3D Multi-Object Online Tracking with Multi-View Clustering

Abstract

Multi-target tracking is an increasingly hot research topic. Since target tracking under a single view is difficult to solve the problem of target tracking, we propose an online tracking method that simultaneously locates and correlates the objects under multiple views. This idea consists mainly of a tracking-by-detection method and an internal association mechanism between cameras. First, we trained the LSTM classfication model for the tracking under a single view, in which we also retained the best features of unoccluded objects for re-identification. Then the hierarchical clustering is used for connecting the objects under different views with the accurate track fragments after camera calibration, and the output of clustering will be made for improving the effect of track fragments in the online tracking. In experiments, our method proved to be more robust to solve the occlusion problem, and can make superior performance in comparison with the state of the art methods on the Terrace videos and Passageway videos from EPFL CVLAB.