Research On 3D Target Pose Estimation Approach Based On Vision Multi-feature Fusion

The development and production of large-scale equipment in the industrial manufacturing fields such as aerospace and shipbuilding require the three-dimensional measurement of various large components.Due to the complex structure and large size of the components,cooperative targets are usually used to assist the measurement for 3D parameters.In the process of measurement,the measuring instrument should aim at the target on the surface of the components,which can be implemented by obtaining the target image and estimating its pose.Affected by the measurement environment,the target image is usually defocused,and because the imaging system is far away from the target,the imaging area of it in the image is very small,and the spatial resolution of the area is low,which results in the low quality and resolution of the target image.The stability and accuracy of the pose estimation results by the existing pose estimation methods for such images will be seriously degraded,and difficult to aim at the target accurately.Therefore,researching the pose estimation method for low-quality and low-resolution target images to improve the robustness and accuracy of the pose estimation algorithm is of great significance and application value for the realization of cooperative target aiming and 3D geometric parameters precise measurement of large components.Based on the analysis of the impact of low-quality and low-resolution on the stability of pose estimation,this thesis proposes a method of target detection and image restoration based on CAD model and an improved weighted EPnP(wEPnP)pose estimation method based on multi-feature fusion.First,aiming at the problem that low-quality images will overwhelm target edge features and reduce the accuracy of feature extraction and pose estimation,the CAD model of the target is used for target detection to effectively segment the its imaging area,and then the target area are restored based on the feature information in the CAD model to enhance the edge features of the target and provide accurate and reliable data source for feature extraction.Secondly,analyze the impact of low resolution for the target area to feature extraction and pose estimation.In the EPnP process,the reliability of feature points is considered,and the importance of different points is adjusted by weight to reduce the influence of the feature extraction error on the stability and accuracy of pose estimation.The reliabilities and weights of feature points are mainly measured and determined by the line and surface features,and then we can realize the wEPnP pose estimation method by fusing the point,line and surface features of the target.The experiments are implemented by the 3D target images and low-resolution simulation images to verify the proposed method,the results show that the proposed CAD model-based target detection and image restoration method can effectively segment the target area and enhance its edge features.The experimental results of low-resolution simulation image pose estimation show that the wEPnP algorithm is more robust,when the performances of the EPnP are degraded,the wEPnP algorithm can still effectively estimate the poses of the targets,and the estimation errors of rotation matrix and translation vector of the wEPnP are decreased by 3.56%-4.47%and 1.79%-2.30%respectively compared to EPnP,experimental results of actual 3D target images show that the CAD model-based target detection and image restoration method and the wEPnP pose estimation method proposed in this paper can realize the pose estimation for low-quality and low-resolution target image,and they can be used for high-precision aiming of 3D targets.