Deep learning-based identifying and processing method for laser radar point cloud data

Abstract

The invention belongs to a deep learning-based identifying and processing method for laser radar point cloud data. The result of a neural network model is computed by using a computing method of yolov2 when identification is carried out by using the trained neural network model; anchor values corresponding to the categories are utilized separately when the width and height of the result are computed; and average width and height of the same category of all objects in annotation data are preferred as the anchor values. According to the deep learning-based identifying and processing method, thepreprocessing process of the point cloud data is avoided; the categories of the objects can be identified; three-dimensional characteristics of the point cloud data are kept through converting the point cloud data into the forms of a distance matrix and a reflected light intensity matrix as input of a deep learning neural network, thereby ensuring relatively high identification rate on small objects and controlling the false identification rate at a very low level; and the finally obtained result is the result in a matrix dimension and can be converted according to the actual scene.

Description

Technical field [0001] The invention relates to the field of deep learning, in particular to a deep learning-based recognition processing method for lidar point cloud data. Background technique [0002] LiDAR-Light Detection And Ranging, that is, laser detection and measurement, also known as LiDAR, is a radar system that emits a laser beam to detect the location and speed of the target. Its working principle is to transmit a detection signal (laser beam) to the target, and then compare the received signal (target echo) from the target with the transmitted signal. After proper processing, relevant information about the target can be obtained, such as Target distance, azimuth, height, speed, attitude, even shape and other parameters, the measured data is a digital surface model (Digital Surface Model, DSM) discrete point representation, the data contains three-dimensional information and laser intensity information. According to the number of laser beams, lidar can be divided int...

AI Technical Summary

Problems solved by technology

Unsupervised algorithms usually require cumbersome preprocessing during the implementation process, such as first fitting the ground, then filtering the points on the ground, and then filtering out the points around the road that are not related to driving, and then can be recognized. The effect of preprocessing It will have a great impact on the final recognition effect, and the unsupervised algorithm cannot directly recognize the object category

However, the existing deep learning algorithms do not have a good method for converting point cloud data into the input of the deep learning neural network, resulting in a very low recognition rate for small objects, which cannot be recognized basically, and there are also high errors. Recognition rate, the background point cloud will be recognized as the target point

Images