Learning Deep Appearance Feature for Multi-target Tracking

Abstract

Multi-target tracking is a worthy studying issue in computer vision. For surveillance video, frequent occlusion and dense crowds complicate the issue. To resolve these difficulties, this paper proposes an effective algorithm of multi-target tracking in videos. Firstly, the faster Rcnn is proposed with the residual network to extract the objects of pedestrians in surveillance videos. The proposedment can effectively eliminate invalid target detection frames, separate peer targets and resist partial occlusions. Then, this paper put forward an accurate and efficient appearance-feature matching network model that is inspired by pedestrian re-identification theory. The deep learning feature-extraction module is composed of the stem Cnn and the Resnet blocks, therefore it can load res-50 caffemodel as pretraining model to increase the accuracy of the featureextraction. Meanwhile, the proposed network can decrease the time of train and test comparing with Resnet. Finally, the obtained multiple target tracking trajectories are further optimized by the strategy of occlusion distinction, deduplication and merging. The experiment results of the 2D MOT 2015 benchmark, KITTI dataset indicate that this proposed algorithm outperforms alternative multiple objects trackers in terms of multiple indicators.