A Center-Based Light and Simple Method for Multi-Object Tracking

Traditionally, multi-object tracking is a technique for tracking points of interest through space and time. With the rise of powerful deep networks, this situation has changed. Nowadays, tracking mainly consists of pipelines that perform object detection and time correlation (also called tracking by detection). However, in order to extract more representative functions, many State-Of-The-Art (SOTA) models use a lot of complex backbones, which hindered the speed of inference. We proposed a new multi-object tracking model named Squeezed-Excited-Center-Based tracking (SECB-tracking), and replace the original IoU loss with the new Complete IoU loss, which can help the model converge faster. We evaluated our proposed method on the MOT17 dataset. Compared with CenterTrack, we have increased MOTA by 7 points (a relative increase of 10.3%) and reduced inference time by 20%.