Attention-based smart-camera for spatial cognition

Bio-inspired attentional vision allows to reduce post-processing to few regions of the visual field. However, the computational complexity of most visual chains remains an issue for an embedded processing system such as a mobile and autonomous robot. We propose in this paper an attention-based smart-camera and neural networks for place recognition in the context of navigation missions in robotics. The smart-camera extracts points of interest based on retina receptive fields at multiple scales and in real-time thanks to a dedicated hardware architecture prototyped onto reconfigurable devices. The place recognition is computed by neural networks, inspired by hippocampal place cells, that code for both the descriptors ('what' information) and the locations ('where' information) of the points of interest provided by the smart-camera. We experimented the addition in the recognition process of a coarse-to-fine approach and obtained improved results during robot localisation experiments.