AE-Net: A High Accuracy and Efficient Network for Railway Obstacle Detection Based on Convolution and Transformer

The incursion of railway obstacles poses a serious risk to train operations, and numerous accidents occur during train shunting. However, existing algorithms still struggle with finding a compromise between detection accuracy and speed during train movement. Moreover, their accuracy and robustness are inadequate, specifically when handling small objects in complicated railway scenarios. To overcome these issues, this paper proposes an efficient network using convolution and transformer (AE-Net) for performing accurate and real-time detection of railway obstacles to ensure driving safety. First, the Enhanced and Lightweight Transformer Module (ETM) is constructed to strengthen the model’s global modeling ability. Then, the Lightweight Feature Integration Module (LIM) is presented to integrate multi-branch feature information and reduce model complexity. Finally, the Reinforced Multi-Scale Feature Fusion Module (RFM) is utilized to enhance the multi-scale object detection capability, especially for small obstacles. The presented algorithm realizes 95.29% mAP and 145 FPS on the railway dataset, which is superior to YOLOv5s. In addition, the experiment on MS COCO further shows that AE-Net can perform a considerably better detection than current state-of-the-art models. Hence, it is practicable to employ AE-Net in actual railway and further more complex multi-target scenarios.