Motion-based Person Re-identification Algorithm Of Video Surveillance System

In recent years, with the development of video surveillance technology, more and more cameras are being deployed in security field, producing massive surveillance videos. The existing video surveillance systems can record video, or provide limited intelligence monitoring ability, such as flow monitoring, cross-border alarm and illegal capture. The identification and tracking of specific targets after the case occurred, mainly rely on a large number of police manual screening. Tracking specific targets in video surveillance systems is a promising work.Because of the particularity of the video surveillance systems, tracking problems in network video surveillance systems are actually target recognition problems. The existing target recognition algorithms based on extracting static features from images, which are not stable, and easy to camouflage. Considering pedestrian movements such as gait often contain rich information, new method based on the characteristics of motion to identify the pedestrians is proposed in this paper.The main work and contributions of this paper are summarized as follows:(1) The framework of the Re-Identification (Re-ID) algorithm is discussed, considering the structure of the video surveillance systems and the tracing task. According to the characteristics of surveillance video, improved Gaussian mixture background model is proposed for video surveillance camera background modeling and foreground extraction. The foreground objects are filtered and morphological processed to get intact regions, in order to satisfy the requirements of the subsequent feature extraction algorithm, also the gait cycle is determined.(2) A gait feature extraction algorithm based on optical flow is proposed in this paper. Method based on region segmentation and parameter motion models is used to compute dense optical flow information which is used to reflect the movement of each part of the human region at every sampling time. Then, original gait feature is descripted by synthesizing gait flow of a whole gait cycle with gait energy image. After that, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) methods are deployed to reduce the feature dimensions and produce well-performed feature. While the recognition rate of gait feature from different perspectives decreases significantly, the Truncated Singular Value Decomposition (TSVD) are deployed for feature transformation in order to increase the recognition rate in the recognition of gait features from different views.(3) A gait feature extraction method based on joint information is proposed in this paper. To extract human joint points from 2D image, an algorithm based on the proportion of human body is proposed. Then gait feature is extracted by studying the joint information. Because the gait characteristics reflected by the joint points are the specific patterns of time variation, and the sampling and gait cycles are not constant, alignment problems of gait feature will be produced. To solve this problem, the Dynamic Time Warping (DTW) algorithm are deployed to extract gait feature, and show a better performance in recognizing person of video surveillance systems in comparison with the time domain characteristics, frequency domain characteristics and the Hidden Markov Model (HMM) based gait recognition algorithm. In the end, Large Margin Nearest Neighbor (LMNN) method is deployed to learn a better distance metric. While using the learned distance metric, better recognition rate is achieved.