Panoramic camera attitude estimation method based on depth learning

Abstract

The invention discloses a panoramic camera attitude estimation method based on depth learning. Based on a depth convolution network and a vanishing point, the three-dimensional orientation of the panoramic camera is estimated, and the rotation matrix of the camera relative to the world can be quickly estimated only by ordering a single panoramic picture. The method includes image data acquisition,image data processing, establishing a new network structure PanoNet for panoramic camera attitude estimation, training PanoNet and predicting panoramic camera attitude. The invention uses a single panoramic picture, and can estimate the rotation of the panoramic camera relative to the world coordinate system under the scene of occlusion, void, weak texture and the like. The method has high robustness.

Description

technical field [0001] The present invention relates to a panoramic camera pose estimation method, in particular to a new deep learning-based panoramic camera pose estimation method, which can accurately estimate the rotation matrix of the camera relative to the world coordinate system, namely the panoramic camera, from a single panoramic picture attitude. Background technique [0002] Camera pose estimation has always been a popular research direction in computer vision, and it is also a basic problem in computer vision. It has a wide range of applications in our lives, such as drones, autonomous driving, and virtual reality. There are three main approaches to existing camera pose estimation. [0003] The first method is mainly based on multi-viewpoint geometry. Given multiple pictures, first extract the feature points of each picture, mainly including SIFT, SURF, ORB and other feature point extraction methods, and then match the feature points of each frame, and then Acc...

AI Technical Summary

Problems solved by technology

However, there are also problems with this method. One is that it is difficult to obtain a large amount of labeled data. Secondly, the selection of the predicted label, if the Euler angle is directly output, the Euler angle has singularity and periodicity, and the effect will be relatively poor. If the quaternion is output, the semantics of the quaternion is relatively weak, and it is difficult to get better results

[0006] Therefore, the above-mentioned existing three methods all have relatively large problems, and are difficult to be applied to practice.

Images