Camera attitude estimation method based on deep neural network

Abstract

The invention discloses a camera attitude estimation method based on a deep neural network. The method comprises the following steps: 1) constructing a camera attitude estimation network; 2) constructing an unsupervised training scheme, respectively reconstructing corresponding images from the input front and back frame images by using the estimated depth map, the inter-frame relative pose and theoptical flow, and constructing a loss function of the network by using a luminosity error between the input image and the reconstructed image; 3) sharing a feature extraction part by the pose estimation module and the optical flow estimation module, and enhancing the geometrical relationship of the features to the interframes; and 4) inputting a to-be-trained single-viewpoint video, outputting acorresponding inter-frame relative pose, and reducing the loss function through an optimization means to train the model until the network achieves convergence. According to the model provided by theinvention, the camera pose of the corresponding sequence is output by inputting the single-viewpoint video sequence, the training process is carried out in an end-to-end unsupervised mode, and pose estimation performance is improved through optical flow and pose joint training.

Description

technical field [0001] The invention belongs to the field of computer vision and relates to a camera pose estimation method, in particular to a camera pose estimation method based on a deep neural network. Background technique [0002] Camera pose estimation, as the most important part of Simultaneous Localization and Mapping (SLAM) technology, has attracted extensive attention in the field of computer vision and the robotics community in the past few decades. At present, it has been widely used in various aspects such as GPS global positioning system and inertial navigation system (INS) of various robots. [0003] Although traditional pose estimation algorithms such as ORB-SLAM and VINS-mono can achieve quite high accuracy, these algorithms often cannot cope with scene changes, and the effect of image processing with sparse textures will be greatly reduced. Convolutional neural network (CNN) has achieved good results in traditional computer vision tasks such as target dete...

AI Technical Summary

Problems solved by technology

[0004] The pose estimation method based on deep learning is very dependent on the extracted features. The nature of the feature representation determines the effect of pose estimation. The pose network trained using features related to image surface information often performs poorly in unfamiliar scenes.

Images