Research On Multi-human Tracking Based On Tracklet Association

ABSTRACT:Multi-human tracking in videos is a research field which tracks multiple targets in a video sequence by using video analysis technologies, and maintains the unique ID for each target. Target tracking has many applications in video analysis field, such as smart video surveillance, computer vision based human computer interaction and so on. However, target tracking attracts more attention, since it is a key technology in the video semantic analysis.The challenge of tracking multiple targets is to face the frequent inter-occlusions of multiple targets due to their movements. This is a big challenge to track each target correctly. Recent years, some new ideas have appeared to solve multi-human tracking named global association. Some researchers also call this idea as tracking by tracklet association. However, the tracking methods based on tracklet association is not mature. Some research works need to do in building tracklets, feature extraction, and tracklet association. For the difficulties in multi-human tracking in complex scenes, the research of this paper includes the following three aspects:1) Particle filter based adaptive tracklet building method:In the framework of tracking by tracklet association, building tracklets is the first step of the whole framework. The traditional tracklet building method links the detection responses in consecutive frames to grow tracklets. The basic idea of this method is that the size and appearance of targets have no significant changes. However, in the complex scenes, due to the complex background and human pose to cameras, the image classification based human detector is difficult to detect human in each frame. So, the tracklets can not grow sufficiently. This paper proposes an adaptive tracklet building method based on particle filter. In the process of building tracklet, we use particle filer and combine detection results to grow tracklets. Then, according to the posterior distribution of particle filter, tracklet is terminated adaptively. This can improve the construction of tracklets and is good for feature extraction and association of tracklets.2) Appearance feature extraction:Appearance feature extraction is an important problem in video tracking. In this paper, we propose an appearance representation method named spatial-temporal appearance model. In each tracklet, according to adaptive clustering, this method clusters appearance pixels of each detection in space and time, to form subregions which cross several frames. Compared with previous appearance features which are extracted in single frame, and do not consider the correlation in consecutive frames, our appearance model provides not only the dynamic subregions of target appearance, but also the duration time for each subregion.3) Tracklet association:In order to associate tracklets to form the final tracking results, we propose an adaptive appearance selection based on spatial-temporal appearance model, and combine with target motion feature based on fuzzy search range to associate tracklets. Adaptive appearance selection is the feature selection to compute the minimal appearance distance between two tracklets. If the minimal appearance distance can not achieve the global minimal, the two tracklets can not be associated. The motion feature is used to help tracklet association, and the fuzzy search range can reduce some effect of the imprecise motion feature due to non-linear motion.In summary, this paper focuses on the research of multi-human tracking based on tracklet association, including tracklet building method, appearance feature extraction, and tracklet association, to improve the performance of multi-human tracking by tracklet association.