TRGS-SLAM: IMU-Aided Gaussian Splatting SLAM for Blurry, Rolling Shutter, and Noisy Thermal Images

Spencer Carmichael; Katherine A. Skinner

cs.ROMar 20, 2026

TRGS-SLAM: IMU-Aided Gaussian Splatting SLAM for Blurry, Rolling Shutter, and Noisy Thermal Images

Authors: Spencer Carmichael, Katherine A. Skinner

Organizations: University of Michigan, Ann Arbor, MI USA

Abstract

Thermal cameras offer several advantages for simultaneous localization and mapping (SLAM) with mobile robots: they provide a passive, low-power solution to operating in darkness, are invariant to rapidly changing or high dynamic range illumination, and can see through fog, dust, and smoke. However, uncooled microbolometer thermal cameras, the only practical option in most robotics applications, suffer from significant motion blur, rolling shutter distortions, and fixed pattern noise. In this paper, we present TRGS-SLAM, a 3D Gaussian Splatting (3DGS) based thermal inertial SLAM system uniquely capable of handling these degradations. To overcome the challenges of thermal data, we introduce a model-aware 3DGS rendering method and several general innovations to 3DGS SLAM, including B-spline trajectory optimization with a two-stage IMU loss, view-diversity-based opacity resetting, and pose drift correction schemes. Our system demonstrates accurate tracking on real-world, fast motion, and high-noise thermal data that causes all other tested SLAM methods to fail. Moreover, through offline refinement of our SLAM results, we demonstrate thermal image restoration competitive with prior work that required ground truth poses.

Figures & tables

Figure 1: The main elements of TRGS-SLAM. Left: Microbolometer-aware rendering. Right: 3DGS map and B-spline trajectory.

Figure 3: Microbolometer-aware rendering process. A sum of multiple rasters, each multiplied with precomputed pixel-wise weights, is combined with a learned FPN estimate to obtain an approximation of the degraded input image.

Figure 4: A depiction of the Gaussians targeted by our view-diversity-based opacity resetting. The color scale is mapped to the log of the condition number of $\mathbf{V}_{g}$ for each Gaussian.

Figure 5: TRGS-SLAM result on the medium indoor sequence.

Type	Method	SO	MO	FO	SI	MI	FI
Monocular	ORB-SLAM3	0.8	9.4	$\times$	$\times$	$\times$	$\times$
	DSM	9.4	$\times$	$\times$	76.7	$\times$	$\times$
	DROID-SLAM	0.5	2.6	$\times$	0.7	$\times$	$\times$
	MASt3R-SLAM*	1.6	94.5	$\times$	2.7	88.4	$\times$
	MonoGS	7.7	$\times$	$\times$	83.9	$\times$	$\times$
Mono-Inertial	ORB-SLAM3 (inertial)	1.1	$\times$	$\times$	$\times$	$\times$	$\times$

Table 1: RMSE ATE in cm on the TRNeRF dataset.

Figure 6: Qualitative restoration results. With offline refinement, TRGS-SLAM achieves restoration quality near that of TRNeRF.

*Utilize ground truth poses. Results copied from [ 7 ] .
Method	MO	FO	MI	FI
GS on the Move*	0.093	0.283	0.206	0.328
TRNeRF*	0.042	0.078	0.068	0.117
TRGS-SLAM + Refinement	0.067	0.098	0.095	0.116

Table 2: LPIPS results on the TRNeRF dataset (lower is better)

Version	Device	Number of Timestamps Evaluated	Rotation Loss Compute Time (ms)		Ang. Vel. Loss Compute Time (ms)
Version	Device	Number of Timestamps Evaluated	Forward	Backward	Forward	Backward
PyPose	CPU	1k	5.45	4.97	5.64	5.71
LieTorch	CPU	1k	1.68	1.39	1.82	1.48
C++	CPU	1k	0.59	0.41	0.51	0.27
PyPose	GPU	1k	14.30	13.74	15.44	15.55
LieTorch	GPU	1k	3.11	2.12	4.28	2.28

Table 3: Timing comparison between different versions of our rotation spline.

Description	Value
Trajectory
Spline order for the positions and rotations ( $k$ )	4
Knot interval for the positions and rotations ( $\Delta t_{c}$ )	$\sim 8$ milliseconds (half the frame period)
Learning rates for the position and rotation control points	$1\times 10^{-3},1\times 10^{-4}$
Number of predicted poses used in spline fitting ( $N_{q}$ )	10 per knot in the time range of the predictions
IMU Parameters & Loss

Table 4: SLAM parameter values.

Ablation Method	SO		MO		FO		SI		MI		FI
	ATE (cm)	# of Trials Completed	ATE (cm)	# of Trials Completed	ATE (cm)	# of Trials Completed	ATE (cm)	# of Trials Completed	ATE (cm)	# of Trials Completed	ATE (cm)	# of Trials Completed
Monocular*	22.5	3	$\times$	0	$\times$	0	7.9	3	$\times$	0	$\times$	0
No Gyro. Before IMU Init.	51.8	3	32.6	1	18.0	3	10.2	3	8.3	3	$\times$	0
Gyroscope Only*	3.6	3	5.7	3	3.7	3	3.3	3	4.6	1	$\times$	0
No Roll/Blur Modeling	9.0	3	7.3	3	32.8	1	4.2	3	5.7	3	$\times$	0
No Pixel-wise FPN	22.7	3	$\times$	0	$\times$	0	52.1	1	$\times$	0	$\times$	0

Table 5: Ablation study results.

Task				Avg. Time Per Run (ms) $\times$ Avg. Number of Runs per Parent Task	% of Parent Task
Tracking				193.27 $\times$ 9654	36.9
	B-Spline Extension			27.32	14.1
	Optimization Iteration			5.12 $\times$ 32	84.8
		Rendering		1.88	36.8
			Pose Evaluation	0.60	31.7
			Rasterization	0.57	30.1

Table 6: Timing analysis on the MI sequence.

Simultaneous Localization and Mapping

Simultaneous localization and mapping estimates a robot’s motion while building a map of an unknown environment.

See all 193 papers

LST-SLAM: A Stereo Thermal SLAM System for Kilometer-Scale Dynamic Environments
Feb 24, 2026Zeyu Jiang, Kuan Xu, Changhao ChenDynamic SLAM Visual SLAM
PEAK-Lab, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 511453, China · School of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798
PL-LIT: A LiDAR-Inertial-Thermal SLAM Using Point-Line Features and Thermographic Mapping
Jun 28, 2026Jiawei Xia, Yixiao Feng, Yongliang Shi +4Thermal Imaging Simultaneous Localization and Mapping
Tsinghua University · The Hong Kong University of Science and Technology (Guangzhou) · University of California, Berkeley
Thermal Image Refinement with Depth Estimation using Recurrent Networks for Monocular ORB-SLAM3
Mar 16, 2026Hürkan Şahin, Huy Xuan Pham, Van Huyen Dang +2Depth Estimation Monocular Depth Estimation
Automatic Control Group (RAT), Paderborn University, 33098 Paderborn, Germany · Department of Electrical and Computer Engineering, Aarhus University, 8000 Aarhus C, Denmark, and also with Upteko ApS, Denmark

Feb 24, 2026J/K move · Enter open · S save

LST-SLAM: A Stereo Thermal SLAM System for Kilometer-Scale Dynamic Environments

Zeyu Jiang, Kuan Xu, Changhao Chen

PEAK-Lab, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 511453, China · School of Electrical and Electronic Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798

Dynamic SLAM Visual SLAM Thermal Imaging Loop Closure Detection

Open paper AskPDF arXiv

Thermal cameras offer strong potential for robot perception under challenging illumination and weather conditions. However, thermal Simultaneous Localization and Mapping (SLAM) remains difficult due to unreliable feature extraction, unstable motion tracking, and inconsistent global pose and map construction, particularly in dynamic large-scale outdoor environments. To address these challenges, we propose LST-SLAM, a novel large-scale stereo thermal SLAM system that achieves robust performance in complex, dynamic scenes. Our approach combines self-supervised thermal feature learning, stereo dual-level motion tracking, and geometric pose optimization. We also introduce a semantic-geometric hybrid constraint that suppresses potentially dynamic features lacking strong inter-frame geometric consistency. Furthermore, we develop an online incremental bag-of-words model for loop closure detection, coupled with global pose optimization to mitigate accumulated drift. Extensive experiments on kilometer-scale dynamic thermal datasets show that LST-SLAM significantly outperforms recent representative SLAM systems, including AirSLAM and DROID-SLAM, in both robustness and accuracy.

Abstract

Simultaneous Localization and Mapping

Explore similar work

LST-SLAM: A Stereo Thermal SLAM System for Kilometer-Scale Dynamic Environments

PL-LIT: A LiDAR-Inertial-Thermal SLAM Using Point-Line Features and Thermographic Mapping

Thermal Image Refinement with Depth Estimation using Recurrent Networks for Monocular ORB-SLAM3