Object Detection System For Autonomous Vehicles In Park Using Sparse 3D Point Cloud

Environmental perception systems are essential for autonomous vehicles.Li DAR is widely used in sensing systems with its precise spatial measurement and anti-interference adaptability,but the high price hinders its application in the commercialization of autonomous vehicles.The low-cost Li DAR which has low-beam get sparse data and insufficient objects’ features.Therefore,the study of object detection systems based on sparse 3D point clouds is of great significance to promote the development of autonomous vehicles.This dissertation focuses on the research of autonomous vehicles object detection in the park.The main work includes:For ground removing,the method of combining plane fitting and area growth improves the accuracy of uneven ground separation.The plane fitting method controlled by the inclination is used to extract the ground point cloud of the nearby area,which is used as the initial seed data for regional growth,and the ground discrimination of other point clouds is performed according to the horizontal inclination.Enhance the tolerance to ground undulations through angle judgment;reduce the calculation amount of plane fitting through point cloud preextraction;ensure the correctness of plane extraction through inclination threshold.For object segmentation,the accuracy of sparse point cloud clustering segmentation is improved through multi-feature evaluation.The three data characteristics of depth map coordinate distance,spatial distance and angle are used to establish clustering evaluation standards,which suppresses the influence of excessive vertical distance of sparse point clouds,while retaining the height difference feature;improving real-time through effective neighborhood search.Compared to traditional single-feature clustering method improves the segmentation effect in the application of sparse point cloud,and enhances environmental adaptability and real-time performance.Combined with pre-processing such as denoising and filling,subsequent processing such as bounding box regression,false alarm suppression,a complete object detection system is constructed,and the program is accelerated through multi-threaded programming and structured data storage.The system performance is verified through experiments,and an average detection frequency of 18.6 Hz is achieved on the NVIDIA Jetson TX2 embedded platform,and it has good detection results in a variety of campus scenarios.The target detection method based on sparse point cloud studied in this paper has practical application significance for park unmanned driving and related autonomous mobile platforms.