Swarms for robot vision: The case of adaptive visual trail detection and tracking

Abstract

Previous work has shown that a pheromone-based visual saliency map can be computed by a swarm of simple agents inhabiting the robot's input image.It was also shown that, with a proper set of behaviours controlling the agents, the saliency map can be used to localise trails present in the robot's visual field.Under the assumption that the robot starts its autonomous operation already on the trail, this paper extends that work by enabling the agents to learn online an appearance model of the trail.The learned model is then used to increase the level of pheromone deployed in the regions of the input image that are more probable of belonging to the trail.This is motivated by the well-known importance that a priori object knowledge has to improve visual search.The outcome of this extension is a self-organising behaviour capable of detecting trails in 98% of the evaluated situations, outperforming the original work.The agents being simple their computation is fast, resulting in a 12 Hz performance.Thus, by introducing a parsimonious learning mechanism, this paper contributes to increase robustness of swarm-based robot vision systems.