Evaluation of Alternative Advanced Convolutional Neural Networks for Train Driver Action Recognition
Junming Miao, Jinjin Jiang, Qili Chen, Miao Wang, Guangyuan Pan · 2021
Action accuracy of train drivers is closely related to the train safety. However, at present, the video of cabs of train drivers is difficult to be monitored in real time, along with poor alarm accuracy, long response time and difficulties in analysis of video data. Therefore, action recognition in video of the train driver is studied in this paper. First, based on the clips of collected video, the emergence of key actions in the whole video can be distinguished, and image enhancement and segmentation technology can be used to extract the images of key actions. Second, three improved convolutional neural networks (C3D, R(2+1)D and R3D) are compared to realize accurate recognition of actions. Third, a comparative experiment based on pre-training models and non-pre-training models is designed to analyze the application effects of different algorithms in this problem. The results show that the method proposed can improve the accuracy of action recognition, which effectively reduce the accidents of train driving.