A learning approach to swarm-based path detection and tracking

Abstract

This dissertation presents a set of top-down modulation mechanisms for the modulation of the swarm-based visual saliency computation process proposed by Santana et al. (2010) in context of path detection and tracking. In the original visual saliency computation process, two swarms of agents sensitive to bottom-up conspicuity information interact via pheromone-like signals so as to converge on the most likely location of the path being sought. The behaviours ruling the agents’motion are composed of a set of perception-action rules that embed top-down knowledge about the path’s overall layout. This reduces ambiguity in the face of distractors. However, distractors with a shape similar to the one of the path being sought can still misguide the system. To mitigate this issue, this dissertation proposes the use of a contrast model to modulate the conspicuity computation and the use of an appearance model to modulate the pheromone deployment. Given the heterogeneity of the paths, these models are learnt online. Using in a modulation context and not in a direct image processing, the complexity of these models can be reduced without hampering robustness. The result is a system computationally parsimonious with a work frequency of 20 Hz. Experimental results obtained from a data set encompassing 39 diverse videos show the ability of the proposed model to localise the path in 98.67 % of the 29789 evaluated frames.