Abstract: To address the issues of wide lane line distribution range, sparse pixel coverage, and difficulty in feature extraction, this essay constructs a residual lane line detection network based on multi-scale feature fusion. This network is built upon a residual bilateral network and incorporates a bilateral feature aggregation module. It leverages the contextual information from the semantic branch to guide the feature responses of the detail branch within the same stage, thereby integrating information from both branches. Different stages operate at varying scales, and a multi-scale adaptive feature alignment fusion module is used to construct a sampling pre-and post-offset vector index table, reducing detail information loss caused by simple sampling. Additionally, a spatial attention mechanism is introduced to enhance the model's ability to capture long-distance features. Experimental results show that the proposed method performs well across three public datasets, achieving an accuracy of 77.89% on the CULane dataset, which is 2% higher than the current mainstream algorithms.

Key words: lane line detection, bilateral segmentation network, multi-scale, attention, end to end