Non-Maximum Suppression Performs Later in Multi-Object Tracking

Abstract

Multi-object tracking aims to assign a uniform ID for the same target in continuous frames, which is widely used in autonomous driving, security monitoring, etc. In the previous work, the low-scoring box, which inevitably contained occluded target, was filtered by Non-Maximum Suppression (NMS) in a detection stage with a confidence threshold. In order to track occluded target effectively, in this paper, we propose a method of NMS performing later. The NMS works in tracking rather than the detection stage. More candidate boxes that contain the occluded target are reserved for trajectory matching. In addition, unrelated boxes are discarded according to the Intersection over Union (IOU) between the predicted and detected box. Furthermore, an unsupervised pre-trained person re-identification (ReID) model is applied to improve the domain adaptability. In addition, the bicubic interpolation is used to increase the resolution of low-scoring boxes. Extensive experiments on the MOT17 and MOT20 datasets have proven the effectiveness of tracking occluded targets of the proposed method, which achieves an MOTA of 78.3%.

Keywords

• multi-object tracking
• deep learning
• person re-identification