Point cloud data processing method, device and electronic equipment

Abstract

The present invention provides a point cloud data processing method, device, and electronic equipment. The method includes: acquiring an initial point cloud data set and a front view corresponding to the initial point cloud data set, performing a convolution operation on the front view, and obtaining the corresponding The convolution feature map; wherein, each feature point in the convolution feature map is determined by the following method: according to the depth of each front view feature point in the convolution range, determine the first feature point corresponding to the convolution range Set; feature aggregation is performed on the first set of feature points to obtain the feature points corresponding to the convolution range. In the process of extracting the features corresponding to the convolution range, the depth of the feature points in each front view is considered. Based on the depth information The determined feature map of the target object can avoid including the features of other objects at the connection of multiple objects, reduce the possibility of wrong feature extraction, and then improve the accuracy of the feature map and enhance the robustness of the feature.

Description

technical field [0001] The present invention relates to the technical field of data processing, in particular to a point cloud data processing method, device and electronic equipment. Background technique [0002] With the rapid development of 3D sensor technology (lidar, RGB-D camera, multi-eye camera, etc.) and the gradual reduction in cost, more and more autonomous driving devices begin to use 3D sensors as necessary perception sensors. Compared with traditional 2D sensors (such as color cameras), 3D sensors can obtain rich geometric position information in the traffic environment, improve the perception performance of unmanned vehicles, and thus ensure the safety and efficiency of autonomous driving equipment. Among them, based on the three-dimensional point cloud data obtained from the lidar and the method of machine learning (neural network), the target detection, tracking, segmentation, etc. of the traffic environment are an important part of the autonomous driving en...

AI Technical Summary

Problems solved by technology

However, in the front view, the target object may have situations such as near large and far small, mutual occlusion, etc. Compared with rasterization or original point cloud data in 3D space, the mainstream 2D ​​convolution feature extraction method ignores the geometry of the target object. Size information and relative position information between objects, feature extraction is performed on the entire feature area (front view point cloud), resulting in target features of different position sizes being mixed at the connection, thereby reducing the robustness and accuracy of the feature

Images