MTD-Map: Single-Stage
Long-Term LiDAR Map Maintenance Framework
via Mixture Transition Distribution

TaeYoung Kim, Gilhwan Kang, Tae Ihn Kim, Seungwon Song, Hun Keon Ko^*

Robotics Lab, Advanced Vehicle Platform Division, Hyundai Motor Company

IROS 2026

TL;DR : A single-stage LiDAR map maintenance framework
unifying dynamic object removal and change detection
via Mixture Transition Distribution.

Abstract

While robust map maintenance has advanced significantly, existing studies have focused on specific tasks, especially dynamic object removal or change detection. In this paper, we take a holistic view of the map maintenance problem and propose MTD-Map, a single-stage framework that handles both dynamic object removal and change detection without separate task-specific modules. MTD-Map employs an explicit representation that compactly encodes the direction and duration of occupancy transitions through Mixture Transition Distribution (MTD) modeling. We develop a recursive MTD formulation that encodes historical occupancy patterns into an augmented state to capture high-order temporal dependencies. Furthermore, a stability-driven adaptive strategy balances noise suppression with the preservation of quasi-static structures. Extensive experiments verify that MTD-Map robustly removes dynamic objects and achieves competitive change detection performance, subsequently reducing computational costs.

System Overview

Qualitative Result

MTD-Map with dynamic objects — **Left**: MTD-Map result with dynamic points, **Right**: MTD-Map result without dynamic points.
Visualizing MTD-Map results via the temporal transition map $\Psi$ , showing static (gray) and quasi-static (blue, red) components throughout, with dynamic objects (olive) present only in the unfiltered map (left).

MTD-Map without dynamic objects — **Left**: MTD-Map result with dynamic points, **Right**: MTD-Map result without dynamic points.
Visualizing MTD-Map results via the temporal transition map $\Psi$ , showing static (gray) and quasi-static (blue, red) components throughout, with dynamic objects (olive) present only in the unfiltered map (left).

Dynamic Object Removal Result

Qualitative DOR results for SemanticKITTI. Points are visualized based on classification accuracy: True Positive (green), False Negative (blue), and False Positive (red).

Quantitative Result

Quantitative evaluation of dynamic object removal (DOR) in point cloud maps.
Bold = best, underline = second best. Units are in %.

Method	KITTI 00			KITTI 07			HeLiMOS 6593			MOE 02
Method	SA↑	DA↑	HA↑	SA↑	DA↑	HA↑	SA↑	DA↑	HA↑	SA↑	DA↑	HA↑
Removert	99.08	68.53	81.02	99.12	34.06	50.70	97.42	40.65	57.37	96.78	27.20	42.46
ERASOR	90.44	96.37	93.31	89.19	97.37	93.10	82.58	95.61	88.62	94.55	99.38	96.90
DUFOMap	99.19	83.46	90.64	99.14	55.60	74.25	98.08	67.64	80.07	96.42	48.13	64.21
OTD	94.93	93.74	94.33	94.64	84.54	89.31	97.93	73.27	83.83	95.67	87.50	91.40
HMM-MOS	99.92	58.45	73.76	99.82	45.57	67.44	99.56	65.62	79.10	99.24	5.84	11.04
ELite	92.09	89.24	90.64	93.77	90.27	91.99	89.28	95.80	92.43	82.93	99.76	90.57
MTD-Map (Ours)	99.47	95.28	97.33	92.71	95.61	94.13	98.51	82.95	90.06	95.95	94.84	95.39

Acknowledgement

We used Uni-Mapper framework for map merge during our experiments.

BibTeX citation

@inproceedings{kim2026mtdmap,
  author    = {Kim, TaeYoung and Kang, Gilhwan and Kim, Tae Ihn and Song, Seungwon and Ko, Hun Keon},
  title     = {{MTD-Map}: Single-Stage Long-Term {LiDAR} Map Maintenance Framework via Mixture Transition Distribution},
  booktitle = {Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  publisher = {IEEE},
}