Abstract: In recent years, with the rapid development of deep learning technology, autonomous driving technology has received increasing attention in the field of intelligent transportation. To enhance the real-time performance of vehicle detection and reduce the requirements for memory and computational resources, this paper proposes a lightweight improvement to the YOLOv5s algorithm. Firstly, the PP-LCNet lightweight neural network is introduced to replace the backbone network of YOLOv5s, resulting in the construction of the lightweight model YOLOv5s-PP. Secondly, the loss function is improved by adopting EIOU Loss to replace the original CIOU Loss, which improves the efficiency and accuracy of bounding box regression. Lastly, the SimOTA strategy is introduced for dynamic label assignment to adapt to complex and diverse road environments. Experimental results demonstrate that, the improved YOLOv5s-PPES algorithm exhibits significant improvements in detection speed and parameter count, with overall superior performance.

Key words: vehicle detection; lightweight; YOLOv5s; PP-LCNet

摘要： 近年来，随着深度学习技术的快速发展，自动驾驶技术在智能交通领域受到越来越多 的关注。为了提高车辆检测的实时性、降低对内存和算力资源的要求，文章针对 YOLOv5s 算法进行了轻量化改进。首先，引入 PP-LCNet 轻量化神经网络替换 YOLOv5s 的骨干网络， 构建了轻量化模型 YOLOv5s-PP；其次，改进了损失函数，采用 EIOU Loss 替代原有的 CIOU Loss，提高了边框回归的效率和精度；最后，引入 SimOTA 策略进行动态标签匹配，以适应 复杂多变的道路环境。实验结果表明，改进后的 YOLOv5s-PPES 算法在检测速度和参数量上 都有显著提升，整体性能更加优越。

关键词: 车辆检测；轻量化；YOLOv5s；PP-LCNet