Abstract: Drivable area detection is a critical task for autonomous vehicle navigation, as it involves pixel-level classification of the environment into drivable and non-drivable regions. This fine-grained segmentation enables the vehicle to distinguish safe road surfaces, providing essential spatial understanding for real-time path planning, obstacle avoidance, and safe decision-making in complex traffic scenarios. However, conventional approaches often adopt an oversimplified definition of drivable areas, typically considering entire road surfaces as uniformly navigable. This assumption fails to account for real-world traffic constraints imposed by lane markings, thereby impacting the reliability of decision-making. To address this problem, we redefine pixel-level drivability based on traffic rules to interpret the scene and construct the lane line constrained drivable area detection (LCDAD) dataset. Based on this dataset, we conduct comprehensive evaluations of existing state-of-the-art methods. Considering the limitations of existing road segmentation methods, we propose a cross-attention fusion architecture to further enhance road scene understanding. Within this architecture, a self-attention mechanism is applied across the feature dimension rather than the spatial dimension, calculating the cross-covariance across feature channels to generate attention maps, making it more suitable for high-resolution drivable area detection. Extensive experiments on three public datasets demonstrate the effectiveness of the proposed framework, while detailed ablation studies confirm the superiority of the proposed fusion structure.