Point cloud 3D target detection method based on key point multi-scale feature fusion

Abstract

The invention belongs to the field of 3D target detection, and particularly relates to a point cloud 3D target detection method based on key point multi-scale feature fusion, and the method comprises the steps: obtaining to-be-detected point cloud data at a current moment, inputting the obtained point cloud data into a trained point cloud 3D target detection model, and obtaining a target detection result; and improving distance sampling global features and an extraction algorithm of feature sampling global features in a point cloud 3D target detection model, and improving the efficiency and accuracy of target detection. According to the method, a feature farthest point sampling sequence extraction module is added, feature-based farthest point sampling acts on different voxel sparse convolution layers to obtain features of different scales, and the influence of background point and target detection is reduced.

Description

Technical field [0001] The present invention belongs to the field of 3D target detection, and is specifically involved in a point cloud 3D target detection method based on a key point multi-scale feature fusion. Background technique [0002] With the rapid development of 3D scene acquisition technology, 3D detectors such as 3D scanners, radar detectors, and depth cameras become more affordable, which provides a full advantage of using a 3D detector using a 3D detector in automatic driving field. Laser radar (LIDAR) sensor enters people's field of view. The large-scale data acquired by the LIDAR sensor is called, and the data group typically contains the laser beam emitted by the LIDAR to return the three-dimensional coordinate positioning and beam of the surrounding object to the laser intensity. [0003] In recent years, two-dimensional (2D) target detection under the camera system has achieved extraordinary achievements, but there is also some problems with objective detection ...

AI Technical Summary

Problems solved by technology

The main reason is that when sampling point clouds, 1. Ignore the information distinction between foreground points and background points. 2. Ignore the relationship between different scale features. 3. Ignore the accuracy of detection on heavily occluded objects.

Images