Swarm-based visual saliency for trail detection

This paper proposes a model for trail detection that builds upon the observation that trails are salient structures in the robot´s visual field. Due to the complexity of natural environments, the straightforward application of bottom-up visual saliency models is not sufficiently robust to predict the location of trails. As for other detection tasks, robustness can be increased by modulating the saliency computation with top-down knowledge about which pixel-wise visual features (e.g., colour) are the most representative of the object being sought. This paper proposes the use of the object´s overall layout instead, as it is a more stable and predictable feature in the case of natural trails. This novel component of top-down knowledge is specified in terms of perception-action rules, which control the behaviour of simple agents performing as a swarm to compute the saliency map of the input image. For the purpose of multi-frame evidence accumulation about the trail location, a motion compensated dynamic neural field is used. Experimental results on a large data-set reveal the ability of the model to produce a success rate of 91% at 20Hz. The model shows to be robust in situations where previous trail detectors would fail, such as when the trail does not emerge from the lower part of the image or when it is considerably interrupted.