融合目标检测和人体关键点检测的铁路司机行为识别

Railway driver behavior recognition based on fusion object detection and person keypoints detection

摘要:

随着我国经济的快速发展，铁路运输在交通运输的地位愈为重要，在传统人工监管无力应对铁路司机安全监督的情况下，使用机器实现自动实时司机行为识别早已成为了一项极有意义的工作。为实现随车部署、实时进行铁路司机行为识别的目的，基于目标框检测算法实现目标检测和关键点检测的融合，搭建了一种可以同时检测司机人体关键点和手机的神经网络。经过网络运行输出人体姿态后，通过分析人体各关节角度和人体关键点与手机目标的位置关系等后处理对六类司机行为进行了分类识别，并通过TensorRT框架对模型进行了模型推理速度的加速和体积上的压缩。实验表明，该模型在嵌入式设备TX2上推理速度为25ms，可以达到较好检测效果下实时运行的目标。实现了实时进行铁路司机行为识别的目的。

Abstract:

With the rapid development of China's economy, the role of railway transportation in transportation becomes more important. In the case that traditional manual supervision is unable to cope with the safety supervision of railway drivers, using machines to realize automatic real-time driver behavior recognition has already become a very meaningful task. In order to realize the real-time railway driver behavior recognition on the embedded device, a neural network that can simultaneously detect key points of the driver's human body and mobile phones is constructed based on object detection and person keypoints detection. Six types of driver behavior are identified by post processing of analyze the relationship between the joint angles of the human body and the key points of the human and the target of the mobile phone. And the model was accelerated and compressed for operation on em-bedded devices through TensorRT framework. Experiments show that the inference time of model is 25ms on the embedded device TX2, which can achieve the goal of better accuracy and real-time op-eration. The purpose of real-time identification of railway driver behavior was achieved.