Research And Application Of Pose Estimation Method Based On Visual Odometer

Visual odometry is a technology that uses adjacent frames to select and match feature points,tracks image changes between adjacent frames and then estimates the position and attitude of the target in space.Monocular odometer is a low-cost,environmentally adaptable solution for pose estimation,which is widely used in intelligent robots,autonomous driving,UAV navigation and so on.However,there are still some problems,for example,it has fuzzy scale and it is easy to be affected by environment.Therefore,for the purpose of improving the accuracy of pose estimation,the ultimate objective of my project is to enhance the auxiliary role of depth information in attitude estimation,improve traditional geometric methods and deep network feature learning ability and explore a visual odometer method that can achieve accurate pose estimation.The main contents includes the following three parts:(1)Depth estimation method based on camera perception multi-scale convolution.Due to the lack of scene depth information in the monocular method,it is hard to achieve scale-consistent pose estimation and global trajectory reconstruction.Firstly,the camera perception multi-scale module and mixed domain attention module are used to map the camera intrinsic matrix and depth features in the depth estimation subnetwork to learn the depth dependence of the scene in this method.Secondly,this method uses the pose estimation sub-network to assist in the construction of the scene depth map and weights all output feature channels,so as to help the network to better selectively focus on the channel features related to pose estimation and indirectly improve the depth estimation accuracy.Finally,the two subnetworks are trained jointly to obtain the scene depth results of the same scale.Experimental results show this method can effectively improve the accuracy of depth estimation and can be adapted to different camera models.Therefore,It has strong practical application value.(2)Hybrid monocular odometer method based on photometric calibration.Aiming at the problems that visual odometer is easily affected by environmental illumination and texture and that deep learning methods tend to ignore important geometric constraints,this method not only carries out in-depth research based on traditional geometric models and uses improved optical flow and depth estimation network as corresponding feature extraction and matching module in geometric models,but also uses the hybrid method for pose estimation.Firstly,this method improves the traditional photometric calibration model to calibrate the input video frame.Secondly,this method uses the optical flow and depth feature tracking module to extract highquality matching points and uses the geometric information to estimate the initial camera pose.Finally,this method estimates the accurate camera pose by aligning the pose information obtained by the traditional geometric method with the depth information obtained by the depth estimation module.The experimental results show that this method can not only effectively solve the problem of scale drift in monocular visual odometry and improve the accuracy of pose estimation,but also maintain good photometric consistency between consecutive frames,thereby reducing the impact of illumination changes in different scenes on pose estimation.(3)Based on the above studies,the pose estimation system built based on visual odometer was developed using pycharm platform and Pyqt5 framework.The system can realize depth estimation and pose calculation of video shot in different scenes,which has high accuracy and robustness.