Construction area target identification method and device based on radar overhead angle segmentation

Abstract

The application provides a construction area target identification method and device based on radar top view angle segmentation. The method comprises: acquiring radar point cloud data in a construction area; dividing the spatial range of the construction area into a plurality of grids with preset length, width and height thresholds, and performing voxelization processing on the acquired radar point cloud data to generate a binary voxel grid matrix representing whether each grid has point cloud data; inputting the binary voxel grid matrix into a BevSeg model to output a category label representing a target in the construction area and a 3D target range frame corresponding to the target, wherein the BevSeg model is a BEV angle point cloud segmentation model based on a convolutional neural network (CNN) and sparse 3D convolution, and target identification and parameter regression are completed through a U-Net full convolution structure combining 3D convolution and 2D convolution; and determining the category of all radar point cloud data to complete target identification in the construction area. Through the application, automatic identification of targets in the construction area can be realized, and the hardware requirements and costs are reduced.

Images