Visual-Inertial SLAM Algorithm with Fusion of Point-Line Features and Stepwise Marginalization

Xinyao Chai; Li Li; Rong Tang; Dongxuan Han; Zhangjun Peng; Zhigui Liu

doi:10.3788/LOP242268

Laser & Optoelectronics Progress, Volume. 62, Issue 14, 1428002(2025)

Visual-Inertial SLAM Algorithm with Fusion of Point-Line Features and Stepwise Marginalization

Xinyao Chai¹, Li Li^2,3, Rong Tang^1,3, Dongxuan Han², Zhangjun Peng^2,3, and Zhigui Liu^1,3、*

¹School of Information and Engineering, Southwest University of Science and Technology, Mianyang 621010, Sichuan , China

²School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang 621010, Sichuan , China

³Sichuan Engineering Technology Research Center of Industrial Self-Supporting and Artificial Intelligence, Mianyang 621010, Sichuan , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Frame diagram of the EPLD-VINS algorithm

Download full size

Fig. 2. Comparison of the line features extraction effect between EPLD-VINS and PL-VINS

Download full size

Fig. 3. Illustration of sliding window optimization of visual inertial information

Download full size

Fig. 4. The ATE heatmaps of the MH_03_ medium sequence in 3D space

Download full size

Fig. 5. The ATE heatmaps of the V1_02_ medium sequence in 3D space

Download full size

Fig. 6. The ATE distribution plot of VINS-Mono, PL-VINS, and EPLD-VINS in MH_03_medium sequence

Download full size

Fig. 7. The ATE distributionst of VINS-Mono, PL-VINS, and EPLD-VINS in V1_02_easy sequence

Download full size

Fig. 8. The effects of line feature extraction and merging in different scenes

Download full size

Table 1. Comparison of the absolute trajectory error

View table

Table 1. Comparison of the absolute trajectory error

Sequence	VINS-Mono			PL-VINS			EPLD-VINS
Sequence	SD	median	mean	SD	median	mean	SD	median	mean
MH_01_easy	0.0590	0.0576	0.0634	0.0379	0.0836	0.0962	0.0219	0.0378	0.0414
MH_02_easy	0.0397	0.0373	0.0444	0.0286	0.0663	0.0737	0.0278	0.0782	0.0787
MH_03_medium	0.0449	0.0707	0.0810	0.0305	0.0487	0.0552	0.0246	0.0524	0.0554
MH_04_difficult	0.1073	0.1139	0.1125	0.0465	0.0984	0.1036	0.0395	0.0941	0.0962
MH_05_difficult	0.1034	0.1407	0.1401	0.0584	0.1182	0.1231	0.0336	0.0973	0.0949
V102_medium	0.0420	0.0766	0.0851	0.0198	0.0515	0.0510	0.0182	0.0276	0.0382
V202_medium	0.1032	0.1149	0.1382	0.0514	0.0722	0.0815	0.0318	0.0484	0.0519

Table 2. Comparison of the relative pose error

View table

Table 2. Comparison of the relative pose error

Sequence	VINS-Mono			PL-VINS			EPLD-VINS
Sequence	SD	median	mean	SD	median	mean	SD	median	mean
MH_01_easy	0.0269	0.0328	0.0391	0.0754	0.0886	0.1053	0.0753	0.0883	0.1052
MH_02_easy	0.0790	0.0938	0.1098	0.0109	0.0028	0.0043	0.0027	0.0023	0.0035
MH_03_medium	0.1246	0.1926	0.2083	0.0179	0.0047	0.0076	0.0049	0.0043	0.0057
MH_04_difficult	0.1304	0.1819	0.2116	0.0298	0.0054	0.0089	0.0062	0.0047	0.0061
MH_05_difficult	0.1299	0.1874	0.2104	0.0195	0.0051	0.0084	0.0074	0.0049	0.0067
V102_medium	0.0719	0.1273	0.1296	0.0695	0.0040	0.0046	0.0019	0.0037	0.0038
V202_medium	0.0674	0.0946	0.1062	0.0064	0.0029	0.0039	0.0063	0.0028	0.0037

Table 3. The RMSE of ATE of different algorithms in this sequence

View table

Table 3. The RMSE of ATE of different algorithms in this sequence

Sequence	VINS-Mono	VINS-Fusion^［21］	PL-VINS	EPL-VINS^［21］	EPLD-VINS （w/o decompose）	EPLD-VINS
MH_01_easy	0.0738	0.1582	0.1129	0.0675	0.0511	0.0468
MH_02_easy	0.0527	0.0942	0.0876	0.0710	0.0954	0.0919
MH_03_medium	0.0941	0.0607	0.0631	0.0634	0.0623	0.0606
MH_04_difficult	0.1234	0.1444	0.1106	0.1416	0.1167	0.1068
MH_05_difficult	0.1517	0.1430	0.1353	0.1120	0.1185	0.1004
V102_medium	0.0831	0.0513	0.0547	0.0466	0.0433	0.0354
V202_medium	0.1725	0.0657	0.0964	0.0805	0.0781	0.0608

Table 4. Comparison of running time of different algorithms
View table
Table 4. Comparison of running time of different algorithms
Thread Module VINS-Mono VINS-Fusion PL-VINS EPLD-VINS
1 Point detection and tracking 12.46 10.39 12.71 12.76
2 Line detection and tracking 27.65 22.19
3 Local VIO 24.58 20.54 23.75 29.02
4 Loop closure 76.91 79.91 74.68 74.40

Table 5. Comparison of ATE RMSE ablation on point and line feature modules

View table

Table 5. Comparison of ATE RMSE ablation on point and line feature modules

Sequence	PL-VINS	EPLD-VINS
MH_01_easy	0.0507	0.0407
MH_02_easy	0.0961	0.0688
MH_03_medium	0.0593	0.0465
MH_04_difficult	0.1347	0.1072
MH_05_difficult	0.0871	0.0850
V101_easy	0.0453	0.0404
V102_medium	0.0366	0.0303
V103_difficult	0.1766	0.1544

Table 6. Ablation comparison of the backend optimization module in terms of time on the EuRoc dataset

View table

Table 6. Ablation comparison of the backend optimization module in terms of time on the EuRoc dataset

Sequence	PL-VINS	EPLD-VINS
MH_01_easy	16533.17	8700.76
MH_02_easy	10642.57	6343.86
MH_03_medium	12259.10	8674.75
MH_04_difficult	6802.97	5231.64
MH_05_difficult	8606.01	6341.97
V101_easy	12726.89	9723.76
V102_medium	5698.64	3513.98
V103_difficult	5131.97	3123.76

Tools

Get Citation

Copy Citation Text

Xinyao Chai, Li Li, Rong Tang, Dongxuan Han, Zhangjun Peng, Zhigui Liu. Visual-Inertial SLAM Algorithm with Fusion of Point-Line Features and Stepwise Marginalization[J]. Laser & Optoelectronics Progress, 2025, 62(14): 1428002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites