Adaptive Lane Angle Prior and Feature Enhancement for Robust Lane Detection

Authors:

Abstract
Lane detection is a crucial component of intelligent driving perception. Among the various lane detection methods, the line anchor-based approach has demonstrated state-of-the-art (SOTA) performance. However, this method still faces challenges when dealing with blurred or occluded lane markings. To enhance the quality of line anchors, we propose an Adaptive Lane Angle Prior, which enables localized refinement of curved lane geometries, thereby more accurately capturing the variations in lane lines. By dynamically adjusting the angle parameters based on angle variance threshold, our method significantly improves lane detection precision in scenarios with curve. To address the common occlusion issues in lane detection, we design a Feature Enhancement Module that integrates both Non-local Blocks and Large Kernel Attention mechanism. The Non-local module enhances the global consistency of features through global information interaction. Furthermore, the Large Kernel Attention mechanism captures large-scale contextual information through large-sized convolutional kernels. To more accurately capture lane line variations, we integrate an Angle Loss together with adaptive angle prior to enhance geometric constraints within the model. Specifically, our method achieves an F1 score of 97.89% on the TuSimple dataset and 80.27% on the CULane dataset. These experimental results demonstrate the effectiveness of our approach on both benchmarks.
Funding Information
  • National Natural Science Foundation of China (62276167)