基于MobileNetV3与ST-SRU的危险驾驶姿态识别

针对危险驾驶行为引起的交通安全事故频发的现状,提出一种基于MobileNetV3和ST-SRU的危险驾驶姿态识别系统.首先,修改MobileNetV3的网络结构使其适用于人体姿态估计任务,输出关节点的热力图和偏移量图,用来估计J个关节点的二维坐标位置;其次,定义ST-SRU骨架动作识别算法,利用动作的骨架序列数据对动作进行分类.实验结果表明:MobileNetV3姿态估计算法在自建的AI Challenger上肢姿态数据集上测得PCP值(percentage correct parts)达到95.6%,测试1 000次用时仅为5.03 s;利用自建的危险驾驶行为数据集将训练好的姿态估计和动作识别模型移植到嵌入式平台,实现了实时的危险驾驶姿态识别系统.

Dangerous driving pose recognition based on MobileNetV3 and ST-SRU

In the face of frequent traffic accidents caused by dangerous driving behaviors, this paper proposes a dangerous driving pose recognition system based on MobileNetV3 and ST-SRU. Firstly, the network structure of MobileNetV3 is modified to be used for human pose estimation, and the heatmaps and offsets of joint points are output to estimate the 2D coordinate positions of J joint points. Then, the ST-SRU skeleton action recognition algorithm is defined, and the actions are classified by using skeleton sequence data. The experimental results show that the PCP(percentage correct parts) of MobileNetV3 pose estimation algorithm is 95.6% on the self-built AI Challenger upper limb attitude dataset, and the time of 1000 tests is only 5.03 seconds. By using the self-built dangerous driving behavior dataset, the trained pose estimation and action recognition model is transplanted to the embedded platform, and the real-time dangerous driving pose recognition system is realized.