The Research Of Multiple Object Tracking With Mixed Strategies

Intelligent Video Surveillance system is a system which can detect and handle abnormal events automatically. And multiple object tracking is one of the core technologies in the system. And it is of vital important in many advanced applications, for example, type recognition of objects.We mainly focus on multiple object tracking under surveillance videos. Before tracking objects, we have to extract them. This paper introduce some classic algorithm of extracting objects and compare all of the strength and weakness, and for real-time and robustness we choose the VIBE(Visual Background Extractor) algorithm, which can handle the suppression of ghosts and static objects. For the foreground segmentation may result in no-connected objects, this paper use bounding box method to solve that problem. In tracking model, we choose different tracking methods vary from different occlusion occasions, and for that we propose a method to detect the state of merging and splitting between objects. When there is no occlusion between objects, the high efficient overlap method would be good because the correspondence between objects and detections is one-to-one. When there is light partial occlusion between objects, we employ a single object tracking method, KCF(Kernelized Correlation Filters) with improvement in knowing whether it is tracking the right object in this paper. When there is heavy partial occlusion between objects, we directly merge these objects to one big objects, and after these objects split, we assign the detections to these objects by feature similarity. In this paper, we use the color feature in HSV, and motion features to represent objects.Extensive experiments show that the propose method is very effective in real-time and robustness.