Indoor Localisation with Regression Networks and Place Cell Models

Animals use a variety of environmental cues in order to recognise their location. One of the key behaviours found in a certain type of biological neuron – known as place cells – is a rate-coding effect: a neuron’s rate of firing decreases with distance from some landmark location. In this work, we used visual information from wearable and handheld cameras in order to reproduce this rate-coding effect in artificial place cells (APCs). The accuracy of localisation using these APCs was evaluated using different visual descriptors and different place cell widths. Simple localisation using APCs was feasible by noting the identity of the APC yielding the maximum response. We also propose using joint coding within a number of automatically defined APCs as a population code for selflocalisation. Using both approaches we were able to demonstrate good self-localisation from very small images taken in indoor settings. The error performance using APCs is favourable when compared with ground-truth and LSD-SLAM, even without the use of a motion model.