Next: An Integrated Vision Up: Sensor and Processor Previous: Exploring With A

A Real-Time Attentional Mechanism for Robotic Vision

Authors: [tex2html_wrap4226]G. Sela, M.D. Levine

Investigator username: levine

Category: perception

Subcategory: sensor and processor design

Real-time computer vision systems are burdened by extremely large amounts of data which must be processed in a small amount of time. To this end, an attentional process allowing computational resources to be concentrated on salient regions in an image would allow scenes to be processed faster and more efficiently. As well, data reduction is also being accomplished through the use of sensors with nonuniform sampling resolutions modeled after the primate visual system. In these implementations, a central high resolution foveal region is surrounded by a much coarser peripheral region whose resolution decreases with the distance from the centre. In such a system, an attentional mechanism is necessary to ensure that the interesting parts of a scene fall within the fovea in order to be analyzed at the highest resolution possible.

This research involves the development and real-time implementation of a general purpose, context-free algorithm to determine interest points in a scene at which a robot-mounted, foveated camera should focus its attention. The objective is to continuously position the camera so that the interesting areas in the scene lie within the foveal region. The algorithm being developed for this purpose is based on psychophysical findings which indicate that human gaze tends to fixate at the centres of mass of enclosed or partially enclosed, symmetric objects. The system uses context-free edge information as its input and is implemented in real-time on a network of parallel processing nodes based on the Texas Instruments TMS320C40 processor.

Tests have shown the algorithm to have a variety of uses beyond visual attention. In particular, it has proven well-suited to finding the location of human facial features over a large range of scales and poses in a real-world environment. This makes the algorithm suitable for use as a fast front-end processor for face recognition and face animation programs.



Next: An Integrated Vision Up: Sensor and Processor Previous: Exploring With A


baron@cim.mcgill.ca