Research On Feature Point Detection And Point Cloud Matching Algorithm For Indoor Scene Image

With the rapid development of computer vision,3D reconstruction of scenes has been widely used in the computer field,and the difficulty of 3D reconstruction lies in the registration of point clouds.The feature-based registration method is the most commonly used registration method.The key is to extract the feature information of the point cloud and then calculate the feature similarity,and register it with the similarity information of the point cloud.The traditional 3D reconstruction algorithm has the disadvantages of complicated point cloud process,slow operation speed and low matching precision.In this paper,the Kinect depth sensor is used as the acquisition device for depth information and color information.The data acquisition,preprocessing,feature point extraction and rough matching,and point cloud exact matching process involved in the 3D reconstruction process are studied and theoretically analyzed.A 3D reconstruction scheme based on the Kinect depth sensor.(1)A fast and high precision NCC image matching algorithm is proposed.By constructing the image wavelet pyramid structure,the feature point search matching time is reduced,and the NCC image matching algorithm is used to match to obtain the rough matching points on the original image.Then the improved relaxation iterative algorithm is used to remove the mismatched points,and the adaptive part winner is introduced.Take all-in-one strategy,increase the iteration speed,and get more one-to-one matching point pairs.Then,the form of calculating the weight coefficient in the FPFH feature extraction algorithm is modified.The exponential function is used instead of the original reciprocal form,and the weight value is standardized within a reasonable interval,which improves the problem of weight coefficient overflow in the original algorithm.(2)An improved ICP matching algorithm is proposed to divide the vertebral body region of the camera.The Euclidean distance formula is used to assign the corresponding weight coefficient to the feature points existing in each region,and the weight coefficient of the corresponding point is taken.The solution into the mean square error function effectively improves the quality of the three-dimensional reconstruction result obtained by continuous multi-frame image after point cloud registration.Then,the traditional ICP algorithm consumes too much time,and the BBF algorithm accelerates the ICP,which can reduce The registration time greatly improves the registration efficiency of the 3D point cloud.The experimental data shows that the three-dimensional reconstruction scheme proposed in this paper can use ordinary personal computers and Kinect to digitize common objective objects.The reconstruction process of this system is simple and easy to operate,and the required cost is low.No high-performance registration computer is needed.Compared with the original algorithm,the proposed algorithm has a great improvement in efficiency and precision,which proves the effectiveness of the proposed algorithm.