Detection-Based Online Multi-target Tracking via Adaptive Subspace Learning

摘要

Multi-target tracking is a challenging task and becomes more so when both camera and targets are in motion and the targets have similar appearances with frequent occlusions. To maintain a proper track in such scenarios, individual target representation and accurate data association methods are prime requirements for a robust multi-target tracker. We observe that a target can be modeled as a subspace by using its feature vectors over several consecutive frames. We propose an adaptive sub-space model to handle the large range of target variations throughout the track. We also develop a novel two-step parallel scheme for data association which exploits scale and location information along with appearance information to distinguish the targets. The track results for challenging videos (containing occlusions and variations in pose and illumination) indicate that the proposed method achieves better/comparable tracking accuracy in comparison to several recent trackers.