Research On 3D Object Recognition In Cluttered Scenes With Local Surface Feature Descriptor

Object recognition in cluttered scenes is a basic research in the field of computer vision.It has important applications in numerous fields,e.g.,intelligent surveillance,automatic assembly,remote sensing,mobile operation,robotics,bioanalysis and medical treatment.Compared to the traditional 2D image,the 3D range image can provide more geometrical information,and the 3D features are not affected by the scale,rotation and illumination.Besides,6DoF(six degree of freedom)transformation information corresponding to the object can be obtained by the 3D range image.Therefore,the research of object recognition with 3D range image has potential advantages.In addition,in recent years,a large reduction in cost of the 3D data acquisition scanners and the rapid-improvement of the computing hardwares greatly improve the time efficiency of 3D object recognition.All these factors make the 3D object recognition have irreplaceable value in the field of computer vision,and promote it to become the current-popular research.However,in practical applications,due to the defects of the hardware equipments,the 3D range scenes contain noise,sparse point cloud,clutter and occlusion,so the current research of 3D object recognition is still a very difficult challenge and task.Therefore,it is imminent to improve the description ability of 3D feature descriptors and solve the problem of 3D object recognition in cluttered scenes.Aiming at above issues,this dissertation analyzes and studies the problem of object recognition in cluttered scenes with local feature descriptors.The main innovative works of this dissertation are as follows:1.Aiming at the problems of feature mismatching in the cluttered scenes which include rotation,translation,noise,sparsity,occlusion and clutter,this dissertation begins with the research and analysis of the 3D distribution information and geometry information of the 3D objects and scenes,and proposes Feature Fusion Information Statistics(FFIS).In the process of generating feature descriptors,in order to improve the invariance ability of rotation and translation of the feature descriptors,this dissertation firstly constructs a Local Reference Frame(LRF)on a Local Surface Patch(LSP).Unlike previous methods,this dissertation constructs the LRF by projecting the eigenvectors of LSP's scatter matrix to the plane which is perpendicular to the normal vector of the LSP.Based on this LRF,we combine mesh distribution information and point distribution information to generate FFIS feature descriptors,and this results in the easy implementation and the high efficiency of the feature descriptors' computation.Recall vs 1-Precision curve is adopted as an algorithm evaluation criterion to verify the effectiveness of the algorithm in the cluttered scenes.The experimental results show that the FFIS method can extract more correct feature matching corresponding pairs in cluttered scenes than other methods(Spin Image,3DSC,FPFH,PFH,USC,RoPS,and SHOT).Especially in the noise-free scenes,FFIS achieved a 95%Recall rate.2.In view of the weak ability of feature to describe object in cluttered scenes,and in order to improve the accuracy of 3D feature matching in cluttered scenes,Histograms of Gaussian Normal Distribution(HGND)is proposed.Firstly,we adopt the two eigen-vectors of LSP's scatter matrix to construct LRF to obtain the rotation and translation invariance for the feature.The HGND feature descriptors are based on geometrical and spatial information,this method generates HGND feature descriptors through the geometrical projection distribution of the spatial distribution points' corresponding nor-mal vectors,which ensures the descriptiveness and robustness of HGND in cluttered scenes.Through the experiments to compare with Spin Image,3DSC,FPFH,PFH,USC,RoPS and SHOT,the HGND method's descriptiveness in different scenes is ver-ified.It also proves that the HGND method can effectively improve the accuracy of 3D feature matching in cluttered scenes.Especially in low noise scenes,HGND achieves a 90%rate of Recall.In addition,FFIS has obtained the best performance in terms of computation efficiency.3.To solve the problem of 3D object recognition and its pose estimation in cluttered scenes which include a large number of false feature matching corresponding pairs,high occlusion levels and sparse point clouds,this dissertation proposes the Hypothesis Generation(HG)method to extract potential object hypothesis instances in the scene.Unlike the existing methods,the HG proposed in this dissertation is based on the Hough Transform and Hough space,and the "self-adapted-tuning" method is used to assist HG to generate potential object hypothesis instances.Based on this HG,this dissertation presents the Hypothesis Verification(HV)based on the ICP(Iterative Closest Point)with double validation processes to verify the potential hypothesis instances generated by the HG.Eventually,only when the hypothesis instances pass through the HV processes,which are identified as the correct hypotheses.The experimental results show that the HG and HV methods proposed in this dissertation can correctly identify the object and its corresponding 6DOF pose information in the cluttered scene with high FP(False Positive)rates,high occlusion levels and sparse point clouds.Especially when the FP rate is less than 95%,the HG and HV methods can nearly achieve a recognition rate of 100%in cluttered scenes.The method proposed in this dissertation realizes the feature matching between the objects and cluttered scenes by the local feature descriptor,which solve the problem of 3D object recognition and its pose estimation in cluttered scene.Our method reduces the complexity of the algorithms,improves the efficiency of the methods,and achieves a high accuracy of object recognition,it results in a good future application in the computer vision research area.