CN-121616759-B - Laser radar point cloud data complement system and method based on PCN model

CN121616759BCN 121616759 BCN121616759 BCN 121616759BCN-121616759-B

Abstract

The invention discloses a laser radar point cloud data complement system and method based on PCN model, relating to the technical field of point cloud data processing, which is used for carrying out data complement on laser radar point cloud data with complex shielding, the system comprises a data loading module, a data processing module, a data complement module, a model training and optimizing module and a data visualization module, the method comprises the steps of obtaining an original point cloud data file, reading and data preprocessing on a multi-format point cloud file, the method comprises the steps of constructing a data set, constructing a point cloud data complement model based on a PCN model, training and verifying the model based on the data set, obtaining a weight file, and carrying out data complement on the point cloud data with complex shielding based on weight parameters, so that the problems of poor compatibility of the data format, overflow or unstable training of the numerical value, low matching and loading robustness of the point cloud file and the like when the coordinate range of the point cloud data is large in the conventional PCN model are solved, and the accuracy of sparse point cloud data complement is improved.

Inventors

LIU ZHEN
MA XIN
MA YUE
ZHANG HENG
ZI YILONG
YANG HAORAN
GUO LINGZHI
HAN YIFAN
ZONG YUCHENG
CHENG FAZHI
ZHOU YUHANG

Assignees

山东科技大学

Dates

Publication Date: 20260505
Application Date: 20260130

Claims (6)

1. A laser radar point cloud data complement system based on a PCN model comprises a data loading module, a data processing module, a data complement module, a model training and optimizing module and a data visualization module, and is characterized in that the data loading module reads and format analyzes a point cloud data file in any one of LAS, LAZ, TXT, PLY formats, pairs the point cloud data file based on a file name matching mechanism to obtain incomplete complete paired data, and constructs and divides a data set; the data processing module performs data preprocessing on incomplete pairing data to obtain standardized incomplete pairing data, wherein the standardized incomplete pairing data comprises incomplete point cloud data and complete point cloud data; The data complement module builds a point cloud data complement model based on the PCN model, performs data complement on the incomplete point cloud data based on the encoder and the decoder, wherein the encoder is connected with the decoder through feature mapping, the input of the encoder is the incomplete point cloud data, the output is global features and point level features, the input of the decoder is global features and point level features, and the output is the point cloud complement data; The model training and optimizing module compares and analyzes the point cloud completion data and the complete point cloud data, and carries out training supervision and super-parameter optimization on the point cloud data completion model based on the loss function to obtain a weight file; the visualization module performs visualization display on the point cloud completion data; Setting paths, categories and divisions of point cloud data files in a data loading module, defining data loading logic, constructing a multi-source data reading and fusing interface, finding the point cloud data files according to the categories of the point cloud data files, reading and format analyzing the point cloud data files in LAS, LAZ, TXT, PLY format by adopting a unified interface, acquiring three-dimensional coordinates of each data point in the point cloud data, and constructing Is a data matrix of the (c) data, Representing data points in point cloud data Is a sum of the number of (c), First representing Point cloud data Data points; Resolving scaling factors and offset parameters for point cloud data files in LAS format and LAZ format; performing separator identification and fault tolerance processing on the point cloud data file in the TXT format; Based on a file name matching mechanism, matching corresponding incomplete point cloud files and complete point cloud data files in the point cloud data files by adopting a three-level matching strategy, outputting a matching log, obtaining incomplete complete matching data, constructing and dividing a data set, and matching rules comprising complete matching, suffix stripping matching and prefix replacement matching; Carrying out data preprocessing on incomplete pairing data based on a data processing module, wherein the data preprocessing comprises data sampling, data denoising and data normalization, so as to obtain standardized incomplete pairing data; The data sampling comprises the steps of defining fixed points in a data processing module, carrying out point cloud sampling on incomplete complete paired data based on the fixed points, adopting repeated sampling on point cloud data with the total number of data points being smaller than the fixed points, retaining all data points in the point cloud data, adopting furthest point sampling on the point cloud data with the total number of data points being larger than the fixed points to carry out downsampling on the point cloud data, and ensuring that the sampling quantity of the data points in each point cloud data is the fixed points to obtain sampling paired data; The data denoising comprises defining a neighborhood radius and a lowest threshold in a data processing module, denoising the sampling paired data based on neighborhood density statistics, removing isolated points and neighborhood sparse data points in the residual defect cloud data, counting the number of residual data points with the distance not exceeding the neighborhood radius between the residual defect cloud data and the central data point by taking the data points as centers, obtaining a statistical value of the central data points, deleting the central data points with the statistical value smaller than the lowest threshold, and obtaining denoising paired data; The data normalization comprises the steps of eliminating geographic coordinate offset based on dynamic coordinate translation and the scale of self-adaptive normalization unified point cloud data, wherein the elimination of geographic coordinate offset based on dynamic coordinate translation comprises the calculation of the mass center of each point cloud data in denoising paired data, taking the mass center of the point cloud as an origin, and constructing a coordinate system; The data complement module comprises an encoder and a decoder, the encoder and the decoder are connected in cascade, connection is realized through feature mapping, the input of the encoder is incomplete point cloud data, and based on a KNN neighborhood aggregation algorithm, each data point of the incomplete point cloud data is searched in a neighborhood mode by using KD tree or ball query to find out the data point and the data point Personal and data points Nearest neighbor data point, build data point Is a neighborhood of (a) Calculating neighborhood data points relative to Based on the offset, obtaining data points by sharing a multi-layer perceptron or one-dimensional convolution Neighborhood characteristics of (v) Aggregating neighborhood features using symmetric functions to generate structural clues ; Constructing a shared multi-layer perceptron, learning data points Is the initial point characteristic of (a) And the original point features and the structure clues Splicing to obtain structural characteristics ; Introducing packet vector attention mechanism, characteristic of initial point Is divided into The group is used for calculating the attention weight of each group of initial point characteristics, obtaining the weighted group characteristics corresponding to each initial point characteristic based on the attention weight, summarizing the weighted group characteristics and obtaining data points Point-level features of (2); Processing point-level features of all data points in the incomplete point cloud data through maximum pooling to obtain global features ; The decoder comprises a rough point cloud generation stage and a dense point cloud generation stage; In the rough point cloud generation stage, the input of the decoder is global characteristic Generating rough point cloud data with fixed points through a multi-layer perceptron as a skeleton for supplementing the point cloud data; The input of the dense point cloud generation stage is global feature The point level characteristics of each data point in the residual point cloud data and the rough point cloud data, the full strength and the fine point budget of each data point in the rough point cloud data are determined based on the point level characteristics, the full strength of each data point is dynamically allocated with the full weight through a normalization function, and the full weight and the fine point budget are calculated on the data points based on the full weight The surroundings generate a minutiae.
2. The PCN model-based lidar point cloud data completion system of claim 1, wherein for each data point in the coarse point cloud data, a local two-dimensional grid is generated within the neighborhood of the data point based on the fine point budget, and global features are registered in the grid And the point-level characteristics of the data points and the complement weights are spliced to generate fine points, folding displacement of the local two-dimensional grid to the three-dimensional grid is predicted based on a multidimensional sequence folding technology of direction vector and lattice tiling, the generated fine points are superimposed on the data points corresponding to the rough point cloud data to generate dense point cloud data, and the dense point cloud data belongs to point cloud complement data.
3. The laser radar point cloud data supplementing system based on the PCN model according to claim 2, wherein the model training and optimizing module comprises a training stage and an reasoning stage, and in the training stage, the model training and optimizing module trains and super-parameter optimizes the point cloud data supplementing model of the data supplementing module based on incomplete complete pairing data and a loss function, and compares and analyzes the point cloud supplementing data with complete point cloud data to obtain model training parameters and obtain a weight file; and in the reasoning stage, a weight file is called, weight parameters trained by the model are imported into a data complement module, the point cloud data to be complemented is subjected to data complement through the point cloud data complement model, point cloud complement data are generated, and the point cloud complement data are visually displayed through a visualization module.
4. A PCN model-based lidar point cloud data completion system according to claim 3, wherein the loss function is a joint loss including a rough loss and a dense loss, obtained by calculating a chamfer distance between rough point cloud data and complete point cloud data, and between dense point cloud data and rough point cloud data; ; ; ; Wherein, the Indicating the distance of the chamfer angle, Indicating the loss of the association, As a coefficient of the loss of roughness, In order to have a dense loss coefficient, Representing the rough point cloud data, Representing the dense point cloud data, Representing the complete point cloud data, Data points representing coarse point cloud data, Data points representing dense point cloud data, Data points representing the complete point cloud data, Representing the square of the euclidean distance.
5. A method for supplementing laser radar point cloud data based on a PCN model, characterized in that the point cloud data supplementing system according to claim 1 is used, comprising: s1, acquiring an original point cloud data file, wherein the original point cloud data file comprises a data set construction file and a point cloud data file to be complemented, and the data set construction file comprises a residual point cloud data file and a complete point cloud data file; s2, carrying out format analysis on the original point cloud data file according to the file type to obtain three-dimensional coordinates of each data point in the original point cloud data file; s3, pairing the incomplete cloud data file and the complete point cloud data file based on a file name matching mechanism to obtain incomplete complete pairing data, and constructing a point cloud data complement data set; s4, carrying out data preprocessing on the point cloud data complement data set and the point cloud data in the point cloud data file to be complemented to obtain standardized incomplete complete pairing data and standardized point cloud data to be complemented; S5, constructing a point cloud data complement model based on the PCN model, dividing a point cloud data complement data set into a training set and a verification set according to a proportion, training the point cloud data complement model based on the training set, acquiring weight parameters, constructing a weight file, and performing super-parameter optimization and performance evaluation on the point cloud data complement model based on the verification set; s6, inputting the point cloud data to be complemented in the S4 into the point cloud complement model of the S5, calling the weight file to complement the data of the point cloud data to be complemented, obtaining the point cloud complement data and outputting the point cloud complement data in a visual mode.
6. The PCN model-based laser radar point cloud data supplementing method as claimed in claim 5, wherein the laser radar is used for collecting point cloud data files to be supplemented, the LAS, LAZ, TXT, PLY format point cloud data files to be supplemented are subjected to format analysis, the method comprises the steps of analyzing scaling factors and offset parameters of the LAS format and LAZ format point cloud data files to be supplemented, and carrying out separator identification and fault tolerance processing on the TXT format point cloud data files to obtain the point cloud data to be supplemented , Representing the total number of data points in the point cloud data to be complemented, the data points Three-dimensional coordinate information and category labeling information of the data points are stored in the data points; For a pair of Carrying out data preprocessing, including denoising the point cloud data to be complemented based on neighborhood density statistics, removing isolated points and sparse points in the point cloud data, eliminating geographic coordinate translation of the point cloud data to be complemented based on dynamic coordinate translation, and unifying coordinate scales of the point cloud data to be complemented based on self-adaptive normalization; Invoking training weight file, and based on weight parameters, performing data complement module pairing Performing data complementation and obtaining the completed point cloud data Through the pair of visualization modules And (5) performing visual display.

Description

Laser radar point cloud data complement system and method based on PCN model Technical Field The invention discloses a laser radar point cloud data complement system and method based on a PCN model, relates to the technical field of point cloud data processing, in particular to the technical field of three-dimensional point cloud data complement or data generation, and specifically relates to a point cloud data complement system and method which support multiple point cloud formats and can adaptively process large-scale geographic coordinates such as a laser radar. Background The laser radar point cloud data accurately depicts the real three-dimensional structure of the surrounding environment of the vehicle through millions of laser points per second, data storage is carried out through files in an LAX format and an LAZ format, each data point of the laser radar point cloud data comprises accurate three-dimensional point cloud geographic coordinates, the point cloud coordinate range reaches hundreds of meters, and the laser radar point cloud data is a basis for realizing automatic driving control. PCN is a classical point cloud completion network, and point cloud data completion is realized rapidly through point-by-point feature coding, but the accuracy of point cloud data completion is not high, and the PCN is suitable for point cloud data processing with regular shapes and smaller coordinate ranges, but has the problems of geometric detail loss and the like when processing point cloud data with complex shielding and data sparseness problems, has the problems of data overflow, normalization failure and unstable training when processing large-scale coordinates, and is difficult to cope with laser radar point cloud data completion of vehicles in the under-forest environment. The existing point cloud data complement system, especially the point cloud data complement system based on PCN model, mostly only supports PLY format file reading and processing when carrying out three-dimensional point cloud data processing, has poor data format compatibility, can not directly process laser radar (LADAR) point cloud data, and when carrying out data processing on the laser radar point cloud data, a format analyzer is required to be additionally arranged to convert the point cloud data file format, can not directly read different types of data and multiple types of mixed data, and has low data reading efficiency and high risk of data reading conversion errors. Therefore, a system and a method for supplementing point cloud data of a vehicle in a forestation environment are needed, which can read and analyze multi-format point cloud data, improve the point cloud data supplementing precision with a complex shielding structure, and are suitable for the point cloud data supplementing of the vehicle in the forestation environment. Disclosure of Invention The invention aims to provide a laser radar point cloud data complement system and method based on a PCN model, which are used for solving the problem of low point cloud data complement precision caused by poor data format compatibility, unstable training or numerical value overflow when the point cloud data coordinate range is large, low point cloud file matching and loading robustness and the like when point cloud data complement is performed in the prior art. In order to solve the problems, the invention provides a laser radar point cloud data complement system based on a PCN model, which comprises a data loading module, a data processing module, a data complement module, a model training and optimizing module and a data visualization module, wherein the data loading module reads and format-analyzes a point cloud data file in any one of LAS, LAZ, TXT, PLY formats, pairs the point cloud data file based on a file name matching mechanism to obtain incomplete complete paired data, and constructs and divides a data set; the data processing module performs data preprocessing on incomplete pairing data to obtain standardized incomplete pairing data, wherein the standardized incomplete pairing data comprises incomplete point cloud data and complete point cloud data; The data complement module builds a point cloud data complement model based on the PCN model, performs data complement on the incomplete point cloud data based on the encoder and the decoder, wherein the encoder is connected with the decoder through feature mapping, the input of the encoder is the incomplete point cloud data, the output is global features and point level features, the input of the decoder is global features and point level features, and the output is the point cloud complement data; The model training and optimizing module compares and analyzes the point cloud completion data and the complete point cloud data, and carries out training supervision and super-parameter optimization on the point cloud data completion model based on the loss function to obtain a weight file; And the visualization module per