Research On 6D Pose Estimation Algorithm Of Target Based On Keypoint Location

The pose estimation of rigid body object is one of the hot technologies of the current direction of computer vision,the aim of which is to determine the motion of the rigid body object relative to the camera in translation and rotation with a total of six degrees of freedom through the image.Pose estimation has a promising future in multiple fields such as Robotics,Space Satellite Docking,Satellite Capture,Autonomous Driving,and Virtual Reality.The pose estimation technology based on studies,relying on fast computational efficiency and good estimation accuracy,has gradually replaced the traditional pose estimation method with its artificial design features.Limited by the limitations of the RGB-D data acquisition environment,algorithms using RGB-D data are hard to be massively applied in project.By contrast,algorithm using single RGB image has the advantages of insensitive to the environment,large field of view,high computing efficiency,low cost,etc.,and is more suitable for project application.In the single RGB image algorithm,using the two-stage algorithm corresponding to the keypoints is widely studied for its good ease of use and excellent accuracy.However,when objects are cluttered or obstructed by other objects in the environment,it will be challenging to conduct object pose estimation utilizing keypoints correspondence with single RGB image.Hence,this paper studies the problem of using single RGB image to estimate the pose of the object with keypoint correspondence method,and the main study work is as follows:1.In the method of predicting keypoints using heatmaps,this paper constructs the heatmap regression network utilizing HRNet multi-layer feature fusion feature to improve the ability of heatmap regression network to extract image features.Meanwhile,to improve the accuracy and robustness of the predicted keypoints,the loss function Heatmap Wing Loss applicable for heatmap regression is proposed.The loss function has loss function gradient for foreground pixels of heatmap,which enables the network to focus more on the foreground pixels of the image and realize more accurate heatmap regression.With the new heatmap regression network as the basis,the two-stage object pose estimation algorithm is constructed.After verified by experiment,the ADD(-S)accuracy rate of the object pose estimation algorithm in this paper on the LINEMOD and Occasion LINEMOD datasets reaches 89.7%and 40.8%respectively,which is better than other recent algorithms.Also,the algorithm can run at a maximum of 25fps,which means it is good in performance and speed.2.Aiming at the problem that the algorithm mentioned above is prone to produce large errors in the pose estimation of the obstructed objects and lead to low stability,this paper adds a translation component regression branch to the heatmap regression network.Meanwhile,based on the nonlinear optimization of heatmaps,the object pose is cultivated combined with the keypoints and translational components.After experiment verification,the accuracy of ADD(-S)improved by 5.5%compared with the original algorithm in the Occasion LINEMOD dataset,and the deviation from a wrong estimation can be effectively reduced to maintain the algorithm’s stableness.3.On the basis of the pose estimation algorithm based on keypoints mentioned before,conduct the pose estimation study directing at the non-cooperative satellite objects in the space images.This paper firstly trains yolo-v5 object detection network to detect and extract the satellite objects in the space images.Then,the heatmap regression network in the preceding part of this paper is improved,enhancing the perception ability of heatmap regression network to image overall situation features by adding CBAM attention module to the network.Carrying out keypoint positioning to the satellite image module.At last,completing the satellite object pose calculation utilizing nonlinear optimization method based on heatmap.Eventually,pose estimation algorithm for noncooperative satellite scored 0.0193 in the Kelvins Pose Estimation Challenge,standing at the fourth place of all the algorithms.