Research On Visual Odometry Algorithm Based On Unsupervised Deep Learning

Visual odometry is a technology for estimating the motion trajectory of a mobile robot based on the image information obtained by the image sensor of the mobile robot.It is an important link for the Simultaneous localization and mapping(SLAM)technology to be practical,and has important theoretical significance and practical research value.In recent years,visual odometry has made more and more progress,but there are also many problems,including the existence of geometric-based methods that can only be applied to specific environments,unable to adapt to visual degradation or extreme natural environment,the accuracy of the based on deep learning method has yet to be improved.How to effectively solve the above problems are the current hot research direction and the main research content of this thesis.In this thesis,the traditional geometric-based visual odometry method and the deep learning-based visual odometry method are improved,and the evaluation is made on the public data set,which has achieved good results.The main work and innovations are as follows:Firstly,the development status of visual odometry is reviewed,its basic principles are expounded,and the background knowledge based on geometry,deep learning and multi-task deep learning methods are described.Secondly,the traditional geometric method-based visual odometry is improved in the anti-noise performance and low accuracy in response to illumination changes,dynamic target motion,camera parameter changes,and lack of texture or texture.Multi-task deep learning methods,using neural networks to learn image features,binocular images to infer depth information,time-series images for pose regression.At the same time,due to the use of unsupervised deep learning methods,it can solve the problem that the method of supervised deep learning relies on a large number of data annotations.The experimental results show that the proposed algorithm enhances the noise immunity of the model,improves the accuracy,and can be trained through a large amount of unsupervised data,which is more suitable for complex environments.Then,based on the unsupervised deep learning method using time series image to infer the pose,it is proposed to add the deep flow constraint to the unsupervised deep learning method,construct the multi-task learning method,and carry out the two-stage four-constrained learning through the joint training network.The four constraint functions jointly constrain the two-stage network,so that the model obtained in the final training only needs to input the time series image,and the pose information can be inferred.Experimental results show that this algorithm effectively improves the trajectory accuracy.Finally,the work done in this thesis is summarized and the direction to be further studied is pointed out.