HiFi-CaNet: A High-Level-Feature-Guided Cascade Network for Lane Detection

Abstract

With the rapid development of autonomous driving technology, lane detection is crucial for enhancing road safety and the reliability of driver assistance systems. Modern methods that combine advanced image processing with deep learning achieve high precision in recognizing lane lines under various conditions, providing reliable support for autonomous vehicles. However, existing algorithms face significant challenges in complex environments, such as drastic lighting changes, blurry road markings, and occlusions, leading to performance declines and misdetections. To address these issues, we propose a high-level-feature-guided cascaded network (HiFi-CaNet). This network strategically integrates high-level semantic information with low-level visual features to enhance accuracy and robustness. First, a channel–space cooperative feature-enhancement module is introduced to improve global feature extraction to combine spatial and channel representations, reducing semantic information loss. Then, a high-level-guided feature pyramid network is designed to filter semantic information in low-level features with weights generated from high-level features for effective multi-scale feature fusion. Additionally, a distance lane intersection-over-union loss is presented to improve lane localization and curved-lane detection performance. Experimental results on the CULane dataset demonstrate that HiFi-CaNet achieves an F1@50 score of 80.39, compared with the baseline model’s 80.20, showing further improvement despite the baseline achieving excellent performance. HiFi-CaNet shows improvements of 0.36%, 0.9%, 0.32%, and 2.4% in scenes with normal roads, wireless roads, ground traffic markings with arrows, and curves, respectively, proving its effectiveness in complex traffic environments.