Abstract: lightweight neural network algorithm MobileNetV3-YOLOv5s based on hybrid attention mechanism was proposed to address the complex network structure, large parameter quantity, and high hardware requirements of current vehicle identification neural network algorithms. Firstly, the bneck module of MobileNetV3 was used to replace the backbone network of YOLOv5s. Secondly, the large convolutional kernel was replaced with a small convolutional kernel, and meanwhile, the SPPF algorithm was improved with a feature fusion method that required less computation. Finally, SENet and CAM algorithms were integrated into the backbone network to form a mixed attention module, which increased the network's weight on important regions. The experimental results show that on the UA-DETRAC dataset, the parameter amount of the proposed algorithm is reduced by 82.6% compared to YOLOv5s, only 2.34 MB, with an average identification rate of 98.2%. On the Nvidia Jetson AGX NX, the detection speed reaches 31 frames per second, with a speed increase of 10.7%. The proposed algorithm can be better deployed on edge devices and meet the requirements of autonomous driving.

Key words: vehicle engineering; vehicle identification; attention mechanism; SPPF; MobileNetV3; YOLOv5s

摘要： 针对目前车辆识别神经网络算法网络结构复杂、参数量大、对硬件要求高的问题，提出一种基于混合注意机制的轻量级神经网络算法MobileNetV3-YOLOv5s。首先，采用MobileNetV3的bneck模块替换YOLOv5s的主干网络；其次，将其中的大卷积核替换为小卷积核，同时用计算量更小的特征融合方法改进SPPF算法；最后，在主干网络中融合了SENet和空间注意力机制，组成混合注意力模块，提高网络对重要区域的权重。试验结果表明：在UA-DETRAC数据集上，所提出算法的参数量相比于YOLOv5s减小了82.6%，仅为2.34 MB，平均识别率为98.2%，在Nvidia jetson AGX NX上检测速度达到31帧/s，速度提高10.7%，可以更好地部署在边缘设备上，满足自动驾驶的要求。

关键词: 车辆工程；车辆识别；注意力机制；SPPF；MobileNetV3；YOLOv5s