Enhanced Detection Model and Joint Scoring Strategy for Multi-Vehicle Tracking

Abstract

Multi-vehicle tracking is one of the most crucial components of an intelligent transportation system (ITS). However, when it comes to busy traffic flow, tracking targets robustly becomes more problematic due to occlusion, motion blur, high appearance similarity, etc. To achieve accurate and efficient tracking performance, we present a novel multi-vehicle tracking method based on the enhanced detection model and joint scoring strategy. Specifically, the former aims to (1) adopt lightweight yet efficient YOLOv5s to improve detection accuracy and running speed, and (2) incorporate the CBAM and transformer encoder modules into the detection model to generate the refined features for the target localization. The Latter preferentially provides high-confidence detections and tracklets for subsequent data association, significantly reducing the number of identity switches and redundant vehicle trajectories caused by mutual occlusion, similar object interference, etc. We evaluated the proposed multivehicle tracking approach on the UA-DETRAC vehicle tracking dataset and demonstrated its superior capabilities through intensive comparison and analysis. Moreover, our proposed method runs at 24.4 FPS on a single GPU and meets the real-time requirement.

Keywords

• Multi-vehicle tracking
• YOLOv5s
• transformer encoder
• joint scoring strategy
• high confidence