Research on gesture recognition method for train driver

Executing prescribed gestures according to the standard of train operation is a critical procedure of train driver. The driving state and operation quality of train drivers can be effectively evaluated by detecting gestures of drivers, which ensures the safety of train operation. The traditional manual inspection method is inefficient. The algorithm of existing gesture recognition has the problems of considerable number of model parameters, low accuracy and slow speed of detection. With the development of intelligent railways, using deep learning methods to build a lightweight, efficient, and high-precision train driver gesture recognition model has gradually become a demand for industry development. In response to the above demand, a train driver gesture recognition model based on improved YOLOv5 was proposed. First, the lightweight convolution named PConv was introduced to optimize the C3 module for reducing parameters and calculating the amount of the network, and to improve the efficiency of model detection. Meanwhile, the Convolutional Block Attention Module was added to the interference of irrelevant information and enhanced the feature extraction ability. Second, bidirectional feature pyramid network (BiFPN) was introduced to replace the Path Aggregation Network (PANet) in the neck layer, which enhanced the fusion ability of multi-scale features and improved the detection ability of small targets by adding a small target detection layer. Finally, the bounding box loss of the model selected Focal-EIoU, which speeds up the convergence rate of the model and improved the accuracy of gesture positioning. The experimental results show that the mean average precision (mAP@0.5) of the improved model reached 97.7%, and the average detection time of improved model was 23.2 ms. As compared to YOLOv5, the amount of calculation was reduced by 23.1%, the mean average precision of the improved model was improved by 0.6 percentage points, and the average detection time was reduced by 7.1 ms. The model can effectively improve the detection efficiency and accuracy while reducing the number of model parameters, which can provide new ideas for train driver gesture recognition. © 2024, Central South University Press. All rights reserved.