Research On Monocular Visual Odometry Algorithm Based On Hybrid Method

Monocular visual odometry is a method that estimates the camera’s motion by using a single camera to capture a series of continuous images and utilizing deep learning or geometric methods.Geometric-based monocular visual odometry cannot directly obtain the depth of the scene and is susceptible to cumulative errors.Although deep learning-based monocular visual odometry can obtain the camera’s motion endto-end,its generalization ability is limited,and it ignores some important geometric constraints.Therefore,this thesis proposes a monocular visual odometry algorithm framework based on a hybrid method.This framework combines the image depth estimation network based on deep learning with the odometry calculation algorithm based on geometric methods,using the depth estimation network to recover the depth of the extracted feature points to improve the accuracy of visual odometry.This visual odometry algorithm can also estimate the depth of the image while implementing pose estimation.In the depth estimation part,to address the lack of image depth information in monocular visual odometry,an image depth estimation network which based on encoder-decoder network is proposed.Firstly,considering real-time issues,the Mobile Net network is used for feature extraction through depth separable convolution.Secondly,the Laplacian pyramid is combined into the decoder architecture,gradually recovering the depth boundaries through different scales of space,and reconstructing the final depth map from coarse to fine.In the pose estimation part,an improved FAST corner detection algorithm is proposed to address the complex logic of corner detection in the original FAST algorithm,improving the speed of corner detection and eliminating pseudo-corners that have not obtained response maximum in the surrounding neighborhood pixels to improve the quality of corners.Next,the tracked corners are based on the LK optical flow method.Then,the depth information of feature points is recovered using the depth estimation network.Finally,based on the geometric robust information criterion,suitable methods for motion estimation are selected from the epipolar geometric constraint method and the PnP method according to different motion situations to address the problem of poor motion estimation when motion degradation occurs.Experiments on depth estimation and pose estimation are conducted in this thesis.The results show that this monocular visual odometry algorithm based on a hybrid method is feasible and effective.The depth estimation performance is superior in all indicators on the Eigen Split dataset and works well under different lighting conditions and different scenes.The pose estimation performance on sequences 00,02,05,06,08,09 and 10 of the Odometry Split dataset is better than that of the compared geometric and deep learning-based visual odometry methods.