A multi-camera detection algorithm for small obstacle characteristics in the rail under complex and variable illumination conditions

In addressing the challenges of false positives and missed detections of small-scale obstacles within low-illumination orbital environments, a multi-scale detection algorithm MTD R-CNN based on multi-camera is proposed. The proposed model contains three stages. In stage 1, the LCSwin Transformer is proposed to complete aggregating detailed features and global relationships. In stage 2, the SAFPN is proposed to realize hierarchical feature interaction at different scales. In stage 3, dynamic instance interactive head multiplexing and multiple loss sets are used to obtain more decadent detection boxes. The test results of the track scene under different illumination conditions show that: (1) The accuracy of the MTD R-CNN is 95.2 %, surpassing the performance of existing models; (2) The detection accuracy of small obstacles is improved by 3.7 % -26.4 %, thereby highlighting the model’s superior perceptual capabilities for detecting such obstacles; (3) The operation speed of the model is 36.63 ms to meet the real-time processing criteria. In summary, the model effectively improves the detection performance of small obstacles under low-light light conditions and has been applied in Nanning Metro Line 5.