Regions of interest in observing robot hand movement by a cooperative robot

This paper presents a method for generating regions-of-interest in a scene of robot hand movement observed by a cooperative robot for action recognition. In a cooperative work, a robot needs to be aware of an action of its partner robot by detecting some regions-of-interest in the visual field like human visual scanpath. To generate regions-of-interests by a robot, we have applied image processing algorithms based on active top-down feature patterns and bottom-up spatial kernels. The algorithms have produced energy maps from the images observed by the robot and they were combined with different weights to generate algorithmic regions-of-interests. They were compared with human regions-of-interest measured by a psychophysical experiment and the algorithmic predictability of scanpath was evaluated using a positional similarity index. Experimental results showed that presented method is applicable to the detection of regions-of-interests in hand movement.