LT-mapper: A Modular Framework for LiDAR-based Lifelong Mapping
Giseop Kim
1
and Ayoung Kim
1∗
Abstract— Long-term 3D map management is a fundamental
capability required by a robot to reliably navigate in the
non-stationary real-world. This paper develops open-source,
modular, and readily available LiDAR-based lifelong mapping
for urban sites. This is achieved by dividing the problem into
successive subproblems: multi-session SLAM (MSS), high/low
dynamic change detection, and positive/negative change man-
agement. The proposed method leverages MSS and handles
potential trajectory error; thus, good initial alignment is not
required for change detection. Our change management scheme
preserves efficacy in both memory and computation costs,
providing automatic object segregation from a large-scale point
cloud map. We verify the framework’s reliability and appli-
cability even under permanent year-level variation, through
extensive real-world experiments with multiple temporal gaps
(from day to year).
I. INTRODUCTION
During long-term mapping using light detection and rang-
ing (LiDAR) sensor, we encounter changes in an environ-
ment as in Fig. 1. The perceived snapshot of the environment
contains both ephemeral and persistent objects that may
change over time. To handle this change properly, long-
term mapping must solve for autonomous map maintenance
[1] by detecting, updating, and managing the environmen-
tal changes accordingly. In doing so, the challenges in
scalability, potential misalignment error, and map storage
efficiency should be addressed and resolved toward lifelong
map maintenance.
1) Integration to multi-session SLAM for scalability: Some
studies regarded change detection as a post-process of com-
paring multiple pre-built maps associated with temporally
distant and independent sessions. As reported in [2], align-
ment of multiple sessions in a global coordinate may severely
limit scalability. Following their philosophy, in this work,
we integrate multi-session SLAM (MSS) and align sessions
with anchor nodes [2] to perform change detection in a
large-scale urban environment beyond a small-sized room
[3]. Our framework consists of a LiDAR-based multi-session
3D simultaneous localization and mapping (SLAM) module,
named LT-SLAM.
2) Change detection under SLAM error: Change detection
between two maps would be trivial if maps were perfectly
aligned. Early works [5, 3, 6, 7] in map change detection re-
lied on the strong assumption of globally well-aligned maps
with no error and avoided handling this ambiguity issue.
Unfortunately, trajectory error inevitably occur in reality.
1
G. Kim is with the Department of Civil and Environmental Engineering,
KAIST, Daejeon, S. Korea paulgkim@kaist.ac.kr
1
A. Kim is with Department of Civil and Environmental Engineering,
KAIST, Daejeon, S. Korea ayoungk@kaist.ac.kr
Fig. 1: An example of permanent structural changes over ∼ 1.5-year
temporal gap. (Top) KAIST 01 of MulRan dataset [4] (June 2019).
(Bottom) KAIST 04, recently released in extended sequences (February
2021). A construction wall appeared over time; the previously existing
parking spaces and trees disappeared. LT-mapper can accurately register the
temporally disjointed maps and detect pointwise changes (e.g., blue points
in the bottom blue box).
We reconcile this potential misalignment during our change
detection, and enable the proposed method to handle po-
tential alignment error robustly. To deal with the ambiguity,
we propose a scan-to-map scheme with projective visibility,
using range images of multiple window sizes named LT-
removert. By extending an intra-session change detection
method [8], the LT-removert includes both intra-/inter-session
change detection, thereby further segregating high and low-
dynamic objects [5] from the change.
3) Compact place management: In addition to change
detection, we present and prove a concept of change com-
position. Once the change is detected, the decision for map
maintenance should be followed to determine what to include
or exclude. Using this feature, ours not only maintains an up-
to-date map such as existing works [1, 3], but also extracts
stable structures with higher placeness; thereby, we construct
a reliable 3D map with authentically meaningful structures
for other missions, such as cross-modal localization [9] and
long-term localization [10]. This final module, named LT-
map, manages the changes and enables a central map to
evolve in a place-wise manner.
In sum, we propose a novel modular framework for
LiDAR-based lifelong mapping, named LT-mapper. Each
module in the framework can run separately via file-based
in/out protocol. Unified and modular lifelong mapping has
barely been made for 3D LiDAR, unlike recently (but par-
tially) delivered visual-based methods [11, 12, 13, 14, 15]. To
the best of our knowledge, LT-mapper is the first open mod-
ular framework that supports LiDAR-based lifelong mapping
in complex urban sites. The proposed has the following
contributions:
• We integrate MSS with change detection and handle
arXiv:2107.07712v1 [cs.RO] 16 Jul 2021
评论0
最新资源