Research On Unsupervised Visual Odometry Computation Based On Generative Adversarial Networks

Visual Odometry(VO)is a computer vision technique for estimating camera self-motion from continuous image sequences,which can recover the 6 degree of freedom(6-Do F)relative pose of the camera.With the rapid development of deep learning,visual odometry based on convolutional neural networks has achieved satisfactory performance in both driverless and robot visual navigation.Considering that the performance of supervised deep learning methods relies on a large amount of high-quality labeled data,it is often difficult to obtain such large and high-quality labeled data in practical applications.Therefore,this paper conducts research around unsupervised visual odometry methods.The main work of this paper is as follows:(1)Chapter 3 presents an unsupervised visual odometry method based on optical flow-depth-pose joint learning.In order to improve the estimation accuracy of camera pose,this paper combines the optical flow method and the idea of coarse to fine,and designs an accurate camera pose estimation model that combines optical flow motion features,namely Flow-Pose Module(FPM).It allows the neural network to make full use of the intrinsic connection between optical flow and camera pose to enhance the ability to learn the camera pose.Furthermore,an iterative optimization strategy of FPM is proposed,which is a camera pose estimation structure based on camera pose decoupling,and iteratively updates the camera pose through the recurrent unit FPM.In each iterative update,it learns a residual camera pose based on the previously estimated camera pose and generates a more accurate camera pose estimate.(2)Chapter 4 focuses on the impact of non-rigid motion and occlusion on visual odometry in dynamic scenes.Based on Chapter 3,this paper proposes an unsupervised visual odometry framework based on rigid perception and generative adversarial networks.The framework consists of a rigid perceptual generator and a rigid mask discriminator.By taking full advantage of the training mechanism of generative adversarial networks,it combines pixel-level and structure-level rigid perception,which is then used to improve camera pose estimation.Finally,to demonstrate the performance of the proposed network model,the proposed algorithmic model is trained and tested on the KITTI dataset.And the comparison with the results of other visual odometry methods shows that the proposed algorithm model has certain validity.