3D Object Recognition System With Local Shape Description And Correspondence Selection

Object recognition is one of the most important tasks in computer vision,and has been applied in many areas,such as intelligent robotics,self-driving car and surveillance security.Due to the limitation of RGB images,such as sensitive to lighting,traditional 2D object recognition algorithm can not guarantee the reliability.On the contrary,3D object recognition is more reliable due to its ability to validate the recognition results by considering the perfectness of shape registration.With recent advances in 3D ge-ometry acquisition technology,it becomes flexible to acquire depth scans.Thus,object recognition working with 3D data(e.g.,depth scans)receives increasing attention.The aim of 3D object recognition is to recognize the object models in the scene while recover their poses.There still remains many issues and challenges to build a practical 3D object recognition system.On the one hand,point cloud data is very sensi-tive to many interference,such as noise,varying point density and occlusion,thus affect the accuracy of 3D object recognition.On the other hand,due to many feature corre-spondences are not right,and verification itself is time-consuming,thus strictly restrict recognition efficiency.In this paper,we focus on recognition of objects represented in partial shapes in cluttered scenes.To achieve this,we propose an efficient yet robust 3D object recog-nition system,including extracting local shape description,choosing salient features to generate matches and selecting more reliable matches by applying geometric con-straints.Details are as follows:First,on the problem of local shape description,we design a novel "Signature of Geometric Centroids(SGC)" descriptor,which is robust under various conditions such as noise,varying point density and occlusion,and propose a descriptor comparison scheme to support matching two incomplete surfaces.Second,on the problem of partial shape matching,we propose a novel descriptor-graph in which each node has K neighbours to accelerate descriptor matching.In ad-dition,by extracting saliency information from descriptor-graph,and applying them to partial shape matching,thus further improve the accuracy of descriptor matching.Third,on the problem of correspondence selection,we define a novel compati-ble measure between two feature correspondence and introduce an auxiliary set to im-plement the "Auxiliary Set Voting" algorithm,thus achieve effective correspondence selection.Targeting at practical applications,all proposed solutions in this dissertation not only achieve highly robustness,while maintain efficiency.Therefore,all these solutions have practical value.