Towards Better 3D Visual Lane Detection in Traffic Scenes via Leveraging Object Semantics

3D lane detection plays a significant role in autonomous driving applications. However, existing methods fail to fully consider the influence of other traffic participants on lane estimation, resulting in reduced prediction accuracy. To address this limitation, we propose a novel method called Semantic Gen-LaneNet that integrates semantic information from other traffic objects into a deep learning based 3D visual lane detection model. Our approach incorporates semantic masks of moving objects in a new branch and integrates the semantic features with the visual features using a feature fusion module. The fused feature map is utilized to predict top view lane lines, and real 3D lane points are obtained directly through geometric projection. Comprehensive experiments on the Apollo dataset demonstrate that our proposed method achieves at most 3.6% improvement in average precision (AP). In far-range detection, our method even exceeds the original model by up to 24%. Furthermore, various object segmentation algorithms can be successfully applied to our method in practical applications.