Lightweight traffic object detection based on knowledge distillation

Traffic object detection plays a critical role in ensuring intelligent driving safety. While existing lightweight general-purpose models meet the computational resource constraints of onboard systems, they struggle with low detection accuracy in complex and dynamic traffic scenarios. To address this, this work introduces target region priors to guide models toward focusing on traffic-relevant areas, thereby enhancing feature representation capabilities. Furthermore, we propose a knowledge distillation framework incorporating object region distillation and category/position-aware distillation strategies to transfer the teacher model's robust feature characterization to lightweight student models. The proposed methods are simple yet effective, and can be transferred to the existing one-stage detection models. Experimental results show that based on Gelan-S, YOLOv5-s, and YOLOv7-tiny, the detection accuracy of the baseline model is improved by 4.0%, 4.2%, and 4.5%, respectively. The significant performance improvement demonstrates the effectiveness and application prospects of our approach.