A Large Number Of Object Recognition Based On 3D Point Cloud Model

The development of information technology makes object recognition from images an important research topic,with a variety of application scenarios.Meanwhile,the acquisition of 3D information has become increasingly convenient,due to the explosion of 2D image data and the maturity of 3D reconstruction technology.A 3D point cloud model reconstructed from a series of unordered 2D images is a common type of 3D information,which describes the spatial structure and local appearance of 3D objects.3D point cloud models provide a comprehensive representation of object visual information and thus can effectively support object recognition from images.In this paper,we study how to leverage 3D point cloud models to perform fast and accurate object recognition in 2D images.The traditional methods first find 2D-3D correspondences between 2D image and 3D models based on local visual features,and then identify the correct correspondences by camera pose estimation to achieve object recognition.However,the efficiency of these methods usually suffers from the noisy 2D-3D correspondences caused by the increasing number of target objects,which make the computational cost of pose estimation unbearable.In order to break this scalability bottleneck,this paper proposes to filter 2D-3D correspondences efficiently.Starting from 2D and 3D geometric information,the method consists of two stages: lightweight local filtering and fine-grained global filtering.The combination of such two stages filters the noise correspondences,and thus improves both the accuracy and the efficiency of pose estimation.With this method as the core,we have further developed a 3D object recognition system that can achieve fast and accurate object recognition.The quantification experiment is based on a self-built 3D model dataset containing 300 target objects and a test dataset of 200 query images.The experimental results show that the method proposed in this paper can effectively improve the quality of 2D-3D correspondences,thereby improving the effectiveness and efficiency of object recognition.