Fast Object Detection with Foveated Imaging and Virtual Saccades on Resource Limited Robots

This paper describes the use of foveated imaging and virtual saccades to identify visual objects using both colour and edge features. Vision processing is a resource hungry operation at the best of times. When the demands require robust, real-time performance with a limited embedded processor, the challenge is significant. Our domain of application is the RoboCup Standard Platform League soccer competition using the Aldebaran Nao robot. We describe algorithms that use a combination of down-sampled colour images and high-resolution edge-detection to identify objects in varying lighting conditions. Optimised to run in real time on autonomous robots, these techniques can potentially be applied in other resource limited domains.