Research On Object Tracking And Recognition In Video Surveillance

Video surveillance systems have been widely used for object detection, tracking andrecognition based on the existing knowledge of computer vision (CV) and imageprocessing techniques. They are able to provide the observers with large amount of valuableinformation, which is an important way to build a―smart city‖and―safe city‖. Recent afew years, with the development of computer science, researchers have made great effortsin the video analysis algorithms, and played a facilitating role in the application andpromotion of video surveillance systems. However, due to the complex surveillanceenvironment (e.g., illumination change) and the motion of targeted objects (e.g., cameramotion, targets position and scale variation), it is still challenging to design real-time androbust video analysis algorithms. In this thesis, we focus on the key techniques in videosurveillance system, including pre-processing in illumination variation, target tracking,object recognition, etc. The main contributions are as follows:(1)To address the bad effects on target tracking and recognition caused by illuminationchanges in complicated scene, this thesis proposes a Local-to-Global Adaptive illuminationsuppression (LGA). The traditional illumination suppressing methods based on imageenhancement algorithms are divided into two categories: local-feature based algorithms andglobal-feature based ones, without any combinations. The former ones can achieve goodperformance at the cost of high computational complexity, while the latter ones have lowcomplexity but lose some image details. To tackle these problems, the proposed methodfirst utilizes the mean value and global feature of the original image to establish the Gammacorrection coefficient table. An SVLM mapping image is then obtained based on themulti-scale neighbor information, and the local image contrast can be enhanced accordingto the Gamma correction coefficient of each pixel. Furthermore, the histogram smoothingcoefficients of each gray scale can be obtained adaptively using the local similarityhistogram, which results in higher use ratio of gray scales and global contrast enhancement.Experimental results demonstrate that the proposed algorithm performs better than othercontrast enhancing methods as well as improving the detection and recognition rate in facerecognition applications. (2) The second contribution is that a real-time object tracking algorithm is proposedcombining an online sequential Extreme Learning Machine and particle filter(OSELM-PFT). The proposed method is effective in handling the object drift due to objectposition variation. The classification-based object tracking algorithms convert trackingproblem into classifier design, which is capable to adapt to the position variation of objectand constrain the―drift‖phenomenon with sufficient training samples of targets andbackground. The proposed OS-ELM-PFT tracking method has the following advantages:(1)parameters in the input layer of ELM are randomly generated with output layer parameterscomputed using the least square. Therefore, there exist no complex iterating steps, whichmeets the real-time requirement of tracking systems;(2) the training samples are added toOS-ELM neural network individually or in batches without re-training resulting in highadaptivity to target motion as well as higher classification accuracy between the targets andbackground and tracking accuracy;(3) to find the target location, the particle filter isutilized to generate limited candidate samples;(4) the posterior of the classification result iscomputed to further reduce the negative effect of classification errors. Thoroughexperiments illustrate the proposed method performs better than most of the state-of-the-artmethods both in computation efficiency and tracking accuracy.(3) In order to obtain more stable tracking results, an OS-ELM tracking algorithm isproposed using a visual dictionary. The proposed method can handle the problems of targetdisappearing and target scale adaptation existing in the abovementioned work OSELM-PFT.The algorithm regards the visual dictionary as detector, and OS-ELM and particle filter astrackers, and utilizes the feedback technique to improve tracking performance. The trackeris aimed to implement the target tracking as well as providing samples for the establishmentand updating of the visual dictionary. The detector aims to detect and judge targetdisappearing and redetect the target after disappearing. For the purpose to improverobustness to part occlusion of targets, a histogram similarity measurement is proposed tobe introduced into the detector using local random sampling, according to the idea of localpartition and Noisy-NR model in computing the similarity of candidate samples andtraining samples. The proposed measurement has addressed the tracking failure of thetraditional histogram based methods due to part occlusion of targets. In order to be adaptive to target scale variation, the Ransac algorithm is introduced to acquire the scale changecoefficient and update the detector. The experiments demonstrate that the proposed methodis robust to target scale variation, and able to detect if target disappears.(4) In order to build an illumination, position and scale invariant face recognitionsystem, firstly, we apply the abovementioned three works for extracting the illuminationinvariant face images with multi-position and multi-scale after the scale normalizing ofimage. Secondly, a blur detection method is put forward to eliminate the vague faceaccording to the proposed bi-Gaussian model, which can increase the recognition ratesignificantly compared with traditional methods without motion blur. Lastly, we propose aface recognition algorithm based on the image set and probability appearance manifold inorder to handle the position variation problem. The algorithm utilizes the probabilityappearance manifold to obtain accurate partition of the whole face position in the test set,which can form the position manifold of the test set. Then the target categories in the testset is acquired by calculating the geodesic distance between the position manifold and thecorresponding manifold in the training set. The above position label method exploits timedomain relation among images by the probability appearance manifold. Similarly, thegeodesic distance between the manifolds of test set and training set instead of the distanceof a single image to manifold is able to adapt to position variation. Experimental resultsshow that the proposed method performs better than other video-based and image-basedface recognition algorithms.