Research On Visual Odometry Based On Optical Flow And Depth Estimatio

As an important part of visual Simultaneous Localization and Mapping(SLAM),visual odometry is widely used in fields such as unmanned driving and mobile robots.The traditional visual odometry method can be easily affected by the environment based on geometric theory.The use of deep learning in the completion of motion estimation is proven to be feasible.This type of method does not require complex procedures such as camera calibration,but the performance still needs to be improved.As associated tasks of visual odometry,optical flow can provide richer features,and depth estimation can provide distance information.Therefore,the research on visual odometry is conducted based on the optical flow and depth estimation.Aiming at the deficiencies of existing research,this paper uses more accurate optical flow and depth information to effectively improve the performance of visual odometry in different research directions.Specifically,the main work of this thesis includes the following aspects:1.This thesis proposes a visual odometry model based on optical flow and self-attention.First,the model introduces the advanced optical flow estimation RAFT model to extract optical flow rich in motion information.Secondly,the relationship between image sequences is learned by combining Bi-LSTM.Finally,referring to the idea of keyframes in the traditional method,the self-attention mechanism is introduced to obtain the relative pose between the last two frames of images through the features of multiple consecutive frames.Experimental results show that the model significantly reduces the root mean square error of rotation and translation.2.This thesis proposes a visual odometry model based on depth estimation and Transformer.First,the model introduces Pack Net,a new encoder-decoder framework,to estimate more accurate depth.Secondly,the Transformer is introduced to learn the depth and original image features to obtain a richer feature representation and perform pose estimation.Finally,the geometric consistency loss is introduced to solve the scale inconsistent problem.Experiments show that the model has a significant improvement in accuracy.3.This thesis proposes a visual odometry method combining optical flow and depth estimation.First,the method obtains forward optical flow,backward optical flow and depth map through optical flow estimation network and depth estimation network.Secondly,the method establishes more reliable 2D-2D or 3D-2D correspondences.Finally,the pose is solved by Epipolar geometry or Pn P.Experiments prove the effectiveness and robustness of the method.