Research On Pose Estimation Technique Based On Depth Image

The cost of hardware based on visual pose estimation methods is relatively low,so the research on correlation algorithm has made great progress in recent years.The indirect method based on feature point or the direct method based on the intensity has appeared the classical pose estimation algorithm.The pose estimation method based on feature points needs front-end steps like extracting the feature points,establishing descriptors,matching feature point,thus the quality of points is relatively high.The front-end of the direct method based on intensity only needs to extract the points with obvious gradient gradient.In order to guarantee the robustness of estimation,more points will be used to participate in post-processing.Visual pose estimation can be divided into monocular,binocular and RGBD pose estimation by the type of sensor.The monocular pose estimation process deals with the image stream,and estimates the sensor pose which is one scale away from the real value.Binocular pose estimation by matching the left and right images,the depth of the scene can be obtained with the baseline,and the true position of the sensor can be obtained by processing the image stream.The pose estimation based on RGBD simultaneously processes the depth image and the RGB image to obtain pose estimation.There are some shortcomings in vision-based pose estimation algorithms.For example,the pose estimation method based on monocular vision sensor has scale missing problem and initialization problem,and the pose estimation method based on binocular vision sensor has serious computational resource consumption,needing hardware acceleration.For the pose estimation method based on feature points need extraction of feature points,the establishment of descriptors,front-end time-consuming and more affected by environmental problems,this thesis uses direct method to process depth image,avoiding the extraction of feature points,the establishment of descriptors,which solves the problem of pose estimation failure and feature extraction in less texture environment.The difference with the current direct method is that we optimize the depth residual,not the intensity error.In order to solve the problem of the pose estimation based on monocular vision and the resource consumption based on binocular vision pose estimation method,this thesis takes the depth image acquired by RGB-D sensor as the data source,and effectively solves the problem of pose estimation scale distortion,scale drift,the calculation of excessive consumption of resources and other issues.In this thesis,a pose estimation method based on depth image direct method is proposed.By optimizing the depth residual of the depth image,the 6-DOF pose information of the sensor in three-dimensional space is estimated,which reduces the resource consumption,improves the computational efficiency,and ensure real-time capability.When the camera is moving or turning fast,the object will quickly disappear from the measurement range or reappear in pose estimation based on direct depth image method.This will have a great impact on the direct pose estimation,which will cause the pose estimation mutation.For the shortcomings of the pose estimation method based on pure depth image,that is difficult to deal with the fast motion of the object,in this thesis,the robustness of the pose estimation is greatly enhanced by introducing the high frequency inertial measurement unit and the tightly coupled data fusion method.In this thesis,the pose estimation algorithm of deep image direct method is proposed,the pose estimation software based on depth image direct method is designed,and a package of pose estimation method based on Xtion Pro live sensor is developed.For the fusion of IMU information,the pose estimation software based on IMU is designed in this thesis,and the pose estimation software package based on Xtion Pro live and MTx-28A is developed.In order to verify the accuracy,real-time and robustness of the algorithm proposed in this thesis,we compare it with the classic ICP algorithm in open data set provided by the Technical University of Munich,Germany,to prove the accuracy and robustness of this algorithm.In addition,under the same experimental conditions,the pose estimation results based on the fusion IMU data are compared with the pose estimation results based on the pure depth image,and the robustness of the pose estimation method based on IMU fusion is verified.