An Improved Real-Time Multiple Object Tracking Algorithm Based on YOLOv8
Xiao Peng, Chaobing Huang · 2023
The main task of Multiple Object Tracking (MOT) is to detect all the objects from a given scene and maintain the identity information of each target. Many tracking-by-detection algorithms use two deep neural networks for object tracking and identity information extraction separately. Although this approach can achieve better tracking results, due to the large inference delay of the networks, it usually cannot achieve the goal of real-time tracking. In this paper, YOLOv8 is applied for object detection. Additionally, person re-identification (Re-ID) modules are embedded into the detection heads of YOLOv8, enabling us to perform object detection and feature extraction within a single network. However, training object detection and Re-ID tasks concurrently presents significant difficulties. Therefore, a two-stage training method is proposed, which trains the object detection and Re-ID modules separately. Our method can perform real-time multi-object tracking at a speed of 44.9 FPS, and compared to the advanced single stage multi-object tracking method FairMOT, it can achieve better tracking performance (74.7% MOTA vs 69.1% MOTA) on the MOT17 validation dataset.