DGMapping: Degeneration-Guided Adaptive Multisensor Fusion for Cost-Effective Mapping

DGMapping receives multi-sensor input and builds a 2D grid map with compensatory 3D structure — DGMapping receives input data from multiple sensors and follows the proposed standardized process: 1) constructing a probabilistic model to determine the probability distribution of candidate states for each sensor; 2) identifying scene degeneration categories and constructing scene degeneration descriptors; 3) adaptively building a joint probabilistic model. As a result, DGMapping enables real-time optimal state estimation and constructs a high-precision 2D grid map, enhanced with compensatory 3D structural information.

01

Abstract

2D LiDAR is widely employed for map construction due to its cost-effectiveness and reliable measurements. However, its inherent limitation of single-plane scanning constrains the dimensionality and scale of the observed data, resulting in degeneration caused by insufficient geometric constraints. Although LiDAR-inertial fusion enhances mapping by compensating for LiDAR degeneration through motion correction, it struggles with accumulating sensor errors and fails to capture comprehensive 3D structural information.

In this paper, we focus on enhancing environmental data utilization, refining the observation model, and integrating 3D structural information to improve mapping quality, rather than relying solely on low-precision motion compensation. We introduce DGMapping, a cost-effective multi-sensor fusion framework that follows a standardized process from joint probabilistic model construction to optimization. During the model construction phase, we incorporate a sampling-optimized fusion observation model to improve the efficiency and completeness of environmental data utilization. Additionally, we further refine the scene degeneration description to enhance the adaptive optimization of the joint probabilistic model. Experimental results across six real-world scenarios demonstrate that DGMapping outperforms existing algorithms in terms of odometry and mapping performance.

02

Framework

03

Experimental Setup

Mobile platform sensor suite and the six experimental scenes — **Overview of real-world experiments.** Left: The mobile platform equipped with a sensor suite comprising an RGB-D camera, a 3D LiDAR, a 2D LiDAR, and an IMU. Right: The experimental scenarios covered in our dataset, categorized into General Environments and Degraded Environments, each containing three representative scenes. All 2D grid maps shown are constructed using our proposed DGMapping.

04

Quantitative Results

TABLE I: Quantitative evaluation of odometry performance in degenerate environments.
Method		Geometric Degeneration				Range Degeneration				Planar Degeneration				Average CE ↓
		ATE [m] ↓		ARE [rad] ↓		ATE [m] ↓		ARE [rad] ↓		ATE [m] ↓		ARE [rad] ↓
		RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD
ORB-SLAM3		2.731	0.946	0.021	0.013	1.502	0.671	0.025	0.011	0.046	0.009	0.007	0.003	0.722
Hector		13.535	5.292	0.027	0.013	12.357	6.469	0.117	0.063	5.582	2.136	0.021	0.012	6.328
Karto		11.283	5.195	0.026	0.012	10.957	5.264	0.192	0.119	3.435	1.719	0.019	0.010	4.319
CSM	θ = 1.0	10.325	4.879	0.029	0.014	9.773	5.132	0.073	0.044	3.253	1.658	0.017	0.009	3.912
	θ = 0.5	8.858	3.694	0.037	0.025	6.633	3.197	0.168	0.095	1.146	0.732	0.028	0.019	2.812
	θ = 0.1	5.222	1.941	0.084	0.052	5.504	2.677	0.191	0.119	0.946	0.632	0.036	0.026	1.997
Cartographer		4.231	1.289	0.027	0.014	3.643	1.912	0.085	0.036	0.930	0.293	0.029	0.017	1.490
DN-CSM		3.038	1.136	0.015	0.008	3.044	1.573	0.035	0.017	0.913	0.597	0.032	0.021	1.180
DGMapping (w/o)		2.135	0.994	0.013	0.007	2.802	1.307	0.033	0.016	0.441	0.206	0.025	0.014	0.908
DGMapping (ours)		1.997	0.841	0.013	0.007	2.541	1.151	0.032	0.015	0.184	0.080	0.011	0.006	0.796

TABLE II: Quantitative evaluation of odometry performance in general environments.
Method		Science Building				Laboratory				Training Center				Average CE ↓
		ATE [m] ↓		ARE [rad] ↓		ATE [m] ↓		ARE [rad] ↓		ATE [m] ↓		ARE [rad] ↓
		RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD	RMSE	STD
ORB-SLAM3		0.320	0.119	0.020	0.009	2.392	0.814	0.073	0.046	0.446	0.043	0.025	0.015	0.546
Hector		5.242	2.207	0.021	0.013	13.665	6.078	0.069	0.046	1.173	0.597	0.023	0.016	3.366
Karto		4.152	1.623	0.025	0.014	5.783	3.019	0.129	0.084	1.132	0.511	0.021	0.011	1.873
CSM	θ = 1.0	4.543	1.838	0.022	0.013	7.731	3.635	0.085	0.067	0.224	0.104	0.021	0.011	2.104
	θ = 0.5	3.486	1.502	0.026	0.015	6.244	3.156	0.093	0.075	0.961	0.439	0.027	0.018	1.806
	θ = 0.1	3.015	1.094	0.037	0.021	5.551	2.957	0.137	0.098	1.266	0.611	0.034	0.021	1.673
Cartographer		1.845	0.719	0.022	0.011	2.619	0.948	0.099	0.051	0.211	0.099	0.020	0.011	0.802
DN-CSM		0.446	0.141	0.021	0.012	2.508	0.884	0.069	0.044	0.206	0.105	0.019	0.012	0.545
DGMapping (w/o)		0.327	0.146	0.020	0.011	1.936	0.891	0.068	0.044	0.199	0.097	0.020	0.013	0.428
DGMapping (ours)		0.307	0.148	0.019	0.011	1.617	0.721	0.067	0.043	0.173	0.069	0.016	0.010	0.367

Absolute Trajectory Error (ATE, m) and Absolute Rotational Error (ARE, rad) are reported. Proposed method Visual reference (degenerate table). Bold and underlined values indicate the best and second-best results.

05

Mapping Results

Technology Building overview — Technology Building

Technology Building map by DGMapping — Technology Building

Mapping results for Tech Tower and corridor scenes. The top row displays the scene overviews, while the bottom row compares the mapping outcomes of DGMapping and DN-CSM. Images are cropped for clarity.

06

Efficiency & Cost Analysis

Memory footprint monitoring — (a) Memory footprint

SWaP-C metrics comparison — (d) SWaP-C metrics

Computational efficiency and resource evaluation. (a) Memory footprint monitoring shows the memory consumption patterns of different algorithms. (b) CPU usage statistics demonstrate the computational load distribution. (c) Runtime analysis of system components highlights the time efficiency of each module. (d) Comprehensive SWaP-C (Size, Weight, Power, and Cost) metrics comparison against baselines.

07

Citation

@article{11569794,
  author  = {Liu, Yu and Li, Zhongyu and Zhou, Meiqi and Fan, Zipei and Feng, Guanyuan},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  title   = {DGMapping: Degeneration-Guided Adaptive Multisensor Fusion for Cost-Effective Mapping},
  year    = {2026},
  volume  = {75},
  pages   = {9526218-9526218},
  doi     = {10.1109/TIM.2026.3704099}
}