Support Relationship Extraction Algorithm Based On RGB-D Images

Support relationship extraction in scenes is an important part of structure analysis. Through support relationship extraction, the relationship between different target objects can be obtained, which is meaningful for scene understanding. As the development of depth capture devices, the depth information of a scene can be easily obtained. The color image with depth information is called RGB-D image. If the depth information can also be used for scene structure extraction, then the physical relationship between objects can be inferred more easily. At the same time, it meets the needs of human visual system. Therefore, the support relationship extraction in scenes based on RGB-D is a very meaningful research area. Image segmentation, feature classification based on segmentation area, the scene structure relationship model building and relationship extraction are the core problems to be solved.The main work of this dissertation are listed as follows:(1) Some researches are carried out based on the core problems in support relationship extraction using RGB-D images, which include:the acquisition of RGB-D images, the method of image segmentation, the building of structure level and support relationship.(2) Hierarchical structure classification is carried out in different segmentation areas using feature descriptor after image segmentation. Based on the traditional SIFT feature descriptor, using kernel principal component analysis method for dimensionality reduction, a new descriptor is proposed based on the new kernel function, and a better result is achieved.(3) The support relationship extraction method in reference [16] is improved. First, a more accurate scene hierarchical structure and support relationship model is proposed. Then, the images are segmented using decision tree trained with another 5 different features added, and the structure level is obtained using our improved descriptor as region features. Finally, the support relationship is calculated using linear planning algorithm. The experiment result is analyzed based on the evaluation rules.