Research On Obstacle Detecting Algorithm Based On LiDAR Multi-frame Data Fusion

With the rise of techniques such as artificial intelligence,unmanned driving has become an important direction for the future development of automobiles.However,due to limitations in manufacturing costs and safety in use,the application and popularization of unmanned vehicles still have a long way to go.As the main sensor of unmanned vehicles,LiDAR’s high price is one of the factors leading to the high manufacturing cost of unmanned vehicles.One of the main technologies that determine the safety of driving vehicles is reliable and effective road obstacle detection technique,which has not yet reached a practical level.At present,most of the relevant research in the field of unmanned driving uses high-line LiDAR to implement the obstacle detection algorithm.Although this can effectively improve the detection performance,it also causes an increase in cost.Therefore,research on obstacle detection with high accuracy on low-cost LiDAR has become an important research direction in the field of unmanned driving.In this thesis,an obstacle detection algorithm is designed for unmanned vehicles based on multi-frame data fusion of LiDAR.This thesis uses a lower-cost 16-line LiDAR to conduct research on multiple aspects of LiDAR data enhancement,obstacle detection,and data set production,which effectively improves the effectiveness of obstacle detection.The main contents of this thesis are as follows:(1)The point cloud data of 16-line LiDAR is relatively sparse,resulting in low accuracy of the obstacle detection algorithm.Therefore,a multi-frame point cloud data fusion algorithm is proposed based on inertial measurement unit and dynamic target detection.According to the LiDAR’s characteristics of continuous scanning in the actual process,the algorithm uses the inertial measurement unit commonly installed on unmanned vehicles to fuse the historical point cloud information into the current point cloud,so as to enhance the density of the point cloud.(2)The general obstacle detection algorithm generally uses the Cartesian coordinate voxel division method,which does not retain the spatial features well.In this thesis,by introducing cylindrical coordinates,an obstacle detection algorithm based on cylindrical coordinates for voxel division is proposed.In the algorithm,the point cloud is grouped with voxels in cylindrical coordinates,and then the voxel feature extraction and spatial feature extraction are performed.This algorithm improves the obstacle detection performance.(3)At present,there is a lack of 16-line point cloud open source data sets for research.This thesis has designed a method for making 16-line point cloud data sets using open source 64-line point cloud data sets.By analyzing the difference between 64-line point cloud data and 16-line point cloud data,considering the point cloud density,point cloud spatial structure,label size and other factors,the 64-line point cloud data set was transformed intoa 16-line point cloud data set.(4)From the two aspects of data set and actual point cloud testing,the performance of the algorithm designed in obstacle detection is verified.