Towards Bio-inspired Place Recognition over Multiple Spatial Scales Zetao Chen 1 , Adam Jacobson 1 , Uğur M. Erdem 2 , Michael E. Hasselmo 2 and Michael Milford 1 1 School of Electrical Engineering and Computer Science Queensland University of Technology 2 Center for Memory and Brain and Graduate Program for Neuroscience Boston University zetao.chen@student.qut.edu.au Abstract This paper presents a new multi-scale place recognition system inspired by the recent discovery of overlapping, multi-scale spatial maps stored in the rodent brain. By training a set of Support Vector Machines to recognize places at varying levels of spatial specificity, we are able to validate spatially specific place recognition hypotheses against broader place recognition hypotheses without sacrificing localization accuracy. We evaluate the system in a range of experiments using cameras mounted on a motorbike and a human in two different environments. At 100% precision, the multi- scale approach results in a 56% average improvement in recall rate across both datasets. We analyse the results and then discuss future work that may lead to improvements in both robotic mapping and our understanding of sensory processing and encoding in the mammalian brain. 1 Introduction The vast majority of robotic mapping and navigation systems perform mapping at either one fixed spatial scale or over two, a local and a global scale [Bosse, et al., 2003, Kuipers, et al., 2004, Kuipers and Byun, 1991]. A range of recent high profile discoveries in neuroscience have demonstrated that animals such as rodents, and likely many other mammals including humans, encode the world using multiple parallel mapping systems each of which encode the world at a different scale [Stensola, et al., 2012, Hafting, et al., 2005]. In rodents, the mapping system scales from neurons that encode an area of a few square centimetres to neurons that encode an area of several square metres, with many intermediate scales ZC, AJ and MM are with the School of Electrical Engineering and Computer Science at the Queensland University of Technology, zetao.chen@student.qut.edu.au . UE and MH are with Center for Memory and Brain and Graduate Program for Neuroscience at Boston University. This work was supported by an Australian Research Council Discovery Project DP1212775 awarded to MM, and an Office of Naval Research ONR MURI N00014-10-1-0936 and Silvio O. Conte Center Grant P50 NIMH MH094263 to UME and MEH. represented in-between. There is also currently no known upper limit on scale, as practical limitations mean experiments cannot be performed in arbitrarily large environments. Figure 1: Our multi-scale place matcher combines the output of multiple arrays of SVMs trained to perform place recognition at different scales in order to filter out hypotheses not supported at all spatial scales. (a) Camera frames are input into each place recognition module leading to inconsistent place recognition hypotheses at different spatial scales, and hence (b) no match is reported. (c) A later set of camera frames leads to (d) consistent place recognition hypotheses at all scales, and hence the most spatially specific hypothesis is accepted. While a number of theoretical-only investigations have hypothesized possible benefits of such a multi- scale mapping system [Burak and Fiete, 2009, Welinder, et al., 2008], no one has investigated potential mapping and place recognition performance benefits in real world environments. In this paper, we propose that these multiple scales provide a mechanism for effectively combining place recognition hypotheses with varying spatial specificities – for example, combining a spatially specific hypothesis that a rat or robot is located at a specific corner of a room with a spatially broad hypothesis that it is located “somewhere” in that room. We further propose that the Proceedings of Australasian Conference on Robotics and Automation, 2-4 Dec 2013, University of New South Wales, Sydney Australia