Three-Dimension Object Detection And Localization Based On Point Cloud Registration

Three-dimension object recognition is an important research field in computer vision due to its widely application in robot localization and navigation,3D scene perception and understanding,unmanned automatic vehicle driving,augmented reality,etc.A key enabling factor for the development of this technology is represented by the availability of low-cost high-accuracy 3D sensors,such as the widely used Microsoft Kinect and Asus Xtion devices.Due to the complexity of the real scenes with noise,occlusions and clutter,3D object recognition is facing a great challenge.The key issues of 3D object recognition is registration.The more complex the scene is,the larger of amount of more the mismatching correspondences is.So how to deal with point cloud registration in complex 3D scenes effectively and accurately has become the focus study towards 3D object recognition.An improved 3D object detection and localization method is proposed.The method achieves pose estimations of multiple object instances in 3D scenes with some occlusions and clutter.First,the normal vector of point is estimated by computing covariance matrix considering distance between the neighboring points and the feature one within the local spherical support domain.The longer the distance is,the smaller the weight is.The covariance matrix with distance weights is calculated to obtain more precisely point normal vectors.Next,based on improved normal vectors,the point cloud descriptors are decoded,called color signatures of histogram of orientations(C-SHOT).Then,to prove the existence of the multi objects being sought on 3D hough voting space,3D feature correspondences between scenes and models need to match using k-dimension tree(KdTree).Finally,we eliminate wrong feature correspondences and compute rough transformation matrix using random sample consensus(RANSAC).When reliable feature correspondences have been selected,a final transformation matrix based on levenberg marquardt iterative closest point(LM-ICP),can be performed to further refine pose estimations.Experiments on publicly available datasets as well as on real lab 3D data obtained with Kinect v2 provide a thorough validation of the proposed method,and the result demonstrates the recognition accuracy and effectiveness of the proposed method.