Research Of Multi-Target Tracking Algorithm Based On Motion Detection And Feature Fusion

In recent years, MTT(multi-target tracking),as the significant research direction in visual tracking,has gained more and more attention from the schoolars home and abroad.The research heat is rising all the time. In the efforts of many scholars, the research in tracking, especially, whose targets are non-grid object like human body, has achieved great achievements. These made extensive application and long term research meaning in fields of civil use such as traffic flow statistic analysis, medical navigation, surveillance in bank region, and the field of military use such as TMD, AEW, air strike and so on.In the comprehensive analysis of some classic MTT theories,this paper,aiming at solving problems that it’s difficult to track multiple targets in the complex background using only one tracking technique,and go on tracking when overlap among in multiple targets,launches research about the issue-MTT based on motion detection and features fusion.Main works of this paper as follows:(1) The proposed robust motion detection algorithm.To avoid huge comuptation amount in tradional codebook algorithm as well as to take advantage of the merit that signal energy analysis method can reduce periodic noise caused by waving trees,water and shadow,this paper combines two methods that signal energy analysis and yuv codebook.(2) The proposed MTT algorithm based on motion detection and features fusion.Based on motion detection, this paper proposed a MTT algorithm,which take steps as follow to track multi-target steadily.First,it characterized H-S2D histogram and SIFT of the target and matched them. Second,it adopted PF as predictor to predict target position through out the tracking process.The last,it adopted Mean Shift to search target when overlap happened. The comparison experiment among KF, PF Mean Shift and the proposed MTT algorithm shows that not only dose the proposed MTT algorithm based on motion detection and features fusion meet real-time requirement, but also has good perfomance in tracking multi-target both in overlap and out overlap.