Multi-Sensing Drivable Area Detection Algorithm For Urban Intelligent Vehicles

Autopilot system is mainly divided into three parts: perception,cognition and control decision.The drivable area is an important part in the perception,and its detection accuracy plays a crucial role for trajectory planning and making policy decisions.The drivable areas in urban scenes are mainly divided into structured road with clear road edge lines;semi-structured road with unclear road lane lines,and unstructured roads with no structural layers roads surface.In this paper,we mainly propose two innovations points.One is road detection by fusing LIDAR point clouds and camera images for semistructured road,another is multi-task lane detection algorithm based on line spatial characteristics for structured road.In the semi-structured area,vehicle in order to realize the local location and path planning.Because of lack of the lane information,the road areas with disregard obstacles has become one of the most important task of self-driving car.Single camera is easly affected by the level of illumination and shadow occlusion.However,LIDARs use the laser light which are undisturbed by serious lighting conditions and can provide accurate distance in the environment.In order to eliminate the impact of light on the camera and increase the robustness of road segmentation under complex illumination.This paper designs a road segmentation network C-LNet with a multi-fusion of LIDAR and camera.Firstly,aiming at the two heterogeneity informations of laser and image,this paper completes the heterogeneous data in the same spatial.Secondly,considering the different fusion modes of information,this paper designs a multi-scale feature fusion method based on different scale information by deep learning algorithms.Compared with other fusion methods,the MIoU has improved about 2%,and the results in complex scenes such as shadow occlusion and light changes are significantly improved.Urban self-driving car positions itself within the road lanes and plans the path based on lane in structured area.Aiming at the lane in the structured area sensitive to occlusion and light,this paper proposes a segmentation network based on the spatial characteristics of the lane.Firstly,we propose a spatial asymmetric receptive field structure by the spatial characteristics of lane in the image.Meanwhile,in order to adapt different scales of lane,we propose a lane receptive field pyramid structure,in order to improve lane segmentation result.Secondly,we add the instance segmentation branch in the network,and segment the lanes in different lane cases.Our model is training on the Tusimple dataset and compared with the mainstream segmentation network.The result has a 2% to 4% improvement.In order to verify the previously proposed driving area detection algorithm,we constructs a L4 self-driving car experiment platform.The experiment takes the campus road scene as the goal,according to the multi-sensor fusion road detection and structured road lane detection.We obtaine the range of drivable area,meanwhile the yaw angle and the offset distance from the decision trajectory are provided for the decision.We verify the road segmentation algorithm and lane line detection algorithm proposed by real vehicle,and we completed the simple autopilot system.