Research on lane line detection method based on improved feature fusion network

Addressing the current issues in lane detection algorithms related to detection accuracy and speed, this study proposes a lane detection method based on an improved feature fusion network. First, the residual network is replaced with RepVGG network, which applies structural reparameterization to reduce the memory occupation and improve the inference speed. Secondly, BiFPN network is introduced in the feature extraction part, which utilizes bidirectional paths to promote the bidirectional flow of feature information between different scales, effectively realizing the exchange of semantic information of lane lines. Finally, to address the problem of insufficient lane line information extraction in low light at night, the feature fusion network is improved, and the self-attention mechanism is introduced to improve the detection accuracy of lane lines in the night environment. Experimental research conducted on the CULane dataset indicates that the improved network has achieved advancements in both accuracy and real time performance. Compared to the state-of-the-art algorithms in the same category, our method demonstrates a remarkable 2.6-fold increase in inference speed, alongside an enhancement in the F1 score from 80.1% to 80.9%. Notably, under nighttime conditions, the F1 score experiences a specific improvement of 0.6%.