Statistical inference with learning for temporal data analysis in visual computing.

摘要

Analyzing temporal visual data is an important task in computer vision. The fundamental challenges of the task mainly come from three aspects: the high dimensionality of the data, the complexity of the temporal dynamics, and the imperfect image measurements. This research focuses on developing novel computational methods for addressing the challenges in the context of motion tracking and classification in video.;In 2D visual tracking, this research proposes to explicitly model the dependency among different state variables so as to reduce the sampling burden of the Sequential Monte Carlo (SMC) approaches. This results in a novel statistical tracking framework, Rao-Blackwellised Particle Filter (RBPF), which exploits the dependency to marginalize the tractable state components, leaving only the remaining state variables being estimated by the SMC method, leading to increased estimation accuracy and efficiency. Meanwhile, the approach greatly alleviates the "curse of dimensionality" and degeneracy problems associated with the traditional Particle Filter.;Further, for tracking articulated human motion in 3D, this dissertation presents a new learning and inference approach, in which the motion correlation between the left-side and the right-side joint angles is exploited to increase the sampling efficiency, with a learning approach based on Partial Least Square Regression being employed to model the left-right motion constraints. A new RBPF algorithm is designed to integrate the learned motion prior with the online measurements. Experiments show that the approach outperforms the state-of-the-art approaches.;Finally, to address the problem of classification of temporal data with strong underlying dynamics, this work proposes a Discriminative Gaussian Process Dynamical Model (D-GPDM) that learns a discriminative probabilistic low dimensional representation of the high-dimensional data. The resultant novel algorithm is found to be able to recover the latent-space dynamics of the data while keeping maximum separation of different classes and thus high classification accuracy can be obtained.;The main contributions of this dissertation lie in the explicit modeling of the dependency among the state variables with learning, the design of the RBPF tracking framework exploiting the dependency models, and the method of temporal data classification that combines dimensionality reduction with dynamic modeling using a latent space approach.