• Propose YOLOv8n-3SE-PD, a lightweight traffic-object detector for resource-constrained edge environments. • Enhance small-object recognition using a three-layer SE channel-attention mechanism. • Reduce parameters and computation by over 35% through structured channel pruning with knowledge distillation. • Achieve a highly efficient model (1.92M parameters, 3.1 ms latency) suitable for real-time intelligent-transportation applications. With the rapid development of intelligent connected vehicle technology and vehicle-road cooperative systems, it is a key requirement for onboard perception systems to realize high-precision and low-latency object detection under the constraints of edge computing. However, high computational and memory consumptions of deep learning-based detection models indicate great challenges for the deployment on embedded vehicular hardware. In this paper, a lightweight model for small-object detection in smart vehicles, namely YOLOv8n-3SE-PD, is proposed. The model improves feature saliency by introducing a SE attention mechanism, compresses redundant network parameters with a structured pruning strategy, and applies knowledge distillation to transfer semantic information from a complex teacher model to the lightweight student model. Results of experiments with traffic monitoring datasets show that the optimized model is able to reduce parameters by more than 35% in the electric two-wheeler helmet detection task, while keeping accuracy comparable with the baseline model and achieving an inference latency of just 3.1 ms per frame. The results confirm that YOLOv8n-3SE-PD achieves a well-balanced trade-off between detection accuracy and computational efficiency, thus providing an effective and show promising trends solution for real-time visual perception and safety monitoring in intelligent vehicles operating under edge-computing environments.
Wang et al. (Sun,) studied this question.