Lane Detection Method Based on Instance Segmentation of Bi-ATT Network

In order to achieve fast and accurate lane detection without limitations, this paper proposes an instance segmentation lane detection method based on an improved BiseNetV2 network. Firstly, we propose a Bilateral Attention Fusion Network (Bi-Att) that incorporates a lightweight multi-scale feature attention fusion module into the semantic branch of the dual branch network. This module effectively focuses on and aggregates global context information, improving the overall performance. Additionally, a cross-layer fusion module is introduced in the binary segmentation decoding process to recover spatial details lost in high-level feature maps, thereby enhancing semantic segmentation accuracy. Secondly, the Bi-Att network is utilized as the backbone of the LaneNet, a well-known end-to-end lane detection network. Finally, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering method is employed for detecting unlimited lanes. Experimental results demonstrate that the proposed method achieves an impressive inference speed of 108.66 FPS, with a MIOU index of 52.9% on the Tusimple dataset, and an accuracy of 94.1%. Comparative analysis against existing detection methods confirms the superior overall performance of the proposed approach.