Research And Implementation Of Real-time Multi-pedestrian Detection And Tracking Algorithm

Due to the rising demand on safety management of traffic, manufacture, and public security, inefficient traditional monitoring approaches cannot meet the actual processing requirements. The video-based intelligent visual detection and tracking system aims at detecting and tracking multiple pedestrians in video in real-time. It can automatically monitor or assist the monitoring process with improved efficiency and effectiveness, meanwhile needs less manpower and resources. It can generate alarms for unsafe behaviors, provide support for management, and bring economic and social benefits.In this thesis, a real-time multi-pedestrian detection and tracking system is studied and implemented. The background modeling module, pedestrian detection module, and pedestrian tracking module are described. In order to be real-time, novel fast algorithms are implemented in all modules. A background modeling algorithm based on VIsual Background Extractor is studied. In order to improve the accuracy of subsequent pedestrian detection, an background modeling algorithm with postposition of foreground map for denoising and association of connected regions is proposed. Histograms of Oriented Gradients + Local Binary Pattern features and Support Vector Machine classifier are used for pedestrian detection, which is applied to a multi-point long-time railway surveillance video project with complex scenes and unsatisfactory image quality. A full-sampling fast tracking algorithm based on Fourier transform are studied. We propose to aid tracking drift correction and update tracking scale by incorporating foreground information, which can correct the deviation of automatically specified targets quickly and smoothly, so as to prevent loss of target. It improves the scale following ability and the tracking accuracy with small computation cost, without changing the complexity of the algorithm.Experimental analysis is performed for each module in detection and tracking. The algorithm with postposition of foreground map is proved to enhance the performance of pedestrian detection. The false positive and false negative rates of the pedestrian detection algorithm are tested under varying parameters. The speed, precision of position and scale, and self-correcting capacity of six tracking algorithms are tested and compared. Overall, the proposed improved algorithm is superior to other algorithms. Finally the characteristics and the actual usage of the entire system are illustrated.