Project Report on Human Detection and Body Tracking

The estimation of human motion from a monocular sequence of 2D images is difficult for many reasons: the non-linear dynamics of the limbs, ambiguities in the mapping from the 2D image to the 3D model, self occlusions, the similarity of the appearance of different limbs, and image noise etc. In this project, we learn and model background scenes statistically to detect foreground objects, distinguish people from other objects by checking skin-color blobs in the foregrounds. After we estimate the position of the body in the foreground region, we apply a Bayesian approach to track body motion, which is to compute the posterior probability distribution over the parameters of the 2D cardboard body model and dynamic model. Since the project has to be finished in 4 or 5 weeks, we make many simplifications in the realization. There are still many ideas that we plan to add into the current framework in the future.