A Two-Stage Refinement Framework for Robust Vehicle Detection in Traffic Surveillance

Abstract

Traffic surveillance systems face challenges in vehicle detection in dense environments due to occlusions, pedestrians, and adverse conditions such as nighttime glare. Non-vehicle objects, including road signs and billboards, create noise and false positives, reducing detection accuracy and reliability. To address these issues, we propose a two-stage refinement framework: a pre-trained Co-DETR model eliminates irrelevant objects, followed by fine-tuned deep-learning models for precise vehicle detection. Additionally, detection stability is enhanced with Weighted Boxes Fusion (WBF), and image quality is improved through NAFNet for restoration and GSAD for low-light enhancement. Our approach significantly improves accuracy and robustness, achieving a mean Average Precision (mAP) of 0.9022 and a final score of 0.7779, which combines the F1 score and mAP, on the SoICT Hackathon 2024—Traffic Vehicle Detection Dataset.