Abstract:

In current research on lane detection, existing algorithms can efficiently detect lane lines under good lighting conditions. However, lane detection in low light still faces the challenge of a high false negative rate. A detection algorithm called Instance Association Net(IANet) is proposed to address this issue by utilizing the structural relationships between lane lines, which is helpful for low light conditions. The algorithm first generates unique masks for different lane lines using features at the starting points of the lane lines and a global feature map, achieving instance-level feature separation of the lane lines. It employs an instance-level attention mechanism to correlate the separated features, facilitating effective information exchange between instances. Before the correlation, absolute position encoding is introduced to enhance the model's focus on the positional correlation of the lane lines. The network achieves precise lane detection by locating key points on the lane lines and calculating the offset. Experimental comparisons with existing methods on the CULane dataset show that IANet achieves an overall score of 75.7% and a score of 71.9% in night scenes, which is higher than other algorithms. It demonstrates good robustness in various lighting conditions and significantly reduces the false negative rate of lane detection in low-light conditions due to the proposed instance feature association.

Key words: autonomous driving, dark lane detection, instance association, position encoding, attention mechanism