Research On Object Recognition And Pose Estimation Based On Multi-feature Model Base

With the development of global aging,family service robots have attracted more and more attention.In complex family environment,the ability to quickly and accurately identify and locate the target object is the key index to evaluate the advantages and disadvantages of service robot.The recognition algorithm based on point cloud features has a high recognition rate in the face of interference such as weak illumination and poor texture,but the recognition rate is greatly reduced in the face of occlusion or stacked scenes.To solve this problem,an object recognition and pose estimation method based on multi-feature model base is proposed in this thesis.In the aspect of recognition,the model base with single feature is improved to improve the recognition rate of occlusion or overlapping environment.In the aspect of pose estimation,this thesis optimized the selection of local feature key points,which effectively improved the problem of large error in pose estimation.In terms of hardware,this thesis adopts the service robot independently developed by the Rehabilitation Robot Laboratory of our university.The robot is equipped with two high-resolution depth cameras.Multiple cameras will help the robot to recognize and estimate the position and pose of different occlusion or overlapping scenes.The main research work is as follows:Aiming at the problem of low object recognition rate in complex family environment,a recognition algorithm based on multi-feature model base is proposed in this thesis.In order to improve the recognition rate of occlusion or overlapping scenes,local Feature FPFH(Fast Point Feature Histogram)and global Feature VFH(Viewpoint Feature Histogram)are extracted during offline modeling.In order to improve the efficiency of online recognition,the global feature VFH is used for fast recognition.When the occlusion or overlap is encountered,the subsequent recognition tasks are completed on the basis of the global feature recognition.Experimental results show that the proposed algorithm can quickly and accurately identify the occlusion or overlapping scenes.Aiming at the problem of poor accuracy of pose estimation in complex environment,a pose estimation algorithm based on optimized local feature descriptor is proposed in this thesis.Firstly,the more stable ISS3D(Intrinsic Shape Signature 3D)is selected as the key point of local feature calculation to ensure the accuracy of late registration.In order to eliminate redundant key points,the key points of the target point cloud are filtered to improve the computational efficiency of local feature descriptor.Experimental results show that the accuracy of the proposed algorithm is improved significantly.Based on the above methods,a comparative experiment is carried out on the real scene recognition and pose estimation of the service robot vision system.The experimental results show that the proposed algorithm can accurately identify and estimate the position and pose of the object in the occlusion or overlapping environment.