Image Analysis and Recognition

Video denoising is highly desirable in many real world applications. It can enhance the perceived quality of video signals, and can also help improve the performance of subsequent processes such as compression, segmentation, and object recognition. In this paper, we propose a method to enhance existing video denoising algorithms by denoising a video signal from multiple views (front-, top-, and side-views). A fusion scheme is then proposed to optimally combine the denoised videos from multiple views into one. We show that such a conceptually simple and easy-to-use strategy, which we call multiple view fusion (MVF), leads to a computationally efficient algorithm that can significantly improve video denoising results upon state-of-the-art algorithms. The effect is especially strong at high noise levels, where the gain over the best video denoising results reported in the literature, can be as high as 2-3 dB in PSNR. Significant visual quality enhancement is also observed and evidenced by improvement in terms of SSIM evaluations.

Example-based super-resolution has become increasingly popular over the last few years for its ability to overcome the limitations of classical multi-frame approach. In this paper we present a new example-based method that uses the input low-resolution image itself as a search space for high-resolution patches by exploiting self-similarity across different resolution scales. Found examples are combined in a high-resolution image by the means of Markov Random Field modelling that forces their global agreement. Additionally, we apply back-projection and steering kernel regression as post-processing techniques. In this way, we are able to produce sharp and artefact-free results that are comparable or better than standard interpolation and state-of-the-art super-resolution techniques.

The Eigenbackground model is often stated to perform better than pixel-based methods when illumination variations occur. However, it has originally one demerit, that foreground objects must be small. This paper presents an original improvement of the Eigenbackground model, dealing with large and fast moving foreground objects. The method generates background images using the Nelder-Mead Simplex algorithm and a dynamic masking procedure. Experiments show that the proposed method performs as well as the state-of-the-art Eigenbackground improvements in the case of slowly moving objects, and achieves better results for quickly moving objects.

Rain is a complex dynamic noise that hampers feature detection and extraction from videos. The presence of rain streaks in a particular frame of video is completely random and cannot be predicted accurately. In this paper, a method based on phase congruency is proposed to remove rain from videos. This method makes use of the spatial, temporal and chromatic properties of the rain streaks in order to detect and remove them. The basic idea is that any pixel will not be covered by rain at all instances. Also, the presence of rain causes sharp changes in intensity at a particular pixel. The directional property of rain streaks also helps in the proper detection of rain affected pixels. The method provides good results in comparison with the existing methods for rain removal.

Local phase coherence (LPC) is a recently discovered property that reveals the phase relationship in the vicinity of distinctive features between neighboring complex filter coefficients in the scale-space. It has demonstrated good potentials in a number of image processing and computer vision applications, including image registration, fusion and sharpness evaluation. Existing LPC computation method is restricted to be applied to three coefficients spread in three scales in dyadic scale-space. Here we propose a flexible framework that allows for LPC computation with arbitrary selections in the number of coefficients, scales, as well as the scale ratios between them. In particular, we formulate local phase prediction as an optimization problem, where the object function computes the closeness between true local phase and the predicted phase by LPC. The proposed method not only facilitates flexible and reliable computation of LPC, but also demonstrates strong robustness in the presence of noise. The groundwork laid here broadens the potentials of LPC in future applications.

In this paper the sliding wedgelet algorithm is presented together with its application to edge detection. The proposed method combines two theories: image filtering and geometrical edge detection. The algorithm works in the way that an image is filtered by a sliding window of different scales. Within the window the wedgelet is computed by the use of the fast moments-based method. Depending on the difference between two wedgelet parameters the edge is drawn. In effect, edges are detected geometrically and multiscale. The computational complexity of the sliding wedgelet algorithm is

O

(

N

2

) for an image of size

N

×

N

pixels. The experiments confirmed the effectiveness of the proposed method, also in the application to noisy images.

Traditional diffusivity based denoising models detect edges by the gradients of intensities, and thus are easily affected by noise. In this paper, we develop a nonlinear diffusion denoising method which adapts to the local context and thus preserves edges and diffuses more in the smooth regions. In the proposed diffusion model, the modulus of gradient in a diffusivity function is substituted by the modulus of a wavelet detail coefficient and the diffusion of wavelet coefficients is performed based on the local context. The local context is derived directly by analyzing the energy of transform across the scales and thus it performs efficiently in the real-time. The redundant representation of the stationary wavelet transform (SWT) and its shift-invariance lend themselves to edge detection and denoising applications. The proposed stationary wavelet context-based diffusivity (SWCD) model produces a better quality image compared to that attained by two high performance diffusion models, i.e. higher Peak Signal-to-Noise Ratio on average and lesser artifacts and blur are observed in a number of images representing texture, strong edges and smooth backgrounds.

Wavelet domain blur invariants, which were proposed for the first time in [10] by the authors, are modified in order to suit a wider range of applications. With the modified blur invariants, it is possible to address the applications in which the blur systems are not necessarily energy-preserving. Also, for a simpler implementation of the wavelet decomposition for discrete signals, we use a method which preserves an important property of the invariants: shift invariance. The modified invariants are utilized in two different experiments in order to evaluate their performance.

This paper introduces a Super Resolution Hough Transform (SRHT) scheme to address the vote spreading, peak splitting and resolution limitation problems associated with the Hough Transform (HT). The theory underlying the generation of multiple HT data frames and the registration of cells obtained from multiple frames are discussed. Experiments show that the SRHT avoids peak splitting and successfully alleviates vote spreading and resolution limitations.

In order to solve the problems that the image fusion method based on separable discrete wavelet transform is lower in spatial resolution and there is block effect in fused image, a new multispectral image fusion method based on non-separable wavelets with compactly support, symmetry, orthogonality, and dilation matrix [2,0;0,2] is proposed. A construction method of four channels 6 × 6 filter banks is presented. Using the low-pass filter constructed, multispectral images are fused. Three fusion methods called NAWS, NAWRGB and NAWL are proposed in the fusion of multispectral image and panchromatic image. Every fusion method presented outperforms the corresponding fusion method of the AWS, the AWRGB and the AWL in preserving high spatial resolution information respectively, and the higher spatial resolution fused image can be obtained. Of all fusion methods, the non-separable additive wavelet substitution (NAWS) method has the best performance in preserving higher spatial resolution information.

We derive mathematically a class of metrics for signals and images, considered as elements of

R

N

, that are based upon the structural similarity (SSIM) index. The important feature of our construction is that we consider the two terms of the SSIM index, which are normally multiplied together to produce a scalar, as components of an ordered pair. Each of these terms is then used to produce a normalized metric, one of which operates on the means of the signals and the other of which operates on their zero-mean components. We then show that a suitable norm of an ordered pair of metrics defines a metric in

R

N

.

There has been an increasing number of tone mapping algorithms developed in recent years that can convert high dynamic range (HDR) to low dynamic range (LDR) images, so that they can be visualized on standard displays. Nevertheless, good quality evaluation criteria of tone mapped images are still lacking, without which, different tone mapping algorithms cannot be compared and there is no meaningful direction for improvement. Although subjective assessment methods provide useful references, they are expensive and time-consuming, and are difficult to be embedded into optimization frameworks. In this paper, we propose a novel objective assessment method that combines a multi-scale signal fidelity measure inspired by the structural similarity (SSIM) index and a naturalness measure based on statistics on the brightness of natural images. Validations using available subjective data show good correlations between the proposed measure and subjective rankings of LDR images created by existing tone mapping operators.

In this paper, we propose an efficient sparse feature on-line learning approach for image classification. A large-margin formulation solved by linear programming is adopted to learn sparse features on the max-similarity based image representation. The margins between the training images and the query images can be directly utilized for classification by the Naive-Bayes or the K Nearest Neighbor category classifier. Balancing between efficiency and classification accuracy is the most attractive characteristic of our approach. Efficiency lies in its on-line sparsity learning algorithm and direct usage of margins, while accuracy depends on the discriminative power of selected sparse features with their weights. We test our approach using much fewer features on Caltech-101 and Scene-15 datasets and our classification results are comparable to the state-of-the-art.

In this paper, we demonstrate the use of relational space to classify microarray gene expression data. We also show that the transformation of real valued data to binary data is able to produce better class separation with fewer genes.

Spatial pyramid matching (SPM) has been one of important approaches to image categorization. Despite its effectiveness and efficiency, SPM measures the similarity between sub-regions by applying the bag-of-features model, which is limited in its capacity to achieve optimal matching between sets of unordered features. To overcome this limitation, we propose a hierarchical spatial matching kernel (HSMK) that uses a coarse-to-fine model for the sub-regions to obtain better optimal matching approximations. Our proposed kernel can robustly deal with unordered feature sets as well as a variety of cardinalities. In experiments, the results of HSMK outperformed those of SPM and led to state-of-the-art performance on several well-known databases of benchmarks in image categorization, even when using only a single type of feature.

We address the empirical feature selection for tracker-less recognition of human actions. We rely on the appearance plus motion model over several video frames to model the human movements. We use the L

2

Boost algorithm, a versatile boosting algorithm which simplifies the gradient search. We study the following options in the feature computation and learning: (i) full model vs. component-wise model, (ii) sampling strategy of the histogram cells and (iii) number of previous frames to include, amongst others. We select the features’ parameters that provide the best compromise between performance and computational efficiency and apply the features in a challenging problem, the tracker-less and detection-less human activity recognition.

We present a method to classify atomic density distributions using CCD images obtained in a quantum optics experiment. The classification is based on the scale invariant detection and precise localization of the central blob in the input image structure. The key idea is the usage of an a priori known shape of the feature in the image scale space. This approach results in higher localization accuracy and more robustness against noise compared to the most accurate state of the art blob region detectors.

The classification is done with a success rate of 90% for the experimentally captured images. The results presented here are restricted to special image structures occurring in the atom optics experiment, but the presented methodology can lead to improved results for a wide class of pattern recognition and blob localization problems.

Re-identification, that is recognizing that an object appearing in a scene is a reoccurrence of an object seen previously by the system (by the same camera or possibly by a different one) is a challenging problem in video surveillance. In this paper, the problem is addressed using a structural, graph-based representation of the objects of interest. A recently proposed graph kernel is adopted for extending to this representation the Principal Component Analyisis (PCA) technique. An experimental evaluation of the method has been performed on two video sequences from the publicly available PETS2009 database.

2D shape recognition from a set of points is largely used in several imaging areas such as geometric modeling, image visualization or medical image analysis. However, the perceived shape of a set of points is subjective. It is mainly influenced by the spatial arrangement of the points and by several cognitive factors. The Delaunay filtration methods derived from the well-known

α

-shapes, like LDA-

α

-shapes or conformal-

α

-shapes, provide a family of shapes capturing the intuitive notion of “crude” versus “fine” shape of a set of points. In this paper, a quantitative criterion based on shape measurements is defined for extracting the “optimal” shape from this family that best corresponds to the human visual perception. A novel automatic shape recognition method is proposed and successfully evaluated on the KIMIA image database, where the reference shapes are known and sampled by generating 2D point sets.

In this paper, the statistical effect of embedding data on Empirical Matrix (EM) of original and differential images is investigated and a novel steganalysis method, targeted at LSB Matching is proposed. It can be mathematically proven, that embedding data in a digital image, causes its empirical matrix and, also the empirical matrixes of its differential images to smooth. Therefore, the high frequency components of an image empirical matrix are omitted due to data hiding which motivates us to extract the radial moments of EM characteristic function as discriminative features for classification. Support Vector Machine with Gaussian kernel is adopted as an appropriate classifier in classification. Experimental results show that the extracted features are highly efficient in attacking LSB Matching.

A fully Bayesian approach to analyze infinite multidimensional generalized Gaussian mixture models (IGGM) is developed in this paper. The Bayesian framework is used to avoid model overfitting and the infinite assumption is adopted to avoid the difficult problem of finding the right number of mixture components. The utility of the proposed approach is demonstrated by applying it on texture classification and infrared face recognition, while comparing it to different other approaches.

The image classification process is based on the assumption that pixels which have similar spatial distribution patterns, or statistical characteristics, belong to the same spectral class. In a previous study we have shown how we can improve the accuracy of classification of remotely sensed imagery data by incorporating contextual elevation knowledge in a form of a digital elevation model with the output of the classification process using Dempster-Shafer Theory of Evidence. A knowledge based approach is created for this purpose using suitable production rules derived from the elevation distributions and range of values for the elevation data attached to a particular satellite image. Production rules are the major part of knowledge representation and have the basic form: IF condition THEN Inference. Although the basic form of production rules has shown accuracy improvement, in general, in some cases accuracy can degrade. In this paper we propose a “refined” approach that takes into account the actual “distribution” of elevation values for each class rather than simply the “range” of values to solve the accuracy degradation. This approach is performed by refining the basic production rules used in the previous study taking into account the number of pixels at each elevation within the elevation distribution for each class.

This paper presents a new model fitting approach to classify tentative feature matches as inliers or outliers during wide baseline image matching. The results show this approach increases the efficiency over traditional approaches (e.g. RANSAC) and other recently published approaches. During wide baseline image matching a feature matching algorithm generates a set of tentative matches. Our approach then classifies matches as inliers or outliers by determining if the matches are consistent with an affine model. In image pairs related by an affine transformation the ratios of areas of corresponding shapes is invariant. Our approach uses this invariant by sampling matches in a local region. Triangles are then formed from the matches and the ratios of areas of corresponding triangles are computed. If the resulting ratios of areas are consistent, then the sampled matches are classified as inliers. The resulting reduced inlier set is then processed through a model fitting step to generate the final set of inliers. In this paper we present experimental results comparing our approach to traditional model fitting and other affine based approaches. The results show the new method maintains the accuracy of other approaches while significantly increasing the efficiency of wide baseline matching for planar scenes.

A characteristics of imbalanced points is their localities an imbalanced point may be contiguous to some other imbalanced points in terms of 8-connectivity. A two-layer scheme was recently proposed for matching imbalanced points based on localities, where the first layer aims to build locality correspondence, and the second layer aims to build point correspondence within corresponding localities. Under the framework of the two-layer matching, we propose a hybrid representation of imbalanced points. Specifically, an imbalanced point in the first layer is represented by a discriminant SIFT-type descriptor, and in the second layer, the imbalanced point is simply represented by a patch-type descriptor (the intensities of its neighborhood). We will justify the rationale of the proposed hybrid representation scheme and show its superiority over non-hybrid representation with experiments.

This paper proposes an affine invariant contour description for contour matching, applicable to wide-baseline stereo correspondence. The contours to be matched can be either object edges or region boundaries. The contour descriptor is constructed locally using matrix theory and is invariant to affine transformations, which approximate perspective transformations in wide-baseline imaging. Contour similarity is measured in terms of the descriptor to establish initial correspondence, then new constraints of grouping, ordering and consistency for contour matching are introduced to cooperate with the epipolar constraint to reject outliers. Experiments using real-world images validate that the proposed method results in more accurate stereo correspondence for clutter scenes with large depth of field than point-based stereo matching algorithms.

Repeatability is widely used as an indicator of the performance of an image feature detector but, although useful, it does not convey all the information that is required to describe performance. This paper explores the spatial distribution of interest points as an alternative indicator of performance, presenting a metric that is shown to concur with visual assessments. This metric is then extended to provide a measure of complementarity for pairs of detectors. Several state-of-the-art detectors are assessed, both individually and in combination. It is found that Scale Invariant Feature Operator (SFOP) is dominant, both when used alone and in combination with other detectors.

Image retargeting is the process of adapting an existing image to display with arbitrary sizes and aspect ratios. A compelling retargeting method aims at preserving the viewers’ experience by maintaining the significant regions in the image. In this paper, we present a novel image retargeting method based on non-uniform mesh warping, which can effectively preserve both the significant regions and the global configuration of the image. The main idea of our method is sampling mesh vertices based on the saliency map, that is to say, we place mesh vertices more densely in the significant regions, defining different quadratic error metrics to measure image distortion and adopting a patch-linking scheme that can better preserve the global visual effect of the entire image. Moreover, to increase efficiency, we formulate the image retargeting as a quadratic minimization problem carried out by solving linear systems. Our experimental results verify its effectiveness.

We propose a segment based moving edge detection algorithm by building association from multi-frames of the scene. A statistical background model is used to segregate the moving segments that utilize shape and position information. Edge specific knowledge depending upon background environment is computed and thresholds are determined automatically. Statistical background model gives flexibility for matching background edges. Building association within the moving segments of multi-frame enhances the detection procedure by suppressing noisy detection of flickering segments that occurs frequently due to noise, illumination variation and reflectance in the scene. The representation of edge as edge segment allows us to incorporate this knowledge about the background environment. Experiments with noisy images under varying illumination changing situation demonstrates the robustness of the proposed method in comparison with existing edge pixel based moving object detection methods.

Keypoints detection and descriptors construction method based on multiscale Gauss-Laguerre circular harmonic functions expansions is considered. Its efficient acceleration procedure is introduced. Two acceleration ideas are used. The first idea is based on the interconnection between Gauss-Laguerre circular harmonic functions system and 2D Hermite functions system. The further acceleration is based on the original fast Hermite projection method. The comparison tests with SIFT algorithm were performed. The proposed method can be additionally enhanced and optimized. Nevertheless even preliminary investigation showed promising results.

In this paper we address the problem of re-identification of people: given a camera network with non-overlapping fields of view, we study the problem of how to correctly pair detections in different cameras (one to many problem, search for similar cases) or match detections to a database of individuals (one to one, search for best match case). We propose a novel color histogram based features which increases the re-identification rate. Furthermore we evaluate five different classifiers: three fixed distance metrics, one learned distance metric and a classifier based on sparse representation, novel to the field of re-identification. A new database alongside with the matlab code produced are made available on request.

Graph cut algorithms are very popular in image segmentation approaches. However, the detailed parts of the foreground are not segmented well in graph cut minimization.There are basically two reasons of inadequate segmentations: (i) Data - smoothness relationship of graph energy. (ii) Shrinking bias which is the bias towards shorter paths. This paper improves the foreground segmentation by integrating the statistical significance measure into the graph energy minimization. Significance measure changes the relative importance of graph edge weights for each pixel. Especially at the boundary parts, the data weights take more significance than the smoothness weights. Since the energy minimization approach takes into account the significance measure, the minimization algorithm produces better segmentations at the boundary regions. Experimental results show that the statistical significance measure makes the graph cut algorithm less prone to bias towards shorter paths and better at boundary segmentation.

In this paper, we propose a new approach for detecting people in video sequences based on geometrical features and AdaBoost learning. Unlike its predecessors, our approach uses features calculated directly from silhouettes produced by change detection algorithms. Moreover, feature analysis is done part by part for each silhouette, making our approach efficiently applicable for partially-occluded pedestrians and groups of people detection. Experiments on real-world videos showed us the performance of the proposed approach for real-time pedestrian detection.

Image-based fluid motion estimation is of interest to science and engineering. Flow-estimation methods often rely on physics-based or spline-based parametric models, as well as on smoothing regularizers. The calculation of physics models can be involved, and commonly used 2nd-order regularizers can be biased towards lower-order flow fields. In this paper, we propose a local parametric model based on a linear combination of complex-domain basis flows, and a resulting global field that is produced by blending together local models using partition-of-unity. We show that the global field can be regularized to an

arbitrary order

without bias towards specific flows. Additionally, the blending approach to fluid-motion estimation is more flexible than competing spline-based methods. We obtained promising results on both synthetic and real fluid data.

Texture analysis is used in numerous applications in various fields. There have been many different approaches/techniques in the literature for texture analysis among which the texton-based approach that computes the primitive elements representing textures using

k

-means algorithm has shown great success. Recently, dictionary learning and sparse coding has provided state-of-the-art results in various applications. With recent advances in computing the dictionary and sparse coefficients using fast algorithms, it is possible to use these techniques to learn the primitive elements and histogram of them to represent textures. In this paper, online learning is used as fast implementation of sparse coding for texture classification. The results show similar to or better performance than texton based approach on CUReT database despite of computation of dictionary without taking into account the class labels.

In this paper two criteria are presented to compute reduced sets of centers of maximal discs in the weighted <3,4> distance transform of 2D digital patterns. The centers of maximal discs selected by the above criteria are used as anchor points in the framework of 2D skeletonization and, depending on the adopted criterion, originate skeletons with different properties.

This paper presents an interactive method for 3D images segmentation. This method is based on a region adjacency graph representation that improves and simplifies the segmentation process. This graph representation allows the user to easily define some splitting and merging operations which gives the possibility to make an incremental construction of the final segmentation. To validate the interest of the proposed method, our interactive proposition has been integrated into a volumetric texture segmentation process. The obtained results are very satisfactory even in the case of complex volumetric textures. This same system, including the textural features and our interactive proposition, has been manipulated by specialists in sonography to segment 3D ultrasound images of the skin. Some examples of segmentation are presented to illustrate the interactivity of our approach.

This article deals with the problem of detecting the weak-symmetry of a simple polygon. The main application of this work is the automatic reconstruction of 3D polygons (planar or non-planar polylines) symmetric with respect to a plane from free hand sketching 2D polygons. We propose a provable approach to check on the weak-symmetry of a simple polygon. The worst time complexity of the proposed algorithm is

O

(

n

3

) where

n

is the number of the vertices of the input polygon.

In this paper, we introduce a new approach for color visualization of multi/hyperspectral images. Unlike traditional methods, we propose to operate a local analysis instead of considering that all the pixels are part of the same population. It takes a segmentation map as an input and then achieves a dimensionality reduction adaptively inside each class of pixels. Moreover, in order to avoid unappealing discontinuities between regions, we propose to make use of a set of distance transform maps to weigh the mapping applied to each pixel with regard to its relative location with classes’ centroids. Results on two hyperspectral datasets illustrate the efficiency of the proposed method.

Tracking a very actively maneuvering object is challenging due to the lack of state transition dynamics to describe the system’s evolution. In this paper, a coarse-to-fine particle filter algorithm is proposed for such tracking, whereby one loop of the traditional particle filtering approach is divided into two stages. In the coarse stage, the particles adopt a uniform distribution which is parameterized by the limited motion range within each time step. In the following fine stage, the particles are resampled using the results of the coarse stage as the proposal distribution, which incorporates the most present observation. The weighting scheme is implemented using a partitioned color cue that implicitly embeds geometric information to enhance robustness. The system is tested by a publicly available dataset for tracking an intentionally erratic moving human head. The results demonstrate that the proposed system is capable of handling random motion dynamics with a relatively small number of particles.

This paper presents a multi-camera surveillance system for motion detection and object tracking based on Motion History Image (MHI), Color-based Particle Filtering (CPF), and a novel relay strategy. The system is composed of two Pan-Tilt-Zoom (PTZ) cameras completely calibrated and placed on desks. Initially, both cameras work as stationary Scene View Camera (SVC) to detect objects for abnormal human motion events such as sudden falling using MHI. If an object is detected in one camera, the other camera can then be controlled to work as Object View Camera (OVC), follow this object, and get zoom-in images using CPF. The states of the tracked object can be exchanged across cameras so that in case that the OVC loses the object, the SVC has sufficient knowledge of the object location, and it can become a new OVC to run the tracking relay. Meanwhile, the original OVC should be reset to work as SVC in order not to lose the global view. Two scenarios, in which the cameras have large or little overlapping field of view, are proposed and analyzed. Experimental study further demonstrates the effectiveness of the proposed system.

Since there does not exist labelled samples during tracking period, most existing classification-based tracking approaches utilize a “self-learning” to online update the classifier. This often results in drift problems. Recently, semi-supervised learning attracts a lot of attentions and is incorporated into the tracking framework which collects unlabelled samples and use them to enhance the robustness of the classifier. In this paper, we develop a gradient semi-supervised learning approaches for this application. During the tracking period, the semi-supervised technology is used to online update the classifier. Experimental evaluations demonstrate the effectiveness of the proposed approach.

Multiple-target tracking represents a challenging question in uncontrolled scenarios. Due to high-level applications, such as behavioral analysis, the need of a robust tracking system is high. In a multiple tracking scenario it is necessary to consider and resolve occlusions, as well as formations and splitting of object groups. In this work, a method based in a hierarchical architecture for multiple tracking is proposed to deal with these matters. Background subtraction, blob detection, low-level tracking, collision detection and high-level appearance tracking is used to avoid occlusion and grouping problems. Experimental results show promising results in tracking management, grouping, splitting, occlusion events, while remains invariant to illumination changes.

This paper addresses the problem of tracking human motion in a movie sequence involving camera movement. We have developed an approach to track the bounding box of a human in motion without using any particular model. This method exploits motion vector fields from the image, then subtracts the motion caused by the camera to obtain the segmentation of the object. In addition, we introduce a multi-level tracking approach. This approach makes the tracking operation more robust, and less prone to errors. Experiments with movie sequences representing human walk are reported.

This paper presents an approach for object detection and event recognition in video surveillance scenarios. The proposed system utilizes a Histogram of Oriented Gradients (HOG) method for object detection, and a Hidden Markov Model (HMM) for capturing the temporal structure of the features. Decision making is based on the understanding of objects motion trajectory and the relationships between objects’ movement and events. The proposed method is applied to recognize events from the public PETS and i-LIDS datasets, which include vehicle events such as U-turns and illegal parking, as well as abandoned luggage recognition established by set of rules. The effectiveness of the proposed solution is demonstrated through extensive experimentation.