Learning key lines for multi-object tracking

Most online multi-object tracking methods utilize bounding boxes and center points inherited from detectors as the base models to represent targets. Limited performance is obtained with these base models alone for tracking. Complex networks are generally applied on top to extract high-level discriminative features such as appearance embeddings and motion predictions for data association. However, the weakness in the feature representation of bounding boxes and center points degrades the tracking performance. In this paper, we propose a novel base model that represents targets with key lines for tracking, which can provide discriminative features and accurate target affinity measurements. Besides, we use the proposed key lines to select low-scored detections and unmatched tracks to recover missed targets and enhance identity consistency. Based on this, we apply the proposed line-based modeling strategy to existing trackers and propose a line-based Cascade Tracking algorithm to associate targets in three stages, and very competitive results are achieved on MOTChallenge benchmarks. Extensive experiments with improved performances demonstrate the effectiveness and generalization of key lines in providing discriminative features and enhancing tracking performance.