LiDar Point Clouds Environment Perception Based On Deep Learning

With the rapid development of artificial intelligence technology,as well as the urgent demand for new technologies,a large number of intelligent innovation and application have been generated.The construction of intelligent transportation,represented by intelligent vehicle,has become an important subject with great research value.Light detection and ranging(LiDAR)sensors on the basis of the advanced laser ranging principle,can obtain high-precision environmental point clouds information,and its measurements ranged over a wide range.In addition,it is not easily affected by weather,illumination,object surface texture,etc.,and has been widely applied in environmental monitoring,autonomous navigation,path planning and so on.It is the core equipment of intelligent vehicle to realize environmental perception.However,LiDAR point clouds data have large volume,inhomogeneous densities,and non-structural distribution,and the multi-frame data obtained by continuous scanning is disordered in the memory,so it is difficult to use convolution kernel directly to extract geometric information between points.To achieve high-precision environment perception in LiDAR point cloud,this paper completes the following tasks:1)In this paper,an intelligent vehicle loaded with HDL-32 E Velodyne LiDAR sensor is used to scan the real urban traffic road environment,obtain three-dimensional(3D)environmental point clouds,then quickly segment it to generate a series of point clusters.On this basis,a semi-automatic object point clouds labeling tool is developed,and create a LiDAR point clouds dataset for object recognition.This LiDAR point clouds dataset has the advantages of abundant quantity,real data,accurate label,etc.2)Combined with the principle of Hough transform,the object LiDAR point clouds are mapped into Hough space and rasterize,to generate Hough accumulator space,and compute the each cell in the Hough accumulator space corresponding mapping point count as its’ feature.In the experiment,3D object point clouds mapping into 3D Hough space to extract the global Hough features.In order to retain the local spatial structure of the object,the multi-scale key point sampling method based on grid and the dynamic nearest neighbor sampling method based on grid were used for sampling,and then use the combination of Hough transform to extract the local Hough feature.3)In LiDAR point clouds objects classification,first input the global Hough feature into the 3D Convolutional Neural Network for feature learning,and then use the Multilayer Perceptron(MLP)to learn its local feature,finally the two parts of the network are connected and input into the full connection layer for further calculation.By a large number of iterative training,we have achieved high-precision LiDAR point clouds object classification,and the average classification accuracy has reached97.6%.