Three-dimensional target detection method and system based on point cloud weighted channel characteristics

Abstract

The invention relates to a three-dimensional target detection method and system based on point cloud weighted channel characteristics, and the method comprises the steps: carrying out the extraction of a target in a two-dimensional image through a pre-trained deep convolutional neural network, and obtaining a plurality of target objects; based on each target object, determining a point cloud viewcone in the corresponding three-dimensional point cloud space; segmenting the point cloud in the view cone based on a segmentation network of the point cloud to obtain an interested point cloud; And based on the network with the weighted channel characteristics, performing 3D Box parameter estimation on the point cloud of interest to obtain 3D Box parameters, and performing three-dimensional target detection. According to the method, the characteristics of the image can be learned more accurately through the deep convolutional neural network. Based on a network with weighted channel characteristics, 3D Box parameter estimation is carried out on the point cloud of interest, the weight of characteristic drop of unimportant points can be reduced, the weight of key points can be increased, interference points can be restrained, the key points can be enhanced, and therefore the precision of 3D Box parameters can be improved.

Description

technical field [0001] The invention relates to the technical fields of computer vision and image processing, in particular to a three-dimensional object detection method and system based on point cloud weighted channel features. Background technique [0002] 3D object detection based on point cloud is a very important task in unmanned driving. It is required to input point cloud data, sometimes it may also need to input the corresponding RGB image data, and then output the parameters of 3D Box. [0003] Generally speaking, it is first necessary to rasterize the 3D point cloud, and then use the 3D convolution operation in each grid to extract the features of the point cloud, but the 3D convolution operation is very computationally intensive in large scenes, so It will make it difficult to apply to real scenarios. Through observation, it is found that the point cloud in the real scene is very sparse. [0004] Generally speaking, the 3D data output by unmanned driving senso...

AI Technical Summary

Problems solved by technology

Since the segmentation effect cannot be completely accurate, there are more or less interference points in these point clouds, which will have a negative impact on the final 3D Box

Images