RGB-D Based 3D Object Recognition And Pose Estimation

In the field of computer vision,object recognition and pose estimation in scene play an important role and have great application prospects.This technology is widely used in the fields of technology and industry,such as manipulator grasping,automatic navigation and intelligent monitoring.However,complex environmental conditions such as illumination change and occlusion will greatly affect the accuracy of object recognition and pose estimation.In addition,in practical applications,the algorithm needs to ensure high accuracy and real-time,and needs to be further studied.Compared with ordinary RGB imaging equipment,RGB-D sensing equipment can provide abundant 3D information.Researches on RGB-D-based correlation algorithms are expected to achieve better 3D object recognition and pose estimation under weak texture and low illumination conditions.Classic 3D object recognition algorithm achieves object recognition through the matching of 3D descriptors.However,the 3D descriptor can only be used for the recognition of a single target,can not achieve overall description to the objects with complex structure.The neural network-based method is also a research hotspot for object recognition and pose estimation.This kind of work is aimed at known objects,requires a large amount of target data for training,and can not achieve pose estimation for unknown object.In addition,it also needs complex post-optimization steps to optimize the final results,which makes it difficult to meet some high real-time application requirements.In view of the above proposed problems,this paper focuses on the 3D recognition of object with complex structure and real-time unknown target pose estimation.The specific work is as follows:1.A 3D structure descriptor based on spatial correlation is proposed to identify 3D objects with complex structures.The real-time dense SLAM reconstruction of 3D scene is carried out,and the reconstructed scene is segmented based on concavity and convexity.On this basis,the real point cloud of the target is obtained,and then the spatial correlation between the target component and the attribute information of the component itself is extracted to form an overall description of the 3D object.Finally,3D objects with complex structure are identified by matching descriptors.Experiments on self-built data sets show that the proposed 3D structure descriptor has good object recognition effect.2.A real-time pose estimation method for unknown targets across the field of view is proposed,which can be used for real-time pose estimation and 3D point cloud sparse reconstruction of unknown targets in large scenes.This method obtains the real pose of the object relative to the initial camera through specific visual features AprilTag,and then combines the cross-field tracking module to fuse multiple RGB-D information to obtain the optimal input data,calculates the current pose of the camera based on ORB feature point matching,and finally converts the pose of the camera into six-degree-of-freedom pose output of the object.The quantitative analysis of the proposed method and the existing pose estimation method are carried out on the simulation dataset.The results show that the performance of the proposed algorithm is significantly improved compared with the existing algorithm.The qualitative analysis of the real scene also shows the effectiveness of the proposed algorithm.At the end of this paper,the paper summarizes the work of the full text,and looks forward to the next step,including further improving the generalization recognition performance of the 3D structure descriptor,and integrating the object recognition and pose estimation methods proposed in this paper,which is further applied to the robot arm sorting,intelligent monitoring and other scenarios.