Research On Key Technology Of RGB-D Image Analysis

Image Analysis is a process of describing, explaining, classifying or inferring thecontent of scenes revealed by images via pattern recognition or artificial intelligencemethods, and now has been widely used in fields such as security surveillance,industrial detection, military remote sensing, medical imaging, etc. Image analysistechnology is in a booming period and a series of research findings have beenachieved. Rapid development especially the expansion of3D videos and the birth ofconsumer depth cameras, brings new energy to the field of image analysis. Undersuch a background, the problem how to use combined RGB and depth information inimage analysis is deeply researched in this thesis.This thesis begins with an overview of image analysis field. Basic knowledge andthe principle of image analysis are introduced in detail, including the system of imageanalysis and common image features and classifiers. Then, the popular consumerdepth camera Kinect is introduced. The mechanism of depth sensing and the storageformat of depth information are explained later. Corresponding to supervised learningand unsupervised learning in image analysis, two directions are researchedrespectively, namely, human head pose estimation based on depth information andrandom regression forests and image region segmentation based on stereoscopicsaliency.When discussing human head pose estimation, this thesis introduces the trainingand testing of random regression forests predictor and then using Kinect to capturedepth map of human beings. Finally, the location of head center and direction ofhuman faces are estimated in real time. The accuracy is pretty high and the speed isacceptable. In the research of image region segmentation based on stereoscopicsaliency, global region contrast saliency model based on RGB color image isextended using depth pre-segmentation. Meanwhile, the extended method iscombined with a depth weighted stereoscopic saliency model to compute the saliencyof scenes. Experimental results show that the proposed method achieves a goodperformance and the segmented object region is accurate.