Multi-Object Target Recognition And Pose Estimation In Clutter

With the development of science and technology and the rapid improvement of people's living standards,service robots are more and more popular among people.Equipping with multi-freedom mechanical arms and being capable of object grabbing have become one of the standard functional configurations for service robots in the fileds of home service,warehousing and logistics,elderly and disability care.However,unlike the repetitive grabbing of robot arms in industrial production line,one of the key problems for object grabbing of service robots is to identify objects and make pose estimation correctly.In this paper,object recognition and positioning of multi-DOF manipulator grab is studied,which is of positive significance and engineering value to guarantee the accurate motion of service robot.In this paper,firstly,a framework for object recognition and positioning based on sparse features and neural network is presented,then a multi-view RGB-D image database is generated based on the model in dataset,and the feasibility of multiple objects recognition and pose estimation of two schemes is analyzed.Aiming at object recognition and pose estimation based on sparse features,a preprocessing operation was carried out on the disordered desktop with objects placed,the desktop was removed by the random sampling consistency algorithm,and the key points were captured with the down sampling algorithm.In order to improve the accuracy of pose estimation,multiple existing local feature descriptors were fused and the best feature fusion method is selected for feature matching.In the phase of pose estimation,DP clustering algorithm is introduced to improve the robustness of rotating subgroup voting,and non-maximal suppression algorithm is used to determine the best posture.Test results on public datasets show that this algorithm is very feasible.Aiming at object recognition and pose estimation based on neural network,the method based on convolutional neural network does not consider the spatial local correlation of point cloud.A semantic segmentation model based on convolutiondeconvolution is established by using dilated convolution to expand the receptive field of the point convolution layer.Then,the regression vector of the network is determined,and the rotation regression vector is determined by Hough voting method.On this basis of above,the feasibility of point convolution neural network for object recognition and pose estimation is verified through experiments.In this paper,a new improved scheme is used to solve the regression vector problem of point convolution neural network model.The determination of object rotation translation vector is separated from the network.Control points are used as regression vectors and guided by semantic segmentation information,and ICP algorithm is used to determine object posture.Aiming at the pose estimation problem of the actual scene,this paper uses 3d structured light scanner to establish our own model library,captures the test scene through RGB-D sensor,and tests two kinds of pose estimation algorithms in this paper with this dataset.Finally,the research work of this paper is summarized,and the future research work and direction is prospected.