A Place in One Kilobyte: Hyperdimensional Zone Recognition over Egocentric SLAM Maps for Offline Heritage Wearables

Abstract

Heritage sites resist the assumptions of modern localization: satellite positioning degrades among stone streets and terrain, disappears indoors, and visual localization pipelines demand power budgets incompatible with all-day wearable operation. This report introduces and field-tests a zone-level place recognition layer built on hyperdimensional computing (HDC) that operates directly on the semi-dense point clouds produced by egocentric SLAM (Project Aria Machine Perception Services). Local occupancy descriptors computed along an expert-mapped route are projected into binary hypervectors; each route zone is enrolled by bundling into a single prototype, and localization reduces to Hamming-distance search, storing each zone in approximately one kilobyte and requiring only XOR and popcount operations at query time. On recordings from a UNESCO World Heritage walk (Meride, Monte San Giorgio, Switzerland), the method attains 96.0% zone accuracy at dimension D=1024 under interleaved hold-out, matching a floating-point nearest-centroid baseline; under the stricter contiguous-block hold-out it reaches 67.5% exact and 80.6% within one zone, comparable to a rotation-matched Scan Context nearest-neighbour baseline (64.4% exact) while storing three orders of magnitude less. In a cross-device test with two glasses recording simultaneously under independent SLAM frames, a second wearer is localized within one zone 97.0% of the time at 15 m granularity. The intended role is architectural: an always-on, milliwatt-class coarse layer that gates a heavier centimeter-grade visual localizer inside an entirely offline system. Protocols for cross-day evaluation via ECEF-geo-referenced trajectories, heading estimation by circular descriptor shifts, and hierarchical zone nesting are specified for subsequent versions.