Lidar Point Cloud Decomposition And Reconstruction Algorithm With Application On Scene Perception

The development of three-dimensional sensors provides a new effective technology for researches on computer vision.Lidar based on light detection and ranging,can get enough spatial information for visual representations.Plenty of single points are collected and shown as point cloud.Through clustering analysis,point cloud can be used for region of interest extraction and target discrimination,and it can be used to realize the scene awareness.At present,lidar is widely used in intelligent transportation,mobile robot and other fields because of its high ranging accuracy,wide range and strong anti-interference ability.It can capture point cloud data with frame rate to obtain three-dimensional dynamic scene information.This paper takes lidar point cloud as the research object,and focuses on the related technologies of point cloud decomposition.The main work is as follows:(1)According to the overall structure of the automatic driving ship platform,combined with the role of the ship automatic identification system in the navigation,the automatic driving scene perception system based on lidar was designed.The working principle and data conversion of lidar based on time-of-flight ranging were analyzed.,and two dimensional mapping and display of point cloud were implemented.(2)Two different filtering algorithms,spatial filter based on the coordinates and point cloud filter based on projection quantitative,were compared,which can filter out outliers and noises of point cloud and retain the efficient points.The horizontal plane detection algorithm based on point cloud gradient was designed.Based on the characteristics of point cloud gradient,the lowest point representative were selected as the seed points.The grid data were then interpolated to fill the missing points.Then the initial plane point cloud was separated by 3×3 gradient filtering.The proposed algorithm was robust to the sparse point cloud with high estimated accuracy of the plane model.(3)A point cloud classification method based on the Bag of Words model is proposed.Computing Points Feature Histogram characteristics of point cloud and involving Points Feature Histogram characteristics of visual dictionary,point cloud feature representations based on Bag of Words model were generated.The machine learning methods K Nearest Neighbor and Support Vector Machine were used to deal with multiple categories of point cloud features.The experimental results show that Points Feature Histogram characteristics in Bag of Words model model can achieve the goal of point cloud target classification.(4)An improved two-phase registration algorithm based on Point Feature Histogram features was proposed.Firstly,the algorithm computed Points Feature Histogram according to the reference point cloud and point cloud to be registered for feature matching,rigid invariant constraints and random sampling consistency algorithm were used to remove the false matching points.Then the quaternion method of solving the transformation matrix was used to obtain the initial match point cloud.Finally,the Closest Point iteration algorithm was used to obtain accurate match according to the coordinates.Lidar data registration results showed the effectiveness of the algorithm.(5)This paper simulated ship encounter situation for the application of lidar in inland waterway,considering whether there was Automatic identification system information or not.By mapping the target ship model to the scene,a particular ship target segmentation was realized.A ship classifier based on Points Feature Histogram characteristics was built which can realize the decomposition.At last,the position and course of the target ship was obtained after reconstruction of the scene.