Design And Optimization Of Weakly Supervised Visual Odometry Network Model

Visual Odometry(VO)is an important application of computer vision in the field of robot positioning and navigation.The traditional method is mainly based on motion constraints and uses geometric method to solve this problem,which usually includes complex processes such as feature extraction,feature matching and motion estimation.It has poor robustness in scenes with weak textures and dramatic changes in lighting.In recent years,with the wide application of deep learning technology in the field of computer vision,it has become an inevitable trend to apply deep learning technology to VO research.By using deep network to directly learn the geometric relationship between images,the complicated process of traditional method can be greatly reduced,and it has better adaptability to the environment.Early VO based on deep learning is mainly based on supervised methods,and training requires a large amount of data with pose labels.In view of the practical difficulties of obtaining the true value of high precision pose and laborious process of image labeling existing in supervised model,this paper mainly uses the unsupervised or weakly supervised method to study VO problem.The core idea of unsupervised and weakly supervised VO method is to transform the problem of depth and pose estimation into the problem of view reconstruction,and train the network of depth and pose estimation by using the constraints of view reconstruction.However,due to the lack of ground truth of pose supervision,monocular unsupervised VO suffers the problem of low accuracy and scale ambiguity.This paper propose to induce some weak supervision signals such as partial sparse depth or stereo images,which are more easily available than pose ground truth,to tackle this problem.In summary,the study of this paper mainly lies in the following three aspects:(1)On the basis of the Sf MLearner model^[1],the principles of unsupervised training and view reconstruction are studied.The model is further optimized by improving the network structure and optimizing the loss function.The resulting model is used as the baseline of the following study.(2)To handle the problem of scale ambiguity existed in the baseline model,a weakly supervised training model using limited sparse depth information is proposed.The framework of the weakly supervised network model and the acquisition method of sparse depth map are introduced,and the experimental verification is carried out.(3)A visual odometry model based on weakly supervision of binocular image is proposed.Based on the framework of the baseline model,the scale information is recovered by using binocular image training.Based on the latest research results,improvements in network structure,constraint conditions,loss function and other aspects are proposed.Some special optimization measures such as disparity up-sampling reconstruction,minimum reconstruction error and three kinds of masks are presented as well.The experimental results show that this binocular image based wearkly supervised model achieves excellent results,which proves the effectiveness of the method.