Study Of Vehicle Target Detection Algorithm Based On Lidar

With the rapid development of artificial intelligence,autonomous driving technology has been widely used in practical work scenarios such as smart city,transportation,and self-driving vehicles.As one of the key sensors of the automatic driving system,laser lidar(Lidar)can reconstruct 3D space about its surroundings.Besides,Lidar has the advantages of high precision,rapid response,long detection distance and strong anti-interference ability.Vehicle detection based on Lidar has also become a research hotspot in the field of autonomous driving at home and abroad.However,due to the sparseness and irregularity of point cloud data obtained by Lidar,research on vehicle detection based on Lidar faces many challenges.This paper mainly studies vehicle extraction based on machine learning method and vehicle detection algorithm based on deep learning.In order to solve the problem of point cloud loss caused by fixed number of initialization points in iterative self-organizing data analysis(ISODATA),this paper proposes a density adaptive initialization algorithm.The algorithm selects the point cloud by counting the density distribution of the point cloud,and solves the problem of losing the point cloud of the clustering result.In order to solve the problem that ISODATA uses Euclidean distance as the basis of merging two sub clusters,which leads to unsatisfactory clustering effect,this paper proposes a merging algorithm based on multi feature fusion and soft voting.The algorithm combines the intensity,density,covariance,spin image and sign of histogram of orientation(SHOT),which has a good clustering effect.Aiming at the problem of low segmentation accuracy of PointRCNN network using PointNet++ network,this paper proposes a point cloud segmentation network based on attention mechanism,and uses this as a basic framework for vehicle detection.In this paper,RS-CNN network is used as the basic framework to extract more semantic shape information.At the same time,an attention mechanism is added after the first two layers of RS-CNN convolution to obtain the neighborhood points that need to be focused.Laser point clouds have the defect of inconsistent distance and distance.Point clouds with higher neighborhood densities can acquire more semantic features than point clouds with lower neighborhood densities.Therefore,the candidate frames generated by point clouds with higher neighborhood densities are better.In view of this situation,this paper proposes an improved non-maximum suppression(NMS)algorithm.The algorithm calculates the neighborhood density of the point cloud and sets a density influence factor on the confidence of each candidate frame to obtain the density-based confidence.The candidate frames are sorted with distancebased NMS to remove redundant candidate frames.Experiments show that the algorithm has better detection effect.Finally,a vehicle detection system based on lidar is designed in this paper.The system includes a data playback interface,a single-frame point cloud data saving interface,a Lidar point cloud-based vehicle detection interface,and a comprehensive display interface.