Model construction method of entity target, storage medium and computer equipment

Abstract

The invention provides an entity target model construction method, a storage medium and computer equipment, and the method comprises the steps: carrying out the semantic segmentation of an image containing an entity target, and obtaining a target outer contour of the entity target; inputting the target outer contour into the trained outer contour grammar classification network to obtain a target outer contour category corresponding to the target outer contour; obtaining a trained target body model parameter regression network corresponding to the target outer contour category; inputting the target outer contour into a target body model parameter regression network to obtain a camera external parameter matrix and body model parameters corresponding to the target outer contour; and constructing a three-dimensional model of the entity target according to the camera extrinsic parameter matrix and the volume model parameters. According to the entity target model construction method, the storage medium and the computer equipment, the entity target model can be accurately constructed, the operation is simple, and the consumed time is short.

Description

technical field [0001] The present invention relates to the technical field of model building of physical objects, in particular to a model building method, storage medium and computer equipment of physical objects. Background technique [0002] In the field of agricultural remote sensing applications, land resource surveys need to identify types of features, such as monitoring whether there is illegal construction of land without permission, need to identify buildings, and reconstruct the precise shape and size data of buildings. The existing monocular 3D target detection technology can be divided into one stage method that does not generate point cloud and depth map and two stage method that generates point cloud and depth map. [0003] Among them, the one-stage method that does not generate point clouds and depth maps can only obtain the 3D bounding box of the target, but cannot obtain the 3D shape of the target, so it cannot be accurately modeled; while the two-stage met...

AI Technical Summary

Problems solved by technology

[0003] Among them, the one-stage method that does not generate point clouds and depth maps can only obtain the 3D bounding box of the target, but cannot obtain the 3D shape of the target, so it cannot be accurately modeled; while the two-stage method that generates point clouds and depth maps, through the monocular In the 3D target detection technology, the two stages that generate point clouds and depth maps, and the 3D target shape reconstruction technology realize model reconstruction. The disadvantage is that high-precision point clouds or depth maps need to be obtained as label samples to train the deep learning network model. Therefore, the label It is difficult to obtain samples, and the equipment for obtaining labeled samples is expensive, and when this method is applied to target detection and modeling, it also has the disadvantages of high computational complexity and long cost

Images