Research On Sparse Multi-Target Tracking And Trajectory Anomaly Detection In Video

With the increment of accidents, crime and terrorist activities, public safety is becoming increasingly important. Faced with these emergencies, intelligent video surveillance system can give timely warning signal or alarm. Compared with the conventional manual monitoring camera, intelligent monitoring system can save a lot of manpower, material and financial resources and can be more efficient to automatically or semi-automatically interpret and analysis these lawful video surveillance data. In the study of intelligent monitoring system, video foreground detection, multi-target tracking and identification of abnormal behavior as relatively new research directions have been becoming a hot topic in computer vision and pattern recognition field. Their research has a very important significance for improving the performance of intelligent monitoring system.Through algorithm analysis in these areas of the video foreground detection, multi-target tracking and identification of abnormal behavior, this thesis conducted in-depth research on the development of intelligent monitoring system. The following aspects are primarily completed in this thesis.1. Summarize and introduce common methods for the moving object detection in the areas of foreground detection and put forward an improved moving target detection method based on Gaussian mixture model. Compared directly with Gaussian mixture model based, this method improved the robustness and accuracy of the foreground detection, which significantly enhanced anti-jamming capability.2. In the tracking phase, a sparse multi-target tracking system framework is proposed for a single fixed camera. The framework will focus on the combination of a very good single-target tracking algorithms (TLD) and correlation matrix to effectively solve the occlusion problem in a multi-target tracking process with stronger anti-blocking property.In the consolidation phase of the target, a tracking window and bi-cubic interpolation algorithm are introduced to scale the initialization target and its tracking window in the same proportion for the combined state of the target. When the target is large, initialization and tracking by the TLD algorithm are relatively slowly. This operation reduces the tracking time. When the size of the target is less than the TLD initialization conditions, the operation solves the case where the target cannot be properly initialized. Special treatment is done for some special cases of the correlation matrix. Finally, in the filtering stage, the framework replace kalman algorithm with fractional kalman algorithm. It not only can reduce the measurement noise for maneuvering target, but also accurately predict the position of the target when the target appears intermittently lost.3. In the phase of anomaly detection based on the trajectory, the thesis presents a trajectory anomaly detection method with multiple-features representation based on time division. First, we propose a new trajectory feature representation method, which consists of six feature spaces:1) the direction and length of the track,2) the average position trajectory,3) the initial point, segment length of time, the orientations of the segments,4) the average velocity segmentation sequences.5) the average acceleration segmentation sequences.6) maximum acceleration of the entire trajectory. Then the surveillance-type SVM classification algorithm is used to train and detect the trajectory feature set. This method improves the abnormality detection rate and recognition accuracy of trajectory abnormal behavior, and lowers the false alarm rate. Meanwhile, the training and test samples do not need scaling processing, thereby greatly improving the practical value of this method.