Post-Processing Method for Lane Detection Based on Prior Knowledge and Distance Penalty

Applying post-processing techniques to extract lane instances from binary segmentation images is an important step in lane detection based on semantic segmentation. This paper proposes a lane instance extraction algorithm based on prior knowledge and distance penalty modules, improving the random sample consensus algorithm. The prior knowledge module uses existing lane prior knowledge to set filtering conditions in the algorithm iteration process, enhancing the robustness of instance extraction. The distance penalty module applies an evaluation penalty based on the distance of lane feature points, solving the deviation at the near end of the lane model caused by a sudden change in the number of feature points due to perspective transformation and improving the accuracy of instance extraction. The experimental results on a large amount of testing data from various traffic scenarios demonstrate that the proposed algorithm can be applied to multi-lane detection post-processing tasks containing any number of lanes, with an instance extraction accuracy rate of 98.53%. It has good instance extraction capabilities for lanes with any curvature, with an average extraction error as low as 16.85 pixels, and stable extraction performance for various types of lanes, such as solid and dashed lines.