Strongly connected component multi-object tracking
Nii Longdon Sowah, Qingbo Wu, Fanman Meng, Bo Wu, King Ngi Ngan · 2016
Multi-object tracking (MOT) continues to gain more attention due to its relevance in the field of computer vision. Tracking-by-detection is one of the most used techniques in multi-object tracking, and this work follows that trend. Many methods generate tracklets in the initial step of tracking and generate final trajectories by optimization algorithms. An important challenge of such approaches is how to generate reliable tracklets which is crucial for generating final trajectories. Bounding box overlap and optical flow are two of such methods. A new clustering method is proposed for tracklet generation in our multi-object tracking algorithm. We present a novel approach that uses strongly connected component clusters in a K-NN directed graph with no defined class number for tracklet generation. We propose that such clusters of detections can accurately represent the tracklets of targets in a video. Binary Integer Programming is used to merge initial tracklets into their final trajectories. Experimental results on four standard benchmark datasets demonstrate the superiority of the proposed method.