A Lightweight Object Detection Network for Vehicle Environment Perception Based on Improved YOLOv5s

Vehicle environment perception techniques are crucial for ensuring safe driving conditions on roadways. Vision-based systems can achieve high accuracy and real-time processing in detecting and recognizing objects, which makes them versatile across different vehicle types and driving conditions, enhancing their suitability for various applications in intelligent transportation systems. However, existing object detection methods often suffer from a high number of parameters and significant computational demands, which limit their performance when deployed on vehicle hardware. To tackle these challenges, a lightweight target detection network based on YOLOv5s framework is designed for vehicle environment perception. First, the proposed network integrates depthwise separable convolution to reduce network complexity and redesigns the C3 module based on one-shot aggregation to streamline the overall network structure. Then, CA mechanism is also introduced to learn each channel feature adaptively and focuse on key feature areas to improve the detection performance of the network. Third, MPDloU is employed as a bounding box loss function to improve the robust and reliable object detection performance, particularly in challenging scenarios with varied object sizes and aspect ratios. Experimental results on the KITTI public dataset demonstrate that the improved model achieves 34.7%, 35.4% and 33.6% respectively in parameters, GFLOPs and model size.