On-board optical visual perception and intelligent control of electric locomotive based on Transformer

In the visual perception system, sudden light transitions and low-light environments may reduce image quality and the accuracy of feature extraction. However, traditional methods are limited by local receptive fields and hierarchical feature extraction mechanisms, making it difficult to adapt to complex lighting scenes. In order to meet the real-time perception requirements of electric locomotives in complex lighting environments, an improved optical vision perception technology based on Transformer is studied and proposed. This model introduces an alternating architecture of window multi-head self-attention and sliding window multihead self-attention, combined with the cross-scale feature fusion module, to achieve efficient extraction and dynamic fusion of multi-scale features. The results show that the average accuracy of the proposed model reaches 92.63%, which is significantly better than the traditional method. The Swin Transformer backbone further improves performance to an average accuracy of 90.28%. With a lightweight design that ensures real-time processing, the model achieves 45 frames per second at 512×512 resolution on an embedded platform, meeting the stringent latency requirements of onboard systems. The improved Transformer model proposed in this study can be processed in real time, providing a high-precision and low-latency solution for rail transit intelligence.