Research On Single-photon Lidar Imaging And Target Detection Methods

As a new type of 3D target detection technology,single-photon 3D imaging lidar can effectively obtain the information of 3D distance and surface reflection intensity of the target,and has the advantages of long imaging distance and strong anti-electromagnetic interference ability.Traditional optical detection technology,such as visible light detection and infrared detection technology,can only detect the target through the difference in surface texture information between the target and the background or the temperature difference between the target and the background.When the target surface is optically camouflaged or the temperature of the target surface is close to that of the background,the greater limitations shown in traditional optical detection methods.Under normal circumstances,the geometric shape characteristics of the target are difficult to change,so the use of single-photon 3D imaging lidar detection technology can effectively make up for the shortcomings of traditional optical detection methods,and better meet the target detection requirements in the battlefield environment and civilian fields.This dissertation mainly focuses on the four aspects of single-photon 3D imaging lidar imaging method based on deep learning,depth image simulation of lidar,depth image noise suppression,and ground 3D target detection method based on deep learning.The main work results of this article can be summarized into the following three parts:(1)A single-photon lidar imaging method based on depth estimation network is designed.In this part of the work,the imaging principle of the single-photon 3D imaging lidar was analyzed in detail,combined with the avalanche effect principle of the avalanche diode,the lidar detection process was modeled,and the single-photon lidar To F(Time of Flight)was realized by the simulation.The mathematical model of the pattern imaging mechanism and the simulation of the raw timing data of the single-photon lidar on the basis of the Middlebury indoor depth data set to generate a batch of raw data of the single-photon lidar.According to the principle of raw data photon counting imaging,combined with the idea of image semantic segmentation neural network in deep learning,a depth estimation network is designed to quickly generate high-quality depth images from raw photon timing data.(2)A single-photon lidar depth image noise suppression algorithm based on Fast adaptive Bidimensional Empirical Mode Decomposition is proposed.In this part of the work,the external and internal factors that produce noise in the single-photon lidar imaging process under long-distance detection conditions are analyzed,the noise in the depth image is classified,and the distribution probability density function of the depth image noise is summarized.Perform depth images simulation on battlefield targets such as tanks on the long-distance ground,and generate noisy depth images of battlefield ground targets at a distance of 3km under different signal-to-noise ratio conditions.A depth image noise suppression algorithm based on fast adaptive two-dimensional empirical mode decomposition is proposed,and noise suppression experiments are performed on the simulated depth image data and the measured single-photon lidar depth image,and a good range profile noise suppression effect is obtained.It improves the imaging quality of single-photon lidar and provides favorable conditions for 3D target detection.(3)A 3D datasets of battlefield ground targets is independently designed,a 3D target detection algorithm based on Point Pillars neural network model is designed,and the target detection experiment is carried out on the 3D target data.In this part of the work,several classic 3D target detection methods are compared,take the KITTI data set as an example to analyze the data characteristics of 3D point clouds,and the mapping method from 3D point cloud data to BEV 2D image is summarized.Summarizes the mapping method of 3D point cloud data to BEV 2D image,and compares the point cloud feature extraction means Voxel and Point Pillars encoders to get the advantages of Point Pillars encoder.According to the current status of 3D point cloud datasets mostly oriented to civil fields such as urban highway vehicle environments,a 3D point cloud datasets of ground targets in a battlefield environment is proposed and constructed,and field measurement experiments,indoor model measurement experiments and simulation experiments are carried out.A large-scale,multi-type 3D point cloud datasets of ground targets is constructed.And design and use the 3D target detection algorithm based on the Point Pillars network model to verify on the KITTI datasets,and obtain the 3D target detection effect with high accuracy and good real-time performance.