Design Of Multi-Objects Detection And Tracking System Based On Improved Fast Particle Filter

In the intelligent video surveillance system, how to track the moving objects fastly and accurately is a key technology in the low-level analysis of the computer vision. At the same time, it is the basis of high-level computer vision processing such as scene understanding and behavior analysising. It also has various application values in military guidance, aerospace, smart city and the other fields. In recent years, many domestic and foreign scholars have put forward a lot of classical algorithms for targets tracking, but how to improve the robustness, real-time and accuracy of target tracking in complex scene was still a hot and difficult problem in the field of computer vision.This thesis is focused on how to track the moving targets fastly and accurately based on particle tracking algorithm under the interference of complex background, light, shadow, serious cross occlusion, missing-recurrence and the other harsh environment. So the algorithms of background modeling, particle filter, random ferns and online learning are studied in this thesis, then a particle filter tracking framework based on the combination of detector and tracker is proposed. Moreover, a fast particle filter tracking system is designed, which is suitable for high accuracy and real time in complex scene. The innovative results achieved are as follows:1. Moving object detection. The detection method of background modeling and object modeling are studied in this thesis. Firstly, the CodeBook background modeling is improved to reduce the influence of light and shadow, and the false detection rate. Secondly, the random ferns combined with the improved 2Bit-LBP feature training classifier is also adopted, which can absolutely achieve the requirements of learning, training and testing. Then, two improved methods of particle filter are proposed according to the two detection methods.2. Object tracking in static background. In order to reduce the computational complexity of the traditional particle filter in the premise of ensuring particles diversity, the appearance of the non-rigid object is preserved as a particle filter template by introducing the improved CodeBook background modeling. Meanwhile, the range of targets sampling is restricted to improve the accuracy and efficiency of the particle filter tracking algorithm.3. Object tracking in dynamic background. Owing to the severe state of motion and the missing-recurrence in the dynamic background, the particle filter was unable to continue tracking. In order to solve this problem, the random ferns classifier is studied in this thesis to make the detector and tracker running at the same time, and according to both of the confidence of tracking and detection by each other to correct tracking results. As a result, the accuracy and real-time performance are improved, and the objects can be tracked effectively after the recurrence due to serious occlusion.At the end of the thesis, the content of the research is summarized, then the future research direction and ideas are also pointed out.