Abstract: High-performance semantic segmentation algorithms cannot quickly perceive road conditions due to their high latency. This paper proposes a dual-path network model based on attention mechanism. The network model uses a lightweight local contour information extraction module and a semantic information extraction module to replace the complex encoder structure. Aiming at the characteristics of feature maps under different paths, feature optimization modules are designed based on self-attention and channel attention mechanisms. This algorithm effectively improves the ability of lightweight network structures to express detailed features. The designed semantic segmentation network processes images at a speed of 25 fps while maintaining an average cross-to-parallel ratio of 73.9%. The physical verification shows that the algorithm has real-time performance and high value in certain practical application.

Key words: semantic segmentation, bilateral convolutional neural network, autopilot, embedded platform