3D Human Motion Tracking and Pose Estimation using Probabilistic Activity Models

Darby, John (2010) 3D Human Motion Tracking and Pose Estimation using Probabilistic Activity Models. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (13MB) | Preview

Abstract

This thesis presents work on generative approaches to human motion tracking and pose estimation where a geometric model of the human body is used for comparison with observations. The existing generative tracking literature can be quite clearly divided between two groups. First, approaches that attempt to solve a difficult high-dimensional inference problem in the body model’s full or ambient pose space, recovering freeform or unknown activity. Second, approaches that restrict inference to a low-dimensional latent embedding of the full pose space, recovering activity for which training data is available or known activity. Significant advances have been made in each of these subgroups. Given sufficiently rich multiocular observations and plentiful computational resources, highdimensional approaches have been proven to track fast and complex unknown activities robustly. Conversely, low-dimensional approaches have been able to support monocular tracking and to significantly reduce computational costs for the recovery of known activity. However, their competing advantages have – although complementary – remained disjoint. The central aim of this thesis is to combine low- and high-dimensional generative tracking techniques to benefit from the best of both approaches. First, a simple generative tracking approach is proposed for tracking known activities in a latent pose space using only monocular or binocular observations. A hidden Markov model (HMM) is used to provide dynamics and constrain a particle-based search for poses. The ability of the HMM to classify as well as synthesise poses means that the approach naturally extends to the modelling of a number of different known activities in a single joint-activity latent space. Second, an additional low-dimensional approach is introduced to permit transitions between segmented known activity training data by allowing particles to move between activity manifolds. Both low-dimensional approaches are then fairly and efficiently combined with a simultaneous high-dimensional generative tracking task in the ambient pose space. This combination allows for the recovery of sequences containing multiple known and unknown human activities at an appropriate (dynamic) computational cost. Finally, a rich hierarchical embedding of the ambient pose space is investigated. This representation allows inference to progress from a single full-body or global non-linear latent pose space, through a number of gradually smaller part-based latent models, to the full ambient pose space. By preserving long-range correlations present in training data, the positions of occluded limbs can be inferred during tracking. Alternatively, by breaking the implied coordination between part-based models novel activity combinations, or composite activity, may be recovered.