DSNet: Lane Centerline Reconstruction with a Driving Scene Reasoning Network

In the field of autonomous driving, accurate detection and tracking of lanes play pivotal roles in ensuring safe and efficient navigation. Traditionally, navigation relies on pre-existing high-definition maps. However, a revolutionary approach is emerging – one that harnesses the power of vehicle-surrounding images for real-time lane reconstruction. We introduce a novel approach called DSNet, which is designed for real-time inference of driving scene topology and online construction of lane centerlines using topology heatmaps derived from information in the driving scene. DSNet uses topology heatmaps to facilitate the exchange of information between lane centerlines, allowing them to share insights while also enabling information exchange between traffic elements and lane centerlines. We evaluate our approach on OpenLane-V2, a comprehensive benchmark for topology reasoning. DSNet achieved outstanding results in the reconstruction of lane centerlines, particularly in complex scenarios.