Enhanced Curve Perception in Lane Detection via Adaptive Guided Techniques

Lane detection is the prominent perception task for both self-driving and advanced driver assistant system. Recently, lane detection methods based on anchors have received increasing attention, but fixed shape anchors are difficult to model complex lane line shapes. To solve this problem, we propose an Enhanced Curve Perception Network (ECPNet) to detect the shapes of lane lines more flexibly and accurately. Specifically, we propose a Layer-by-layer Context Fusion Module (LCFM) to fully utilize both high-level and low-level features in lane detection, which can simultaneously obtain the global spatial structure relations and the local accurate positioning details. The module includes short hop connections within the initial layer and between feature layers of different scales. More importantly, we propose a novel Structural Correction Prediction Module (SCPM), which enhances the model's detection capability on curved structured lanes by assisting model selection anchor classification mode through adaptive guided techniques. In addition, we design a branch network and construct a Cross-channel Attention Mechanism (CAM) to realize self-information attention, which further enhances the detection accuracy. Experiments on the two most representative datasets demonstrate that the proposed method achieves state-of-the-art performance in a variety of environments, especially in curvy lanes.