STRATA: One Geometry-Free Persistence-and-Periodicity Engine for Selectable 2D/3D Lifelong LiDAR Mapping

Source abstract

A robot that maps one building for weeks must keep walls, forget movers and treat a scheduled door as signal. Occupancy mappers make the first two decisions with a log-odds accumulator whose increments were chosen by habit. We show that the OctoMap increments (+0.85, −0.4) are the increments of Wald’s sequential probability ratio test (SPRT) between the hit rates 0.1975 and 0.4621, so that “Static” means “occupied in more than about a third of the windows”, and that a multiplicative leak removes the property on which an error bound rests. STRATA 0.2 makes the test explicit: the user states two hit rates and two error rates, increments and thresholds follow, and a cell ends Static, Rejected or Undecided. Under independent windows, clutter at a hit rate of at most m0 graduates in the first test of one cell history (up to its first graduation, its first rejection or its retention as a periodic candidate) with probability at most 0.0101 (Ville’s inequality); restarts after a demotion are not covered. The operating characteristic measured on the engine with the Periodic veto off matches the exact computation of the rule and is approximated by Wald’s closed form. At a lag-1 autocorrelation of 0.23, the low end of the range measured on real calibration cells (0.231 to 0.474), the graduation rate per forward test (a history split at each demotion) is 3.49 times α, and at 0.5 it is 10.87 times α. We derived the pair under a pre-registered rule that we amended after the first calibration measurement, and the wall-thickness gate failed on every calibration sequence, so the shipped pair (0.20, 0.70, 0.01, 0.01) is provisional and the 3D pair is not calibrated. On held-out LILocBench, Oxford Spires, KTH-benchmark and Boreas data we compare seven evidence rules, without tuning and without choosing a winner. On LILocBench clutter, 8.81 % of the cells are ever Static under the sequential test (2.93 % at an observed hit rate of at most m0) against 21.16 % under v0.1; both exceed the nominal α = 0.01, which is per history, not per cell. The sequential test rejects none of the 366 wall-run cells, as does v0.1, but it leaves 14.21 %of them not Static at the end against 1.09 % for v0.1, and it needs a median of 33 windows to graduate one, against 10. In the opposite direction, on KITTI 05 (0.2 m voxels, ten frames per window, Periodic veto on) under the policy that exports Static voxels only, it drops more of both static and dynamic points than v0.1 (static accuracy 66.90 % against 90.20 %, dynamic accuracy 98.99 % against 83.19 %); a dynamic accuracy above that of an oracle map is reached only by dropping static points, and in this setting, under the policy that exports every voxel not Rejected, it removes no dynamic point. The engine is one geometry-free state machine for a 2D grid and a 3D voxel hash, with a FreMEn-style periodic label whose false-alarm bound holds for iid clutter; Periodic is validated on synthetic streams only. Code: https://github.com/kjungmo/strata/tree/paper-v2.