Multi-Object Tracking Based on Prediction of Invisible Object Trajectories

In previous 3D Multi-Object Tracking (3DMOT) methods terminated tracking when the tracked object was not associated with new detected object with in few frames (usually less than 5 frames). However, this can lead to premature termination of a tracker if an object simply became dimmer, such as being temporarily occluded by other objects or leaving the field of view. This became a major cause of identity switches in tracking. Additionally, premature termination of trajectory tracking is the main cause of identity switching in modern 3D MOT systems. To address this problem, we propose a method to preserve vanished object trajectories using trajectory prediction: when the life cycle management module performs trajectory management, unlike previous 3DMOT methods, when a tracked object is not matched by association within a few frames, instead of identifying the tracked object as vanishing and terminating its trajectory, our method predicts the vanishing tracked object state by the 3D Kalman Filter, which in turn comes to preserving its trajectory. With this approach, the experimental results on the nuScenes dataset show that the 3D multi-object tracking algorithm proposed in this paper performs excellently, with an Average Multi-Object Tracking Accuracy (AMOTA) of 73.7% at the frame rate of 2hz, and the Identity Switches (IDS) are also significantly reduced compared to the current excellent 3D multi-object tracking algorithms (e.g., about 50% less than SimpleTrack).