Research On Obstacle Detection Based On Multi-lidar Fusion

In the research of automatic driving,laser radar is becoming more and more popular.Mechanical lidar has a range of 360° and is often used to sense information about the environment.Although solid-state lidar has a limited detection range,its point cloud is informative and inexpensive,making it ideal for use as a point cloud augmentation device.Using multiple lidar has become a mainstream solution for many self-driving cars to sense their environment.This scheme makes the perception range wider and can bring more point cloud information to the vehicle.It can also remove the occlusion and improve the reliability and robustness of environmental perception.In this paper,a driverless bus is used as the experimental platform,and the fusion of mechanical lidar and solid state lidar is used to realize the environmental detection on the platform.Therefore,this paper proposes a practical automatic point cloud registration scheme.Firstly,the initial rotation parameters were solved by using the extracted common features,and the solution was optimized by minimizing the nonlinear cost function.Then,time synchronization and distortion removal were carried out for them.Finally,the data fusion between multiple lidar is realized successfully.The automatic calibration method in this paper does not require manual initialization and does not rely on additional sensors at all.The experimental results show that this method can successfully achieve fast,stable and reliable high-precision registration,and the Angle error and distance error are less than 0.01 rad and 0.02 m,respectively.Point cloud filtering is used to cut the fused point cloud data to remove the interference points and reduce the number of point clouds.RANSAC algorithm is used to separate ground points from non-ground points.In this process,the algorithm is optimized by grid-based pre-segmentation and accelerated iteration result judgment.Then,Euclidean clustering is used for non-ground points,and the algorithm is optimized by region pre-segmentation,dynamic clustering threshold detection and region boundary quadratic fitting.Through experiments,the scheme in this paper can solve the problems of incomplete separation of the far end of the ground and unrecognizable low slope,and solve the problems of inconsistent detection effect of the near and far end of the obstacle clustering,easy fragmentation of targets and missed detection.The method in this paper successfully achieved fast,stable and reliable obstacle detection,with the detection accuracy of about 95.52% and the detection time of about 100 ms.The research in this paper realizes the high integration of multi-lidar in time and space.The devices used in this study include MEMS solid-state lidar with unique scanning mode,which is used to provide reliable data sources for environmental awareness.In addition,this paper optimized the traditional algorithm to achieve obstacle target detection,providing reference for the traditional algorithm in automatic driving target detection.