|
1. Introduction
It is one of very difficult problems to understand a scene based on image information in a large scale and uncertain real world. To overcome it, we propose a method that divides and manages the Bayesian network (BN) modules to solve the complex problem. We also suggest a BN structure design method based on activity feature, which categorize the domain knowledge, and build BN structure with them, then define several inference processes as a behavior selection network (BSN), and combine the BN and BSN as an inference model. We present a learning method of inference model, which is based on BN and logic network model, and aims to update and adapt the model by using the interactive data. We have performed several simulations to evaluate the performance of the proposed methods.
2. Methods
- Modular
approach & Combination for recognition of uncertain situation
Figuire
1. The concept of our research
- Hierarchical
object recognition BN design based on domain knowledge
Figure 2. Design process of BN
-
Behavior
network based BN ensemble combination
Figure 3.
General Concept of BSN based BN ensemble
- Interactive
learning based on logic network
Figure 4.
Interactive learning process
3. Experimental Results - Hierarchical object recognition Bayesian network based on
activity
Figure 5. Designed Presentation Activity BN
- BSN based place and object perception
Figure 6. Designed BSN
for Combination
- Interactive
learning based on logic network
Figure 7. Learned
Logic Network
- Performance Test
Figure 8. Webot
Simulation
Figure 9. Comparison of place recognizing probability in 6 places (corridor, dining room, toilet, seminar room, elevator, office). Correct answer is colored by black.
4. Publications
[1] Y.-S. Song, S.-B. Cho, and I.-H. Suh, "Activity-object
Bayesian networks for detecting occluded objects in uncertain indoor environment,"
Int. Conf. on
Knowledge-Based Intelligent
Information & Engineering Systems, Sep 2005. IF02=0.515
[2] Y.-S. Song, S.-B. Cho, "Hierarchical
Bayesian networks based on activity for localizing hidden target objects in
indoor environment,” Proc. of Korea Computer Congress, vol. 32, no. 1, pp. 616-618, Jul 2005.
[3] K.-S. Hwang, H.-S. Park, S.-B. Cho, "Bayesian
probability and evidence combination for improving scene recognition performance,”Proc.
of Korea Computer Congress, vol. 32, no. 1, pp. 634-636, Jul 2005.
[4] S.-B. Im, S.-B. Cho, "Place and object
recognition in uncertain indoor environments using SIFT and Bayesian networks,”
Proc. of Korea Computer Congress, vol. 32, no. 1, pp. 637-639, Jul 2005.
[5] J.-O. Yoo, S.-B. Cho, "Fuzzy Bayesian
network for fusion of multimodal context information,”Proc. of Korea Computer Congress, vol. 32,
no. 1, pp. 631-633, Jul 2005.
[6] K.-S. Hwang, and S.-B. Cho, "Constrained
learning method of Bayesian network structure for efficient context classification,"
Proc. of Korea Information Science Society, vol. 31, no. 2, pp. 112-114, Oct 2004.
[7] J.-O. Yoo, K.-J. Kim, and S.-B. Cho, "Speciated
evolution of Bayesian networks ensembles for robust inference," Proc. of
Korea Information Science Society, vol. 31, no. 2, pp. 226-228, Oct 2004.
[8] J.-O. Yoo, K.-J. Kim, and S.-B. Cho, "Bayesian
inference with fuzzy variables for customized high level context extraction,"
Proc. of Korea Information Science Society, vol. 31, no. 2, pp. 115-117, Oct 2004.
| 
