On the Structure of Dynamic Bayesian Networks for Complex Scene Modelling
We introduce the idea of constructing Dynamic Bayesian Networks (DBNs) with hierarchical structures for modelling complex scenes at both the event level and the activity level simultaneously. Practical issues regarding the structure design of a DBN with multiple hidden processes and hierarchical str...
| Main Authors: | , , |
|---|---|
| Other Authors: | |
| Format: | Text |
| Language: | English |
| Published: |
2003
|
| Subjects: | |
| Online Access: | http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.432.7491 http://www.eecs.qmul.ac.uk/~sgg/papers/XiangEtAl_JVS2003.pdf |
| Summary: | We introduce the idea of constructing Dynamic Bayesian Networks (DBNs) with hierarchical structures for modelling complex scenes at both the event level and the activity level simultaneously. Practical issues regarding the structure design of a DBN with multiple hidden processes and hierarchical structure are identified and discussed. Experiments are presented to compare a Multi-Observation Hidden Markov Model (MOHMM), a Hierarchical MOHMM, a Hierarchical Dynamically Multi-Linked Hidden Markov Model (DML-HMM), and a Hierarchical 2-layer DML-HMM (2L-DML-HMM) for complex scene modelling. It is demonstrated that only the Multi-Observation Hidden Markov Model is able to perform meaningful factorisation in the activity state space and to extract the deterministic temporal structure of activities occurred in a complex dynamic scene. 1. |
|---|