An improved deep learning algorithm for obstacle detection in complex rail transit environments

Onboard obstacle detection is a vital technology to ensure the safety of smart trains. Traditional object detection algorithms have poor detection accuracy for small obstacles and are challenging to cope with complex changes in rail transit environments. To address the above problems, a robust deep learning algorithm RFA-Net is proposed. The improved focal and global knowledge distillation is used to improve the feature extraction capability without additional computational burden. The side-aware boundary localization subnet is used to enhance the detection head to obtain high-quality detection bounding boxes. Experimental results based on a fully annotated rail transit dataset show that the RFA-Net achieves a detection accuracy of 92.7%mAP while keeping the model lightweight. Compared with the baseline model, the overall detection accuracy is improved by 5.8%, and the detection accuracy of small obstacles is improved by 7.5%. The RFA-Net has a TPR of 0.893 and a FPR of 0.016, achieving a good balance in identifying dangers and reducing false alarms. In various complex rail transit environments, the RFA-Net is able to detect obstacles stably, and the detection accuracy for small obstacles is improved by 18.6%. In addition, experimental results based on the KITTI dataset show that the proposed algorithm has good generalization ability in traffic scenes.