Efficient Train Driver Action Recognition Method Based on Improved Temporal Difference Network

The railway transportation system is a complex system composed of “human” and “machine” elements. Among the “human” components, whether train drivers operate according to regulations and procedures is a key factor influencing railway transportation safety. In response to the weak timeliness, high cost, and time-consuming nature of the existing train driver action recognition approaches in practice, this paper proposes an efficient method based on partial convolution and spatiotemporal attention flow for recognizing train driver action, requiring fewer hardware resources and achieving rapid and accurate results. Firstly, we utilize a partial convolutional (PConv) to integrate temporal difference network (TDN), constructing an extremely computationally efficient action recognition model. Subsequently, to address the issue of low accuracy stemming from the model's lack of focus on action information, we devise a spatiotemporal attention flow module to enhance the model's attention towards action regions, thereby improving accuracy. Finally, to further improve the accuracy of the model, the coordinate attention mechanism (CA) is introduced to capture direction-aware and position-sensitive information. Experimental results demonstrate that this detection method achieves high-speed and high-accuracy detection of train driver actions.