The Recognition Of Moved Object With RGB-D Data

Object recognition has always been a significant topic in the field of robot vision. Especially the indoor recognition is an important developing direction. This thesis is built on the probability model that uses a Bayes method based on the Markov assumption. Through analyzing the different probability calculation methods in different status, eventually recognize the moved objects.This thesis makes the two sets of point cloud as the point of penetrate, and analyze the method to make the space registration and correspondence. Considering the problem from the angle of precision and efficiency, we adopt the method that combines the SURF and ICP to register the two sets of scenes that have many differences in space, and make the comparative analysis of the result of the match. We use the plane extract method and the fitting method that based on the zone broadening to get the largest plane when we finished matching the edge of the scene, and then label the data.In order to solve the reconstruction problem and try to improve the effectiveness of the algorithm, we project the data to the spatial grid with the space division method and construct the point model and surface element model. Through the merging of the surfaces we could both reduce the cavity and the overlapping region, and optimize the reconstruction of the object. In order to achieve the judgment about the status of the object, we transform the probability model with the iteration Bayes formula based on the Markov random assumption. We combine the Gaussian distribution, the Poisson’s distribution and the even distribution together to carry out the calculation. Then based on the MRF for optimizing the judgment the surface with its neighborhood surfaces, we successfully realize the recognition of the moved object in the3D scene. Through analyzing the result of the experiment we testify the effectiveness of the method.