Abstract: In order to achieve real-time detection of vehicle object in tunnel environment, an improved YOLOv4 object detection algorithm was proposed. First, the Mobilenetv3 network was used to replace the backbone network of the YOLOv4 algorithm which can reduce the amount of algorithm parameters and improve the detection speed; using a hybrid attention mechanism in the Mobilenetv3 network by combining channel and space dimensions to further improve detection accuracy of the algorithm; a lightweight PANet was proposed which used Ghost convolution operation to replace the ordinary convolution operation so to further reduce the amount of algorithm parameters. In order to verify the effectiveness of the designed model, the urban tunnel vehicle dataset is manually labeled. Experiments show that the object detection model designed in this paper has an average accuracy rate of 92.08% on the tunnel vehicle dataset which is similar to the original YOLOv4 network model. The detection accuracy are similar, but the speed is increased by 45.7%.

Key words: Object detection; Convolutional neural network; Tunnel environment; Intelligent transportation system

摘要： 为了实现隧道环境中车辆目标实时检测的目的，提出一种改进的 YOLOv4 目标检测算 法。首先使用 Mobilenetv3 网络替换 YOLOv4 算法的主干网络，减少算法参数量，提高检测 速度；在 Mobilenetv3 网络中引入混合注意力机制，结合通道与空间维度进一步提高算法的检 测精度；提出一种轻量化 PANet 结构，使用 Ghost 卷积操作替换普通卷积操作，进一步减少 算法参数量。为验证所设计模型的有效性，手工标注了城市隧道车辆数据集。实验表明，文 章设计的目标检测模型在隧道车辆数据集上平均精确率达 92.08%，与采用 YOLOv4 网络的检 测效果相近，速度却提高了 45.7%。

关键词: 目标检测；卷积神经网络；隧道环境；智能交通系统