Research On Moving Object Detection And Tracking In Intelligent Video Survellance

Intelligent video surveillance is a research hotspot in the field of computer vision in recent years. The major goal of the research is to perceive and to understand the information of the scenes by hierarchical computation and analysis of the image data based on the combination of the techniques of image processing, pattern recognition and artificial intelligence. The related researches have created great economic and social benefits in many application fields, such as intelligent security, intelligent transportation, intelligent tourism, smart city, Internet of Things and so on.Moving object detection and tracking is the core foundation and key technique in the field of intelligent video surveillance. Although great progresses have been reported in the area, the technical challenges remain in the aspects of accuracy, robustness and efficiency on the conditions that the complicated scene motion is in consideration. In this dissertation, the key techniques of moving object detection and tracking are explored and studied thoroughly using scene motion pattern as the fundamental starting point. The main work and innovations are summarized as follows:1) With the purpose of perception and expression of local area movement regulations of the scenes, a novel modeling method of scene motion pattern based on grid division is proposed in this dissertation. In the method, the scene is divided into several sub-regions using a square grid, and the corresponding oriented velocity weighted histograms of optical flow is established to describe the probability of the movement behaviors in different directions. Then the scene motion pattern model can be upgraded online utilizing the linear interpolation based voting weighted by the sparse optical flows in the scene, which are calculated using the KLT corner features extracted from the image sequences. Moreover, a weight compensation coefficient is introduced in the voting calculation to overcome the unbalance problem caused by different observation angles. It is indicated by the experiments that the model can accurately describe the movement regulation of different sub-regions, and that the algorithm of online upgrading is proved to be efficient.2) Aiming at the relationship between scene motion pattern and sub-region division, an improved modeling method of oriented scene motion pattern is proposed based on background image segmentation. As the background image of scene extracted by Gaussian mixture model is segmented to super pixels by SLIC segmentation algorithm based on k-mean clustering, the boundaries of the sub-regions are fit to the edges of layout elements in the scene. To describe the motion patterns of the sub-regions, a two-dimensional histogram of velocity and direction based on LK optical flow voting is set up. It is proved by the experiment that the motion probability density distribution of local area can be expressed more effectively by this model with fewer sub-regions. The experiment also proves that the model overcomes the inconsistent problem of the sub-region motion patterns, and thus has higher accuracy of movement regulation description. Meanwhile, the upgrading efficiency of the model can be proved to satisfy the requirement of the real-time surveillance system.3) A scene understanding approach based on the model of oriented scene motion pattern is presented aiming at the relationship between scene motion patterns and scene layout elements. In the beginning, according to the different influences of scene element on object movement, the scene region is classified into four categories:path, entrance, holding area and other ordinary region. For paths labeling, the high-speed motion probability density distributions of the sub-regions are obtained based on the corresponding optical flow histograms. Then transfer probabilities of sub-regions can be calculated using the direction and the voting similarity distances of motion behaviors in adjacent sub-regions. So the paths labeling problem can be simplified to an optimal path searching problem in a directed graph which is built up in terms of the adjacency relations of different sub-regions. The sub-regions at the ends of all the paths are clustered by DBSCAN algorithm to label the entrance region. And the wandering motion properties of the sub-regions are quantized using the variance of the movement probabilities in different direction for holding area labeling. Our scene understanding approach is proved effectively through the experimental results.4) According to the relationship between scene motion pattern and object detection, a parallel cascade oriented AdaBoost detection approach based on the scene motion pattern model is proposed in this dissertation. Firstly, for describing the color and edge qualities of object image, the Haar-like features and the HOG features are extracted using an improved integral image and integral histogram method. Then multiple Gentle AdaBoost sub-classifiers corresponding to different directions of the scene motion pattern model are trained using a cascade learning method. The ultimate decision are made by parallelly combining all the sub-classifiers based on the probabilistic descriptions of local motion in the model of scene motion pattern. It is proved experimentally that the AdaBoost detection is effectively enhanced in accuracy and efficiency of classification by the approach.5) A particle filter based object tracking method combined with an online scene motion pattern model is proposed in terms of the relation between scene motion pattern and object tracking. A channel weighted feature extraction algorithm of HSV color histogram is employed to describe the object image, and a cyclic correlation matrix of the hue channel is defined in Mahalanobis distance calculation of the features for offsetting the color distortion caused by illumination variation. To speed up the converging of the particles, the likelihood probabilities of the particle filter are corrected through the scene motion pattern model. A MCMC resampling algorithm is adopted to overcome the sample degeneration problem and sample impoverishment problem. Meanwhile, a sub-region traversal method based on convex polygon oriented scan line seed filling algorithm is proposed to realize the quick enumeration of the sub-regions overlaid by object motion. The experiments using two kinds of video datasets show that the average error of our tracking algorithm is significantly lower than that of standard particle filter tracking.