Their output served as input for the image segmentation algorithm: agglomerative merging, which is extended to color images. From the resulting coarse segments, boundaries are extracted by pixelwise classification, which are smoothed by erosion and dilation operators. The resulting features of the extracted shapes, completed the data for a <color, texture, shape>-vector. Combined with the intersection distance measure, this vector is used for OBIR, as are its components. Although shape matching by itself provides good results, the complete vector outperforms its components, with up to 80% precision. Hence, a unique, excellently performing, fast, on human perception based, OBIR scheme is achieved.

Applying the divergence operator on the gradient vector field is known as a robust method for computing the local creaseness, defined as the level set extrinsic curvature. Based on this measure, we present a multi-scale method to extract continuous midlines of elongated objects of various widths simultaneously. The scale-space is not built on the input image, but on the gradient vector field. During the iterative construction of the scale-space the current solution keeps thin objects even when they are located near more dominant structures. The representation of the midlines is realised as curves in the image plane, consisting of equidistant sample points. At each sample point the tangential direction of the curve is computed directly with the smoothed gradient vector field.

In this paper we discuss about the applicability of text classification techniques for automatic content recognition of the scenes from news videos. In particular, the news scenes are classified according to a predefined set of six categories (National Politics, National News, World, Finance, Society & Culture and Sports) by applying text classification techniques on the transcription of the anchorman speech. The transcription is obtained using a commercial tool for speech to text. The application of text classification techniques for the automatic indexing of news videos is not new in the scientific literature, but, to the best of our knowledge, no paper reports a detailed experimentation. In our experimentations we considered different issues concerning the application of text categorization and speech recognition for news story classification: in fact, we calculated the overall performance obtained by using text categorization on the ideal transcription, as it could be obtained by employing a perfect speech recognition engine, and the transcription provided by a commercial speech recognition tool; furthermore, in our experimentation we were also interested to characterize the performance in terms of the portion of the news story by which the transcription is obtained. The experimentations have been carried out on a database of Italian news videos. This experimental validation represents the main contribution of this paper.

The aim of this paper is to present an improvement of a previously published algorithm. The proposed approach is performed in two steps. In the first step, we generate the Weighted Adaptive Neighborhood Hypergraph (WAINH) of the given gray-scale image. In the second step, we partition the WAINH using a multilevel hypergraph partitioning technique. To evaluate the algorithm performances, experiments were carried out on medical and natural images. The results show that the proposed segmentation approach is more accurate than the graph based segmentation algorithm using normalized cut criteria.

Color induced texture analysis is explored, using two texture analysis techniques: the co-occurrence matrix and the color correlogram as well as color histograms. Several quantization schemes for six color spaces and the human-based 11 color quantization scheme have been applied. The VisTex texture database was used as test bed. A new color induced texture analysis approach is introduced: the parallel-sequential approach; i.e., the color correlogram combined with the color histogram. This new approach was found to be highly successful (up to 96% correct classification). Moreover, the 11 color quantization scheme performed excellent (94% correct classification) and should, therefore, be incorporated for real-time image analysis. In general, the results emphasize the importance of the use of color for texture analysis and of color as global image feature. Moreover, it illustrates the complementary character of both features.

As content-based multimedia applications become increasingly important, demand for technologies on semantic video object segmentation is growing, where the segmented objects are expected to be in line with human visual perception. Existing research is limited to semi-automatic approach, in which human intervene is often required. These include manual selection of seeds for region growing or manual classification of background edges etc. In this paper, we propose an automatic region growing algorithm for video object segmentation, which features in automatic selection of seeds and thus the entire segmentation does not require any action from human users. Experimental results show that the proposed algorithm performs well in terms of the effectiveness in video object segmentation.

This paper presents an object coding scheme based on varying Bezier polynomials between cubics, quadratics and linears. Extracted data points, without any other overhead, are the end product of this scheme which form set of Bezier control points. Corner detection as a preprocessing phase simplifies subsequent coding operation and properties of Bezier splines are exploited to extract final data points. The proposed method results in high data reduction without any compromise to the quality of reconstructed shapes. The coding scheme is suitable for real time image processing applications due to its high compression ratio, efficient and accurate representation of given shapes.

The

Logarithmic Image Processing

model (LIP) is a robust mathematical framework for the processing of transmitted and reflected images. It follows many visual, physical and psychophysical laws. This works presents a new formulation of a 2D–convolution of separable kernels using the LIP paradigm. A previously stated LIP–Sobel edge detector is redefined with the new proposed formulation, and the performance of the edge detectors programmed following the two formulations (the previous one and the new one proposed) is compared. Another operator, Laplacian of Gaussian, is also stated under the LIP paradigm. The experiments show that both methods obtain same results although our proposed method is much faster than the previous one.

Towards the goal of object/region recognition in images, texture characterization is a very important and challenging task. In this study, we propose a salient point based texture representation scheme. It is a two-phase analysis in the multiresolution framework of discrete wavelet transform. In the first phase, each wavelet sub-band (LH or HL or HH) is used to compute multiple texture features, which represents various aspects of texture. These features are converted into binary images, called salient point images (SPIs), via an automatic threshold technique that maximizes inter-block pattern deviation (IBPD) metric. Such operation may facilitate combining multiple features for better segmentation. In the final phase, we have proposed a set of new texture features, namely non-salient point density (NSPD), salient point residual (SPR), saliency and non-saliency product (SNP). These features characterize various aspects of image texture like fineness/coarseness, primitive distribution, internal structures etc. K-means algorithm is used to cluster the generated features for unsupervised segmentation. Experimental results with the standard texture (Brodatz) and natural images demonstrate the robustness of the proposed features compared to the wavelet energy (WE) and local extrema density feature (LED).

Recognition of rock fractures is crucial in many rock engineering applications. In order to successfully applying automatic image processing techniques for the problem of rock fracture detection and description, the key (and hardest task) is the robust image segmentation of rock fractures. A one-pass valley-edge detection algorithm (“valley” or (“ridge”) means here finding locally dark (or bright) line-like or curve-like features) was studied. The image segmentation algorithm is for delineating rock fractures based on multiple scale and valley-edge detection techniques. Results indicate that this approach is useful in this domain of images.

In this paper, we present an effective and robust shape-based leaf image retrieval system that supports two novel features: improved MPP algorithm and revised dynamic matching method. The improved MPP algorithm reduces the number of points for the shape representation considerably. Moreover, the new dynamic matching method, which is a revised Nearest Neighbor search, reduces the matching time. We implemented a prototype system based on these features and performed several experiments to show its effectiveness. We compare its performance with other known methods and report some of the results.

A wide-area image can be synthesized from an image sequence taken with a moving camera by using image mosaicing techniques. However, motion blur caused by the motion of the camera may significantly degrade the quality of the synthesized image. In this paper, we propose a new method for generating a deblurred mosaic from an image sequence that is degraded by motion blur under the condition that we do not have any information about the intrinsic and extrinsic parameters of the moving camera during input acquisition. In this method, we assume the objects in the scene can be classified into two regions in order to handle depth. In this paper, the displacement vectors of the features, which are computed using the KLT feature tracker on the consecutive frames, are classified into two regions. Here, the classified vectors provide a Point Spread Function (PSF) of the blurred image, and a homography between two consecutive frames for segmentation and mosaicing. Experimental results show that the Signal to Noise Ratio of the generated images can be significantly improved by our proposed method.

The aim of our project is to design an algorithm for counting people in public transport vehicles such as buses by processing images from surveillance cameras’ video streams. This article presents a method of detection and tracking of multiple faces in a video by using a model of first and second order local moments. The three essential steps of our system are skin color modeling, probabilistic shape modeling and bayesian detection and tracking. An iterative process is used to estimate the position and shape of multiple faces in images, and to track them in video streams.

In this paper, we propose an algorithm to characterize camera motion in video sequences based on image feature analysis. The approach predicts camera motion using spatio-temporal information obtained from tracking selected feature points throughout an image sequence. The spatio-temporal information provides the advantage of rich visual characteristic along a larger temporal scale over the traditional approaches, which tend to formulate computational methodologies on a few adjacent frames. The algorithm detects five basic camera motions of stationary, panning, tilting, zooming, and the combination of panning and tilting. We conduct the experiments to verify the proposed approach using real compressed video sequences. The experimental results have demonstrated the performance of proposed approach in determining camera motion.

Shape information is an important distribution to Content-Base Image Retrieval (CBIR) systems. There are two major types of shape descriptors, namely region-based and contour-based. In this paper we present a shape retrieval method that makes use of a contour-based descriptor, Principal Components Descriptor (PCD). In PCD, shapes are aligned on principal axes and described by a combination of the mean shape and weighted eigenvectors. The retrieval is achieved by comparing the weights of the eigenvectors. The developed approach is applied to Sharvit’s Silhouettes database and the results are compared with MPEG-7 standard contour-based descriptor, Curvature Scale Space (CSS). The comparison indicates that PCD shows higher accuracy than CSS.

Surveillance systems that automatically detect illegal behaviors performed by unaware people have a wide range of applications: security, healthcare, conservation of cultural heritage and so on. In particular monitoring public areas such as museums and archaeological sites is a challenging problem that has to be solved in order to avoid irreparable damages to historical heritage. In this paper a system able to check by common digital RGB cameras unexpected accesses to forbidden areas in a public museum is presented. The reliability of the proposed framework is shown by large experimental tests performed in the Messapic Museum of Egnathia (Italy) .

Cyclic strings are strings with no starting or ending point, such as those describing a closed contour. We present a new algorithm to compute a similarity measure between two cyclic sequences based on Dynamic Time Warping. The algorithm computes the optimal alignment between both sequences and is based on the cyclic edit distance algorithm proposed by Maes. The algorithm runs in

O

(

mnlg

m

) time, where

m

and

n

are the lengths of the compared strings. Experiments on a shape classification and shape retrieval with a public database are presented.

In a video surveillance system, moving object detection is the most challenging problem especially if the system is applied to complex environments with variable lighting, dynamic and articulate scenes, etc. Furthermore, a video surveillance system is a real-time application, so discouraging the use of good, but computationally expensive, solutions. This paper presents a set of improvements of a basic background subtraction algorithm that are suitable for video surveillance applications. Besides we present a new performance evaluation scheme never used in the context of moving object detection algorithms.

A novel real-time approach for classification or identification of objects is presented here that is suitable for visual attention system of mobile robots. The proposed method constructs convex hulls for regions found in an image using a new external scanning technique. Then a cleaning step produces refined polygons that are in turn used for extracting shape signatures for the regions. In the training phase, shape signatures are collected from test data to find a mean signature for a particular object. A small database is created for all objects related to a specific context in which classification is to be performed. In classifying phase, signatures obtained from objects found in a given image are compared with those present in the database for identification. Nearest signature from the database to a given one is taken as identity of the later. Results have proved efficiency and accuracy of this method.

In this paper, we investigate how to preprocess bad input face images for robust face recognition, under uneven illumination environments. Proposed filter combination shows nice performance under varying illumination, however, it can not provide the highest performance under normal illumination. We found that the performance of each preprocessing method for compensating illumination is highly affected by working illumination environment. Changing illumination poses a most challenging problem in face recognition. A previous research for illumination compensation has been investigated. This paper proposes a filter block for efficient face recognition. Since no priori knowledge of system working environment can be assumed. The proposed method can decide an optimal configuration of filter block by exploring the filter combination and the associated parameters to unknown illumination conditions. The illumination filter includes Retinex filter, end-in contrast stretching and histogram equalization filter. The proposed method has been tested to robust face recognition in varying illumination conditions (Inha DB, FERET DB). We made in illumination cluster using combined FART. Extensive experiment shows that the proposed system can achieve very encouraging performance in varying illumination environments. We furthermore show how this algorithm can be extended towards face recognition across illumination.

Fusion is basically extraction of best of inputs and conveying it to the output. In this paper, we present an image fusion technique using the concept of perceptual information across the bands. This algorithm is relevant to visual sensitivity and tested by merging multisensor, multispectral and defoucused images. Fusion is achieved through the formation of one fused pyramid using the DWT coefficients from the decomposed pyramids of the source images. The fused image is obtained through conventional discrete wavelet transform (DWT) reconstruction process. Results obtained using the proposed method show a significant reduction of distortion artifacts and a large preservation of spectral information.

Manually labeling video data is not only a labor intensive and time-consuming task, but also subject to human errors. In this paper, we present an automatic video annotation system. The system uses spatial attributions such as color, texture, shape, motion, and temporal hierarchical attributes among video objects. The system includes a new method of automatic video segmentation, object recognition and object-tracking scheme, and hierarchical object-based video representation model.

A moving target produces a coherent cluster of feature points in the image plane. This motivates our novel method of tracking multiple targets by cluster analysis of feature points and multiple particle filters. First, feature points are detected by a Harris corner detector and tracked by a Lucas-Kanade tracker. Clusters of moving targets are then initialized by grouping spatially co-located points with similar motion using the EM algorithm. Due to the non-Gaussian distribution of the points in a cluster and the multi-modality resulting from multiple targets, multiple particle filters are applied to track all the clusters simultaneously: one particle filter is started for one cluster. The proposed method is well suited for the typical video surveillance configuration where the cameras are still and targets of interest appear relatively small in the image. We demonstrate the effectiveness of our method on different PETS datasets.

Image scene classification is an integral part of many aspects of image processing. Indoor and Outdoor classification is a fundamental part of scene processing as it is the starting point of many semantic scene evaluation approaches. Many novel techniques have been developed to tackle this problem, but each technique relies on its own database of images thus reducing the confidence in the success of each method. We attempt here to look at the current field of indoor / outdoor scene classification and develop a benchmark model for evaluating current methods.

We propose a new technique for automatic spinal deformity detection from moire topographic images. Normally the moire stripes of a human body show a symmetric pattern. According to the progress of the deformity of a spine, asymmetry becomes larger. Numerical representation of the degree of asymmetry is therefore useful in evaluating the deformity. Displacement of local centroids and difference of gray value are calculated between the left-hand side and the right-hand side regions of the moire images with respect to the extracted middle line. Extracted 4 feature vectors (mean value and standard deviation from the each displacement) from the left-hand side and right-hand side rectangle areas apply to train a neural network. An experiment was performed employing 1,200 real moire images and 90.3% of the images were classified correctly.

A robust, efficient segmentation algorithm for automatic segmentation of MR images of the metacarpophalangeal joint is presented. A preliminary segmentation detects bones in MR scans and uses histogram analysis, morphological operations and knowledge based rules to classify various tissues in the joint. The second part of the algorithm improves the segmentation mask and refines boundaries of bones using minimization of a sum of square deviations, automatic signal segmentation into an optimum number of segments, graph theory, and statistical analysis. The algorithm has been tested on 9 MR patient studies and detects 97% of all existing bones correctly with an average exceeding 80% mutual overlap between ground truth and detected regions

This paper describes a lung field segmentation method, working on digital Postero-Anterior chest radiographs. The lung border is detected by integrating the results obtained by two simple and classical edge detectors, thus exploiting their complementary advantages. The method makes no assumption regarding the chest position, size and orientation; it has been tested on a non-trivial set of real life cases, composed of 412 radiographs belonging to two different databases. The obtained results and the comparison with more complicate techniques presented in the literature, prove the robustness of the algorithm and demonstrate that rather simple and general methods, properly combined to fit the requirements of a specific application, can provide better results.

Breast area segmentation or skin-line extraction in mammograms is very important in many aspects. Prior segmentation can reduce the effects of background noise and artifacts on the analysis of mammograms. In this paper, we investigate a novel method to estimate the breast skin-line in mammograms. Adaptive thresholding [1] yields a nearly perfect skin-line at the center of the image and around the nipple area with images from the MIAS database [2], but the upper and lower portions of the extracted boundary have been observed to be erroneous due to noise and artifacts. Because the distance from the edge of the stroma to the actual skin-line is usually uniform, we propose a method to estimate the skin-line from the edge of the stroma, with the information provided by the center portion around the nipple from adaptive thresholding. The results are compared with the ground-truth boundaries drawn by a radiologist [3] using polyline distance measure and shape smoothness measure. The results on 83 mammograms from the MIAS database are demonstrated. The proposed methods led to a decrease in a shape smoothness measure based upon curvature, on the average, from 65.6 to 20.0 over the 83 mammograms tested, resulting in an improvement of 69.5%.

More than 2 people out of 10 suffer from allergies, which can take various forms, from eczema to anaphylactic reactions with possible lethal consequences. Diagnosis is achieved through so-called “prick-tests” or IDR (intra-dermo-reaction): the injection of a small quantity of substances suspected to cause the allergic manifestation induces an erythema, the size of which is a useful indicator for the diagnosis. The manual surface measurement is time-consuming and inaccurate. This article presents a method for the semi-automatic measurement of the erythema from a photograph of the skin, taken in such conditions that lighting problems are minimized. The method is based on region growing and takes advantage of the most significant color spaces; the Lab space appears to be the best suited. It was tested on nearly 100 images, taken by various operators, on patients with various skin pigmentations; it gave promising results and proved to be robust.

Histopathological tissue analysis by microscopy is a process that is subjective, prone to inter- and intra-observer variation. This, along with the problems associated with verbalising visual elements of the diagnostic process, make learning the skill quite difficult. Training is long and largely relies on an “apprentice” model, where trainees learn the skill by witnessing an expert at work. Here we present the first findings of a longitudinal study of a group of histopathology trainees. By monitoring the progress of the trainees, we hope to be able to provide information that will improve training and assessment. In this paper we discuss the results of early data collection and analysis, from a web-based study of trainee classification accuracy and classification time.

Automated lung nodule detection through computed tomography (CT) image acquisition is a new and exciting research area of medical image processing. Lung nodules are potentially cancerous growths in the lungs that often appear in CT images as distinct, high intensity spherical objects. We have developed a nodule detection system. The first stage of the nodule detection technique automatically segments the lung regions using a unique 3D region growing approach. The next stage identifies regions of interests (ROIs) by using adaptive multi-level thresholding (MLT) based on the cumulative density function (CDF) of the lung volume. The last stage reduces false positives (FPs) by using unique features such as vessel and lung wall connectivity, a modified bounding box and 3D compaction to compensate for partial volume artifacts due to thick CT slices. We obtain a sensitivity of 80% with approximately 3.05 FPs per slice.

In this paper we show some preliminary results of our research in the fieldwork of classification of imbalanced datasets with SVM and stratified sampling. Our main goal is to deal with the clinical problem of automatic intestinal contractions detection in endoscopic video images. The prevalence of contractions is very low, and this yields to highly skewed training sets. Stratified sampling together with SVM have been reported in the literature to behave well in this kind of problems. We applied both the SMOTE algorithm developed by Chawla et al. and under-sampling, in a cascade system implementation to deal with the skewed training sets in the final SVM classifier. We show comparative results for both sampling techniques using precision-recall curves, which appear to be useful tools for performance testing.

Particle tracking is important for understanding the mobile behaviour of objects of varying sizes in a range of physical and biological science applications. In this paper we present a new algorithm for tracking cellular particles imaged using a confocal microscope. The algorithm performs adaptive image segmentation to identify objects for tracking and uses intelligent estimates of neighbourhood search, spatial relationship, velocity, direction estimates, and shape/size estimates to perform robust tracking. Our tracker is tested on three videos for vesicle tracking in GFP tagged videos. The results are compared to the popular Harvard tracker and we show that our tracking scheme offers better performance and flexibility for tracking.