Research On Railway Driver Behavior Recognition Technology Based On Deep Convolutional Neural Network

With the rapid and comprehensive development of China’s economy and the improvement of people’s living standards,China’s demand for transportation is increasing.With the rapid development of railway transportation productivity,the safety of railway traffic has been paid more and more attention,and the safety and security technology of railway traffic has been put forward higher requirements.In order to meet the higher requirements of locomotive operation safety,realize real-time reminders of driver behaviors,and reduce the waste of human resources in the existing supervision system,this thesis takes the identification and classification of driving behaviors of railway locomotive drivers as research objects,a realtime driver behavior analysis system which can be deployed with the vehicle is designed and implemented by using deep convolutional neural network.The main content and work of this thesis are summarized as follows:First,a deep convolutional neural network is designed to perform human pose estimation and object detection simultaneously.In order to deal with the driver’s use of mobile phones in behavior classification,a new neural network model was designed by integrating human body key points recognition and object detection through object frame recognition.The dataset of railway driver behavior recognition was collected and labeled.The model is trained on this dataset.The evaluation index of the model is designed,and the model recognition effect was tested on the test dataset.Secondly,building a multi-scale pyramid structure enhances the model’s ability to recognize small objects,and the precision tuning of the model is realized.Based on model pruning and half-precision acceleration,the volume and inference speed of model were optimized.The accuracy of the model has been greatly improved,the model volume is reduced to 16.2MB,and the inference speed in the embedded device Jetson Nano is 63 ms,which meets the purpose of real-time operation.Finally,a post-processing method for driver behavior classification is proposed based on the results of the key points of human body and the mobile phone object frame output from the model.The effectiveness of the driver behavior classification method is verified by experiments.The driver behavior recognition system is composed of pre-processing,neural network model and post-processing method.Through the design of multi-process pipeline operation to achieve the overall acceleration of the system.The system’s operating speed in the embedded device Jetson Nano was 14 FPS,which achieved the purpose of real-time operation.