Abstract: In order to detect the safety risk source of highway driving environment and provide the basis for intelligent control of highway risk source and real-time safety risk assessment in random uncertain scenarios, a safety risk source identification algorithm of highway driving environment based on deep convolutional neural network model was studied. By improving the output layer activation function and loss function of MobileNetV3, the number of risk source categories output by the network was increased from one to multiple, which solved the problem of identifying multiple risk sources in the same image. The spatial attention mechanism was used to enhance the feature extraction ability of the MobileNetV3 network, which solved the problem that the MobileNetV3 channel attention mechanism could not pay attention to the feature information of the risk source inside the channel and improved the model recognition accuracy. Through the channel pruning method to remove redundant expansion channel, the number of network parameters was reduced, and the prediction speed of the model was improved. The experiment results show that the proposed method can effectively identify the driving environment safety risk source, with a detection rate of 0.829, an average classification accuracy of 0.833, and a real-time detection effect.

Key words: vehicle engineering; traffic safety; safety risk sources of driving environment; multi-label image classification algorithm; MobileNet

摘要： 为了检测公路行车环境安全风险源，为公路风险源智能化管控和随机不确定场景下实时安全风险评估提供依据，研究了基于深度卷积神经网络模型的公路行车环境安全风险源识别算法。通过改进MobileNetV3的输出层激活函数和损失函数，将网络输出的风险源类别数量由一个变为多个，解决了同一图像中存在多种风险源的识别问题。利用空间注意力机制加强MobileNetV3网络的特征提取能力，解决了MobileNetV3通道注意力机制无法关注到通道内部风险源特征信息的问题，提升了模型识别准确率。通过通道剪枝方法去除冗余扩张通道，减少网络参数量，提升了模型预测速度。实验结果表明：该方法能够有效识别行车环境安全风险源，检测率达0.829，平均分类准确率达0.833，且具备实时检测效果。

关键词: 车辆工程；交通安全；行车环境安全风险源；多标签图像分类算法；MobileNet