A Lightweight Sementic Segmentation Model for Metro Tunnel Scene Based on Vehicle Front Camera
摘要
Metro intrusion seriously affects the safety of metro operation. Semantic segmentation is the main content of the research on metro intrusion. Tunnels account for a large proportion of metro lines, by processing the images captured by the vehicle front camera, the state of the elements in the tunnel can be obtained, among which the perception and understanding of the whole scene is an important part of the image processing. For on-board cameras, we propose a lightweight semantic segmentation algorithm for metro to satisfy the real-time requirements. Finally, for a 2048 × 1024 input, the algorithm achieves 75.21% MIoU on the metro tunnel data set with speed of 54FPS (Frames Per Second) on NVIDIA 3090 Ti card.