Research And Implementation On Visual Multi-target Tracking Algorithms

Since nineteenth century, along with the rapid development of computer technology, radar, infrared, sonar, television, laser and visual target tracking technology have been developed continuously and improved gradually. As the most popular topics in the field of computer vision, visual target tracking has been used in many areas such as intelligent transportation, human-computer interaction, video surveillance, and behavior analysis, etc.Visual tracking system can be divided into two categories, single-target tracking and multi-target tracking. In this paper we study the problem of visual multiple targets tracking. It is a challenging task that we must deal with many problems, e.g., target auto-initialization, track end, track maintain, data association and inter-object occlusions in the multiple targets tracking.First, this paper presents an integrated framework for multi-target tracking, which is different from the traditional multi-target tracking-by-detection algorithms. We use robust visual tracking based on sparse representation and C1 minimization to follow individual pedestrian in the video over time. This single object tracker is initialized and periodically updated by a pedestrian detector. In order to produce longer-term and more accurate data association, the resulting tracklet outputs are fed to the middle-level association which is the second part of a detection-based three-level hierarchical association approach.Secondly, we propose an online tracking-by-detection algorithm for multi-object tracking in this paper. Objects which are initialized automatically are clustered into three categories:active objects, non-active objects and vanished objects, based on cues that are obtained from nearby objects, scene occluders and objects motion direction. The active objects are associated with detection responses using frame-by-frame data association. The non-active object is tracked by object-oriented hypothesize tracking. And the vanished object is deleted from object pool.Finally, we evaluate our method, which are robust visual tracking based on sparse representation and C1 minimization and cues-driven online multi-object tracking, on two challenging datasets:CAVIAR and i-Lids AB, by using MATLAB simulation tool. From the robust tracking results we can see that our two tracking system perform well.