Improved Mask R-CNN for obstacle detection of rail transit

Accurate identification of obstacles shows great significance to improve the safety of automatic operation trains. The ME Mask R-CNN is proposed to improve the accuracy of active identification. The SSwin-Le Transformer is used as the feature extraction network and the ME-PAPN is used as the feature fusion network. A variety of multi-scale enhancement methods are integrated to improve the detection ability of small target objects. PrIme sample attention is used as the sampling method, the anchor boxes size and ratio suitable for the characteristics of train obstacles are adopted. The train obstacle dataset is based on a variety of test scenarios such as Nanning Metro Line 1 test line, tunnel line and night test. The test results show that ME Mask R-CNN achieves 91.3 % mAP with an average detection time of 4.2 FPS, which is 11.1 % higher than that of Mask R-CNN.