SFEP-YOLO: A Track Obstacle Detection Model for Autonomous Electric Locomotives in Underground Mine

The stable and safe operation of autonomous mining electric locomotives (MELs) requires the support of advanced track obstacle detection technology. Due to the problems of slow detection speed and low detection accuracy in traditional sensor based detection methods, Swin Transformer, Focal EIOU loss function, Efficient Channel Attention (ECA) attention mechanism, and four prediction layers were introduced into the traditional YOLO model to develop a high-precision obstacles detection model for MELs in this work, which is called SFEP-YOLO. In terms of network structure, the Swin Transformer model was introduced in the Backbone to design a Multi-scale Channel Swin Transformer(MCST) module, which achieves windowed extraction and interactive calculation of image features; the Focal EIOU loss function was introduced to improve recognition ability for difficult-to-classify samples. In the Neck layer, the ECA attention mechanism was incorporated to effectively capture the interdependence between channels, in order to achieve rapid localization of important feature information. In the Head layer, the four prediction layer structure enhances the fitting ability of the model's prediction box and true box, providing detection accuracy for small obstacles, maintaining high performance detection accuracy. A rich dataset of track obstacles has been established from real complex mining transportation scenes. The transfer learning method was adopted to train the backbone on the VisDrone2019 dataset, and the obtained weights were set to be the pre-training parameters for the SFEP-YOLO model. The proposed SFEP-YOLO model was tested both on the VisDrone2019 dataset and the MELs obstacles dataset. Experimental results show that the accuracy at mAP@50 is 93.8%, the Precision is 93.7%, the Recall is 91.2%, and the detection speed is 117FPS. Compared with other existing detection models, the SFEP-YOLO model achieves a good balance between speed and accuracy, which ensures that the electric locomotive can detect all kinds of obstacles while detecting in real time.