A Multi Object Tracking Framework Based on YOLOv8s and Bytetrack Algorithm

Abstract

In recent years, the YOLOv8 series algorithms have become a research hotspot in many fields, and they can perform excellently in different computer vision tasks. However, YOLOv8 still has room for improvement in multi-target tracking. We integrated it with the Symmetric Positive Definite Convolution (SPD-Conv) module and proposed the YOLOv8s SPD detector, which enhances its detection ability for small targets. The values of [email protected] and [email protected]:95 have both been increased compared to YOLOv8s. Subsequently, the detector was combined with the ByteTrack tracking algorithm, and the IoU and loss function were optimized to achieve superior performance. We refer to this tracking framework as YBTrack. YBTrack was tested on the Multiple Object Tracking (MOT) Challenge MOT17 and MOT 20 datasets, and achieved MOTA metrics of 74.0% and 66.8%, respectively. Compared with existing tracking frameworks with built-in detectors, our tracking framework has better performance.

Keywords

• YOLOv8s
• Bytetrack
• SPD Conv
• MOT
• computer vision