Structure modifiable adaptive reason-building temporal bayesian network (smartbn): theory and application in human activity and three-dimensional vehicle modeling from video
Bir Bhanu, Nirmalya Ghosh · 2007
This dissertation proposes a novel data-driven continuously evolvable Bayesian Net (BN), namely Structure Modifiable Adaptive Reason-building Temporal BN (SmartBN). SmartBN considers very few causal relations, called causal templates, models the current evidence online by Expansion and Instantiation (EI) principle and keep on continuously self-modifying with the streaming information. It is the only current scalable probabilistic graphical framework to effectively model unpredictable dynamic systems or processes by systematic and incremental uncertainty handling. Several application areas, from economics to medical decision systems to database management to computer vision, are distinguished. Performance of SmartBN is reported on two computer vision applications, human activity analysis in streaming movie clips and incremental 3D model reconstruction of vehicles from unpredictable traffic video. SmartBN performance is shown for a complex movie clips with multiple characters and multiple activities based on body-part positions, relations among body-parts and changing relations across the frames. SmartBN based 3D model building uses a novel concept of Directional Template Library (DTL), one single generic 3D vehicle model and several videos of different vehicles in motion. DTL maps 2D features to 3D model parameters. EI principle uses the identified causal templates to instantiate the SmartBN online and keep on evolving it for new features and relations encountered. Performance of SmartBN in 3D model building is also compared with an incremental adaptive clustering based approach proposed in this work. The clustering based approach uses DTL and novel structural congruency (to the generic model) scores at multiple abstraction levels to incrementally reconstruct the 3D model of the vehicles. The SmartBN and clustering based approaches are compared for 3D model building application. SmartBN widens the scope of probabilistic graphical frameworks manifold, specifically the video-based computer vision systems.