Object Extraction And Geometric Relationship Resoning Of Indoor Scenes From RGB-D Images

The scene parsing is directly related to the robot’s working status. Direct measurement of high precision depth information can make up for the lack of traditional RGB texture visual information, which has great potential and attractive prospects for solving the problem of indoor scene parsing. With the development of sensor devices, it’s convenient to obtain the depth image, based on the RGB-D image of the indoor scene analysis develops rapidly. The research of scene parsing focuses on the scene of the object contour extraction, object segmentation and geometry relation resoning.In this paper, based on the practical application, analyzes the steps of the scene parsing, and makes a deep research on the contour extraction, the object segmentation, and the geometric relationship resoning between objects. The main innovations of this paper are as follows:Firstly, In view of the rich information of RGB-D image, a new method based on disk is proposed to make full use of the brightness, color, and texture and depth information. Due to high differential and over fitting from single decision tree, ignore the the accuracy of the single decision tree to meet the diversity of ensemble model. The similar method for structured label is introduced to approximate the information gain criterion, structured forest method is derived. The binary variables indicate whether the corresponding pixel contains boundary information. A RGB-D image contour extraction method based on structure forest is proposed.Next, on the basis of the above research, this paper proposes tree-based instance segmentation. The global model is trained by using the standard segmentation trees and the biased segmentation trees, and the ILP is built to describe the semantic and the mapping problem. The loss function is used to solve the ILP model. This instance segmentation method can effectively avoid the error segmentation, accurate marking segmentation block.Finally, proposed a more simple and accurate geometry relation resoning between objects in the scene based on the results of the instance segmentation above. Objects in the scene were divided into four structured class, combined with the physical constraints and a priori statistical support relationship, inference problem is transformed into an integer programming problem, redefine the constraints, the use Gurobi linear programming to solve this kind of NP problem.