Probabilistic variational methods for vision based complex motion analysis

Ying Wu,Gang Hua

2006-01-01

Abstract:Many emerging applications, including intelligent vision based human computer interaction, intelligent video surveillance, virtual and augmented reality, animation, and biomedical image analysis for computer aided diagnosis and surgery, demand effective and efficient vision-based methods to analyze complex motions, such as articulated motion, deformable motion, and multiple motions. The fundamental challenges of this inverse problem come from two aspects: the high degrees of freedom in these complex motions, and the complications in the image measurements. The high dimensionality of this problem has plagued the scalability and efficiency of many existing methods. In search for a new and scalable solution that overcomes the curse of dimensionality, we view this problem from another angle and conjecture that the complexity of such a problem can be approached by the collaboration among a set of low dimensional motion estimators. Targeting on the two fundamental challenges, in this dissertation, we propose a distributed and collaborative probabilistic reasoning framework for complex motion analysis. The theoretic foundation of the proposed approach is based on probabilistic graphical models. The probabilistic variational inference on these graphical models clearly reveals a set of simple interactive Bayesian motion estimators, which obtains the optimal solution in a collaborative fashion. All computations in the Bayesian inference can be efficiently performed in a distributed and parallel way. The proposed collaborative approach is also a distributed visual measurement integration framework, which handles measurement uncertainty, measurement multi-modality, and measurement inconsistency in a principled way. Extensive experimental results on analyzing different complex motions demonstrate the effectiveness, efficiency, scalability and robustness of the proposed collaborative approach.