A deep learning-based automatic classification method of vehicle road scene point cloud

Abstract

The invention discloses a deep learning-based automatic classification method and system for vehicle road scene point clouds. The construction of a training sample set includes collecting three-dimensional point cloud data of vehicle-mounted lasers in road scenes, labeling category labels, and combining the three-dimensional point cloud data with corresponding The category label information is packaged as a point cloud data set; a three-dimensional deep learning network model based on the octree data structure is designed, and the three-dimensional deep learning network model adopts a U-shaped fully convolutional neural network model; using the weighted loss function for each category, To solve the problem of large differences in the number of points of different types of vehicle laser point clouds in road scenes, the network model is trained based on the training sample set to obtain a trained point cloud classification model; the vehicle laser 3D point cloud data to be classified is input. By adopting the technical solution of the present invention, the laser point cloud in the vehicle road scene with the classification targets of different types, numbers and sizes can be automatically classified stably and accurately.

Description

technical field [0001] The invention belongs to the field of automatic classification of laser point clouds of vehicle road scenes, in particular to a method for automatic classification of laser point clouds of vehicle road scenes based on deep learning. This method learns and trains the 3D laser point cloud datasets that have been labeled with category labels, and the final deep learning model obtained by training can stably and accurately perform laser detection in vehicle-mounted road scenes with different types, numbers and sizes of classification targets. Point cloud for automatic classification. Background technique [0002] Laser point cloud automatic classification technology refers to the process of automatically classifying each three-dimensional point into a specific category by using appropriate feature extraction, learning and classification methods for the original laser point cloud with three-dimensional spatial coordinates and other information. The efficie...

AI Technical Summary

Problems solved by technology

However, this method projects three-dimensional space points to a series of two-dimensional planes, which will inevitably lose a certain amount of spatial detail information.

At the same time, there are still a series of problems such as projection viewing angle selection and self-occlusion.

Therefore, it cannot fundamentally solve the difficult problem of using deep learning to process three-dimensional data

Images