Ground Moving Target Detection System Based On 3D Point Cloud

With the rapid development of artificial intelligence technology in all walks of life in China,driverless technology has become the core technology necessary for the rapid development of transportation industry.In this thesis,domestic traditional automobile manufacturers and high-tech enterprises have released their research results in the field of driverless,such as Jingdong,Suning and other e-commerce with logistics and transportation industries have also launched their own unmanned logistics vehicles.In this thesis,we will focus on the application and Realization of the ground moving target detection technology based on 3D laser point cloud.First,this thesis introduces three-dimensional point cloud data processing technology in object detection from two aspects: traditional method and deep learning method.By comparing the processing process of 3D point cloud data in different methods,it is found that the traditional method has more limitations than the method based on deep learning.Therefore,this thesis chooses voxelnet model as the reference neural network.However,the voxelnet based three-dimensional point cloud target detection method has a huge amount of computation,which can not be applied to mobile embedded systems.In this thesis,the deep separable convolution operation and voxelnet network are combined to reduce the amount of calculation,and a neural network which can meet the needs of unmanned logistics vehicles in the closed park is obtained.On the basis of theoretical analysis,this thesis designs a mobile car based on embedded system according to the operation of the car in the closed park.The mobile embedded platform uses 16 line laser radar as sensor,NVIDIA’s Jetson TX2 as embedded main board,and opencr development board as bottom driver board.The communication between embedded system and underlying driver is based on ROS node,and the communication between remote computer and embedded system is based on SSH protocol.Finally,this thesis tests and analyzes the improved voxelnet model,and finds that although the model sacrifices part of the accuracy,it meets the real-time demand on the mobile embedded platform,and can be effectively used on the unmanned distribution car in the closed park.