Improved YOLOv5s Lightweight Vehicle Axle Recognition Model

The real-time detection of axle types plays a key role in off-site truck overload management. Addressing the challenge of balancing recognition accuracy and real-time performance due to the limited resources, a lightweight vehicle axle type recognition model improved by YOLOv5 architecture was proposed in this paper. Firstly, Ghost convolution is employed instead of conventional convolution to reduce the number of model parameters. Secondly, the coordinate attention mechanism is introduced to enhance the capability of detail extraction by employing a specialized convolution kernel that specifically targets tire features. Finally, The SIoU regression loss function is employed to address angle distortion specifically, and improve the model's convergence speed. The experimental results show that, compared to the traditional YOLOv5 model, the enhanced model achieves a 1.5% increase in average accuracy, a 38% reduction in model size, and a 25% boost in calculation speed. A single detection only takes 15.4 ms, which can meet the real-time requirements of vehicle axle detection.