Lane-Level Three-Dimensional Semantic Mapping Based on Stereo Vision

Autonomous vehicles need to clarify their position and recognize objects in the urban scene, making it increasingly rely on maps to provide them with prior semantics for advanced tasks such as positioning and navigation or planning control. Semantic segmentation and geometric reconstruction techniques required for mapping have gradually developed. They have been combined to a certain extent for applications such as generating semantic maps. Still, they are rarely refined to the lane level, even if the lane can be used as a necessary constraint for vehicles to maneuver on the road. Important clues. This paper proposes a three-dimensional lane-level semantic mapping system that utilizes stereo vision and use the scrolling grid representation to save computing time and memory. We use the deep neural network to obtain the stereo disparities and lane-level semantics and then combined pose and depth to transmit the semantics into the three-dimensional space. Experiments on the KITTI data sequence show that our system can continuously identify and reconstruct objects on the road, even if their strong structure prior or few appearance clues.