Research On Multi-object Tracking Based On Layered Association

In this thesis, the detection based multi-object tracking from single camera is studied. On the basis of on-line layered association framework, the thesis has made detailed analysis and investigation on the local association, global association, and the feature extraction as well. By local association, the relationship between the candidate object of current image and the corresponding continuous track is established. By global association, the relationship between the candidate object of current image and the broken track segment is established. Furthermore, by introducing incremental linear discriminant analysis, the adaptive on-line learning is adopted so as to refreshing the association cost function between the continuous tracks and the candidate objects, more continuous and complete displacement tracks for multi-object are obtained.The main research work is as follows:(1) Local association with occlusion handling is proposed. When the occlusion is detected, the tracker divides the occluded target into blocks and gives a weight to each block. The more seriously the block occluded, the smaller its weight gain. Finally, the tracker determines whether the candidate targets belongs to the target according to all the information. And when extracting the color feature of the target, the thesis puts forward the weighted annular color histogram, taking into account the distribution probability and spatial information of color.(2) Global association based on motion prediction is proposed. First of all, the tracker uses Kalman filter to predict the location of the target where the track interrupted. Then, centering on the predicted location, the tracker does a fuzzy search. Finally, in the process of search, the tracker sees if there is the right candidate target to match according to all the features.(3) Adaptive multi-object tracking method based on incremental linear discriminant analysis is proposed. In order to distinguish between similar targets, this thesis introduces incremental linear discriminant analysis method. By online adaptive updating of the target feature, the matching cost function is adjusted in a timely manner. In this way, it is used to guide the growth of the track segments in the local association and the global association. A continuous and reliable tracking of multiple targets is realized.The performance of the proposed algorithms is examined through the open video sequences.The results show that local association with occlusion handling can handle partial occlusion effectively, global association based on motion prediction can improve the matching ofdisconnected track segments and candidate targets, and adaptive multi-object tracking method based on incremental linear discriminant analysis can effectively deal with the tracking of similar multi-object.The last part of the thesis gives a short summary of the work done, followed by outlook for future direction.