Flower 3D reconstruction method based on ORB and U-net

Abstract

The invention provides a method based on ORB and U-NET, including ORB-based 3D reconstruction of depth camera and u-Net. In the part of 3D reconstruction, the global posture optimization, loop detection and relocation are added to kinectfusion, which greatly improves the accuracy of the camera posture estimation and can get the high-quality flower 3D reconstruction model. Three-dimensional reconstruction is divided into three parts, because the three-dimensional data set of the flower objects is small, so the three-dimensional reconstruction is based on U-NET network structure, can obtain thebetter effect under the condition of having the smaller data set. Deep learning method can fill a large number of holes in the traditional dense slam method. The end-to-end method is implemented in the network to fill the holes, which can save a lot of time to fill the holes.

Description

technical field [0001] The invention relates to a three-dimensional reconstruction technology, in particular to a flower three-dimensional reconstruction method. Background technique [0002] In the current process of 3D reconstruction technology, the 3D reconstruction of small scene objects is mainly based on the dense slam technology of kinectfusion. By using kinect and other depth cameras to collect depth images around scene objects, and then process the depth images to obtain 3D reconstruction of small scene objects Model. The technical process of Kinectfusion is as follows: First, use the kinect camera to collect depth images around small scene objects, use the kinect camera internal reference to convert the original 2D depth image into a 3D point cloud in the camera coordinate system and calculate the normal vector; use ICP to iterate the minimization point The energy function to the plane is used to calculate the camera pose of the current frame; the 3D point cloud i...

AI Technical Summary

Problems solved by technology

[0003]The current 3D reconstruction method, for the scanning of heavily occluded objects such as flowers, due to the TOF or structured light technology adopted by the kinect camera, the occluded part has no depth data and unable to rebuild

In practice, it is also not feasible to ensure that the scan can cover all points on the flower surface (flower objects are often placed in the corner and the bottom of the flower cannot be scanned), thus resulting in the use of existing 3D reconstruction methods. a large number of voids

At the same time, because the kinectfusion method only performs pose estimation by registration between the current frame and the previous frame, there is no global pose optimization, loop detection and loop closure, which makes the pose estimation for the depth map very inaccurate during 3D reconstruction. At the same time The robustness of the fast movement of the camera itself is very low, and only slow movement modeling can be performed, which is easy to lose the camera pose

Images