To realize service robots with the capability of visually guided human-like manipulation in an unstructured environment cluttered with various daily life objects, it is crucial to guarantee dependability in visual perception. In order to solve such critical issues as robustness and real-time processing in vision based 3D recognition and pose estimation, we are developing a novel vision based framework designed under the concept of “behavioral perception."

The main features of our approach as follows:

- Behavioral Perception: Dependability of perception may not come from the perfection of individual components, but from the integration of individual components into dependable system behaviors that guarantees reaching perceptual goals.
- Automatic selection and collection of an optimal set of evidences based on in-situ monitoring.
- A probabilistic and spatial-temporal integration of top-down and bottom-up perceptual processes: multiple evidence-model matching in a sequence of images based on particle filtering till reaching a credible decision.

* Overall Framework

Currently, we are focusing on developing 3D recognition system using FER-CNN and an image translation approach based on deep learning.