The Research On The Feature Description And Registration For RGB-D Images

With the continuous progress of image sensing technology,there are multi-ways to obtain images and their category has been extended from the traditional 2D images to the 2.5D depth images and 3D point clouds.It is known that local feature description is always the fundamental core research topic in computer vision.It has been widely used in several areas including image registration,image mosaic,object recognition,object tracking,and 3D modeling.As for local feature description,many outstanding algorithms proposed by domestic and overseas researchers have shown strong robustness with respect to scale,rotation and illumination,and they have been applied in real world.However,most descriptors only use information of 2D images or 3D point clouds and are represented by high-dimensional float vectors.With the rapid development of portable smart devices and the increasing requirement for efficiency,it is necessary to study discriminative local feature descriptors with low memory consumption and high matching efficiency.Meanwhile,as for image registration,obtaining high registration accuracy is always the direction of effort.Motivated by the new requirements,this paper investigates the description and registration of RGB-D images.(1)This paper proposed a local feature descriptor to combine both RGB information and geometric information.Benefitted by the rapid calculation of local binary pattern and the intrinsic characteristics of containing both RGB information and geometric information for RGB-D images,the proposed algorithm uses a bit string to construct a local feature descriptor.In details,after estimating the scale and orientation of a feature point,we extract a certain number of sampling point pairs in the neighborhood of the feature point.Then,for each point pair,we use 2 bits for representing the constructed two results about RGB information and 1 bit for geometric information.Finally,we concatenate all these bits for all point pairs.The experimental results show that the proposed descriptor is discriminative and occupies less memory and matches rapidly.(2)This paper designed a coarse-to-fine registration method based on the proposed descriptor for RGB-D images.By transforming the depth image to a 3D point cloud,this paper transforms the depth image registration problem to 3D point cloud registration.Before coarse registration,we compute the descriptor on two point clouds using our proposed method and obtain the matched feature point pairs.Then,we obtain the initial transformation matrix through the matched feature point pairs.Finally,in order to get better registration result,we apply the ICP algorithms.The experimental results show that our proposed descriptor can get good registration accuracy.Besides,the experimental results also show that our designed coarse-to-fine registration method can further improve the registration accuracy.