Online Multi-Object Tracking Using Single Object Tracker And Markov Clustering

Jiao Zhu,Shanshan Zhang,Jian Yang

DOI: https://doi.org/10.1007/978-3-030-34113-8_46

2019-01-01

Abstract:In this paper, we address the challenging problem of online Multi-Object Tracking (MOT). We find for those targets that are occluded or too small, the detectors usually fail to locate them. But an SOT tracker always provides a prediction for each target in the next frame. Therefore, we propose to use Single Object Tracking (SOT) predictions as complementary to detections. Also, we solve the data association problem via a new clustering method based on the Markov Clustering Algorithm (MCL). We first build a graph based on the targets, detections and SOT predictions, and then separate different identities by clustering. Experimental results on the MOT17 benchmark shows that our proposed method outperforms previous state-of-the-art methods w.r.t. MOTA and also reduces the number of false negatives and fragments significantly.