Estimating 3-D human body poses from 2-D static images

Our objective is to estimate 3-D human body poses from single 2-D static images. This task is difficult due to the influence of numerous real-world factors such as shading, image noise, occlusions, background clutter and the inherent loss of depth information when a scene is captured onto a 2-D image. We propose a novel fusion of two techniques to form a two-step process: in image preprocessing, an algorithm based on image segmentation and the evaluation of visual cues is used to find immediately identifiable body parts, which we consolidate into 'proposal maps'. This is then fed to a Data Driven Markov Chain Monte Carlo (DDMCMC) pose estimation technique to explore the high dimensional solution space. The best 3-D body pose is then estimated by the Maximum a Posteriori solution. Experimental results show that the DDMCMC is highly accurate in converging to the true solution when given ideal proposal maps. The results show that the DDMCMC is able to converge to the true solution, albeit with some errors. Nevertheless, the technique shows promise in inferring 3-D body poses. We are currently exploring improvements such as a more accurate model of the human body, the ability to estimate poses from images with cluttered backgrounds and improvement in recognition speed.