Computer Vision in Control Systems-4

This chapter contains a brief description of the methods, algorithms, and implementations applied in many fields of computer vision. The graphological analysis and identification of handwritten manuscripts are discussed using the examples of Great Russian writers. A perceptually tuned watermarking using non-subsampled shearlet transform is a contribution in the development of the watermarking techniques. The mobile robot simultaneous localization and mapping, as well as the joined processing of visual and audio information in the motion control systems of the mobile robots, are directed on the robotics’ development. The ambient audiovisual monitoring based on a wide set of methods for digital processing of video sequences is another useful real life application. Processing of medical images becomes more and more complicated due to the enforced current requirements of medical practitioners.

The problem of recognition of handwriting text is still far from its final solution. The existing systems of recognition of handwritten texts are usually developed for some special applications. The difficulties are caused by recognition of the conjoint writing because a variability of handwritings is the highest and often it is necessary to solve the problem of delimitation of the separate letters. In this chapter, along with to the known methods of the handwritten fragments’ analysis, it is offered to use the developed methods of vectorization of raster images and vector dynamic parameterization. Also, a description of the automated information storage and retrieval system for the graphological analysis and identification of unintelligible fragments of handwritten texts is given. The system contains a database of handwriting samples with variants of the author’s calligraphy from the Manuscript Department of the Institute of Russian Literature (Pushkin’s House) of the Russian Academy of Sciences.

Digital watermarking remains the rapidly developed branch of the computer science due to the huge amount of internet resources, requiring a defense. In the past decades, many promising methods for a watermark embedding in frequency domain were elaborated. At the same time, the excellent concept of human visual system properties ought to be applied in these new transforms. The Non-Subsampled Shearlet Transform (NSST) is one the most perspective techniques, providing a high level of visibility and payload for the host image. The perceptual channel decomposition is used for detection of textural regions or regions with the expressed edges that have high contrast values. The embedding process is executed using the NSST and the Singular Value Decomposition (SVD), when the last significant bits in a sequence of eigenvalues are replaced by the embedded watermark in a binary representation. The algorithm is reinforced by the Arnold’s transform of a watermark, the parameters of which are stored in a secret key. The quality of the extracted watermarks under typical attacks, such as the scaling, disproportionate scaling, rotation, translation, and cropping, was estimated by the Bit Error Rate (BER) and the Peak Signal to Noise Ration (PSNR) metrics. The proposed method indicates the highest robustness to rotation and proportional scaling (the BER mean values are 1.2–2.7%) and the medium robustness to translation and cropping (the BER mean values are 10.9–12.4%).

The research considers the implementation of simultaneous localization and mapping algorithm based on the FastSLAM technique and specific problems that are typical for the RGB-D sensor-based solutions. An improvement of the classical FastSLAM algorithm has been obtained by replacing the method of landmarks’ observations filtering with unscented Kalman filters. Instead of linearizing, the nonlinear models through the first order Taylor series expansion at the mean of the landmark state were applied. The proposed algorithm computes a more accurate mean and uncertainty of the landmarks, which are moving nonlinearly. Various data preprocessing issues are discussed, such as the method of calibration of Kinect-like cameras, depth map restoration using a modified interpolation technique, and filtering the noise in the RGB images for more accurate detection of key features. Additionally, the chapter presents an improved resampling algorithm for the particle filtering through the adaptive thresholding based on the data of the effective particle number evolution. The proposed algorithm runs in real time and shows good accuracy and robustness in comparison with other modern SLAM systems using all the advantages and disadvantages of the RGB-D sensors.

Decision making for movement is one of the essential activities in motion control systems of mobile robots. It is based on methods and algorithms of data processing obtained from the mobile robot sensors, usually video and audio sensors, like video cameras and microphone arrays. After image processing, information about the objects and persons including their current positions in area of mobile robot observation can be obtained. The aim of methods and algorithms is to achieve the appropriate precision and effectiveness of mobile robot’s visual perception, as well as the detection and tracking of objects and persons applying the mobile robot motion path planning. The precision in special cases of visual speaking person’s detection and tracking can be augmented adding the information of sound arrival in order to receive and execute the voice commands. There exist algorithms using only visual perception and attention or also the joined audio perception and attention. These algorithms are usually tested in the most cases as simulations and cannot provide a real time tracking objects and people. Therefore, the goal in this chapter is to develop and test the fast parallel algorithms for decision making in the motion control systems of mobile robots. The depth analysis of the existing methods and algorithms was conducted, which provided the main ways to increase the speed of an algorithm, such as the optimization, simplification of calculations, applying high level programming languages, special libraries for image and audio signal processing based on the hybrid hardware and software implementations, using processors like Digital Signal Processor (DSP) and Field-Programmable Gate Array (FPGA). The high speed proposed algorithms were implemented in the parallel computing multiprocessor hardware structure and software platform using the well known NVIDIA GPU processor and GUDA platform, respectively. The experimental results with different parallel structures confirm the real time execution of algorithms for the objects and speaking person’s detection and tracking using the given mobile robot construction.

In this chapter, the methods and algorithms of audio and video signal processing for analysis of indoor human activity are presented. The concept of ambient intelligent space and several implementations are discussed. The main idea is the development of proactive information services based on the analysis of user behavior and environment. The methods of image processing, such as illumination normalization and blur estimation based on focus estimation methods and image quality assessment, are described afterwards. A short overview of face recognition methods including the principal component analysis, Fisher linear discriminate analysis, and local binary patterns is presented. Their efficiency was subjected to comparative analysis both in terms of the processing speed and precision under several variants of the area selected for image analysis, including the procedure seeking face in the image and limitation of the size of the zone of interest. Several approaches to audiovisual monitoring of meeting room participants are discussed. The main goal of the described system is to identify events in the smart conference room, such as the time, when a new user enters the room, a speech begins, or an audience member is given the floor. The participant registration system including face recognition accuracy assessment and recording system involving assessment results of the pointing of the camera, when photographing participants are presented.

Video surveillance has a wide variety of applications for indoor and outdoor scene analysis. The requirements of real-time implementation with the desired degree of recognition accuracy are the main practical criteria for most vision-based systems. When a person is observed by a surveillance camera, usually it is possible to acquire multiple face images of a single person. Most of these images are useless due to the problems like the motion blur, poor illumination, small size of the face region, 3D face rotation, compression artifacts, and defocusing. Such problems are even more important in modern surveillance systems, where users may be uncooperative and the environment is uncontrolled. For most biometric applications, several of the best images are sufficient to obtain the accurate results. Therefore, there is a task for develop a low complexity algorithm, which can choose the best image from a sequence in terms of a quality. Automatic face quality assessment can be used to monitor image quality for different applications, such as video surveillance, access control, entertainment, video-based face classification, and person identification and verification. In practical situation, the normalized images at the output of face detection algorithm are post-processed and their quality is evaluated. Low quality images are discarded and only images with acceptable qualities are received for further analysis. There are several algorithms for face quality assessment that are based on estimating facial properties, such as the estimating the pose, calculating the asymmetry of the face, and non-frontal illumination to quantify the degradation of a quality. Several investigations show that application of a quality assessment component in video-based face identification system can significantly improve its performance. Another possible application of face quality assessment algorithm is to process the images with different qualities in different ways. Proposed face quality assessment method has been applied as a quality assessment component in video-based audience analysis system. Using the proposed quality measure to sort the input sequence and taking only high quality face images, we successfully demonstrated that it not only increases the recognition accuracy but also reduces the computational complexity.

Designing systems to detect the human gestures and movements is an important area in computer vision. In this chapter, a method to detect human eye blink patterns is proposed. Our system detects the user’s eye blink patterns in real time and responds with an action on a mobile device, such as the phone call, text message, and/or an alarm. In this chapter, several image processing techniques are used for detecting human eye blinks. To examine the state of the eyelid, whether it’s opened or closed, the eye state value is used by computing the minimum threshold. The system is able to track the blinking of the eyes efficiently and accurately from the video using the proposed method. This system is user-friendly and easy to operate. The experiment was performed under different conditions by changing the distance from the camera and light in the room. The experimental results showed that the overall detection rate for eye blink is 98%. The proposed method takes only 8 ms as the average execution time for each frame, which makes it work more efficiently in real time applications.

Image processing techniques play an important role in the diagnostics and detection of diseases and monitoring the patients having these diseases. The chapter presents the medical image processing and morphological analysis in the solution of urology and plastic surgery (hernioplasty) problems. Novel methodology for processing medical images using a color coding of contour representation obtained by Digital Shearlet Transform (DST) has been presented. The object contours in the medical urology images are obtained using the conventional filters, and then results are compared. Since medical images can contain some noise, it makes sense to suppress the noise at the preprocessing step. For this purpose, the optimized in implementation algorithms of the most frequently used filters, such as the mean filter, Gaussian filter, median filter, and 2D cleaner filter, had been developed. A comparison of the optimized and ordinary implementations of noise reduction filter shows great speed improvement of the optimized implementations (around 3–20 times). Additionally, the parallel implementation gives 2–3.5 times performance boost. The proposed methodology allows to improve the accuracy and decrease the error of the sought parameters and characteristics by 10–20% on average without a lack of significant details in the structural features of the examined objects. The results of the experimental study show an error decrease in data representation for the plastic surgery (hernioplasty) by 15–25%.

In this chapter, the methods of medical image processing and analysis in Clinical Decision Support Systems (CDSS) are discussed. The main principles of image analysis with the aim of differential diagnostics in the CDSS are determined. The implementation is given through the method of multispectral images automatic processing and analysis for TV system of cervix oncological changes diagnostics. The method provides differential diagnostics of the following changes in cervical tissues as Norm, Chronic Nonspecific Inflammation (CNI), Cervical Intraepithelial Neoplasia in various types of oncological changes (CIN I, CIN II, CIN III). In proposed method, images of different type (fluorescent images and images obtained in white light illumination) are analyzed. The decision rules in the classification task are based on data mining methods. For the border CIN/CNI sensitivity 87% and specificity 75% are achieved. The detail description of main steps is given in the chapter.

Foot Progression Angle (FPA) detection is an important measurement in clinical gait analysis. Currently, the FPA can only be computed, while walking in a laboratory with a marker-based or Initial Measure Unit (IMU) based motion capture systems. A novel Visual Feature Matching (VFM) method is presented here, measuring the FPA by comparing the shoe orientation with the progression, i.e. the walking direction. Both the foot orientation and progression direction are detected by image processing methods in rectified digital images. Differential FPA (DFPA) algorithm is developed to provide accurate FPA measurement. The hardware of this system combines only one wearable sensor, a chest or torso mounted smart phone camera, and a laptop on the same Wi-Fi network. There is no other prerequisite hardware installation or other specialized set up. This method is a solution for long-term gait self-monitoring in a home or community like environments. Our novel approach leads to simple and persistent, real time remote gait FPA monitoring, and it is a core of new bio-feedback medical procedure.