Lightweight intelligent vehicle target detection algorithm based on Yolov4

Abstract

Aiming at the complex and changeable driving scenarios of intelligent vehicles and the need to quickly and accurately identify obstacles, an improved YOLOV4 algorithm is proposed. To limit the number of neural network parameters, the CSP-darknet53 backbone of the original YOLOV4 was replaced with the Ghostnet backbone. In addition, to improve the neural network's accuracy, a lightweight attention mechanism ECA is added to the three effective feature layers generated by the backbone using residual block connections. Experiments have shown that the improved YOLOV4 has a 2.8% increase in mAP compared to the original YOLOV4. Without changing the accuracy, The network model's memory size is lowered by 39%, as well as a 50% improvement in detecting speed. Therefore, the improved YOLOV4 accuracy and real-time performance are better than the original network detection, providing a strong guarantee for intelligent vehicle obstacle avoidance.