Monocular VI-SLAM Algorithm Based on Lightweight SuperPoint Network in Low-Light Environment

Xudong Zeng; Shaosheng Fan; Shangzhi Xu; Yuting Zhou

doi:10.3788/LOP232620

Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1815001(2024)

Monocular VI-SLAM Algorithm Based on Lightweight SuperPoint Network in Low-Light Environment

Xudong Zeng^1,2, Shaosheng Fan^1,2、*, Shangzhi Xu^1,2, and Yuting Zhou^1,2

¹School of Electrical and Information Engineering, Changsha University of Science and Technology, Changsha 410111, Hunan, China

²Hunan Provincial Key Laboratory of Electric Power Robotics, Changsha 410111, Hunan, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. System framework diagram of GS-VINS

Download full size

Fig. 2. Comparison of adaptive image enhancement algorithms before and after processing. (a) Image before enhancement; (b) image after enhancement; (c) grayscale distribution before enhancement; (d) grayscale distribution after enhancement

Download full size

Fig. 3. GN2_SuperPoint Feature Detection Network Framework

Download full size

Fig. 4. Flowchart of dynamic adjustment for feature tracking threshold

Download full size

Fig. 5. Comparison of optical flow tracking effect under different feature extraction thresholds. (a) When the feature extraction threshold is large; (b) the feature extraction threshold is set to 0 01; (c) when the feature extraction threshold is small

Download full size

Fig. 6. Patrol robots with dataset acquisition equipment

Download full size

Fig. 7. Effect comparison of different algorithms on EuRoC image sequence enhancement. (a) Origin image; (b) CLAHE; (c) Zero Dec++; (d) single scale Retinex; (e) dark channel; (f) proposed algorithm

Download full size

Fig. 8. Effect comparison of different algorithms on real environment image sequence enhancement. (a) Origin image; (b) CLAHE; (c) Zero Dec++; (d) single scale Retinex; (e) dark channel; (f) proposed algorithm

Download full size

Fig. 9. Comparison of time consumption of different feature extraction algorithms. (a) MH_04_difficult sequence; (b) V1_03_difficult sequence

Download full size

Fig. 10. Comparison of feature tracking performance in low-light scenes. (a) Comparison of tracking success rates of GS-VINS and VINS-Mono algorithms in the MH_04_difficult sequence; (b) comparison of tracking success rates of GS-VINS and VINS-Mono algorithms in the V1_03_difficult sequence

Download full size

Fig. 11. Comparison of GS-VINS and VINS-Mono motion estimation visualisations in the EuRoc dataset. (a)‒(e) Sequences 01-05 for MH scene; (f)‒(h) sequences 01-03 for V1 scene; (i)‒(k) sequences 01-03 for V2 scene

Download full size

Fig. 12. Ablation experiment for each module of GS-VINS algorithm. (a)‒(e) Sequences 01-05 for MH scene; (f)‒(h) sequences 01-03 for V1 scene; (i)‒(k) sequences 01-03 for V2 scene

Download full size

Fig. 13. Comparison of absolute trajectory errors between GS-VINS and VINS-Mono in realistic scenarios

Download full size

Table 1. Encoding structure of GN2_SuperPoint
View table
Table 1. Encoding structure of GN2_SuperPoint
Conv Ex #out SE Stride
$W \times H \times 1$
Conv BN ReLU 1 16 2
GN2-bottleneck 16 16 1
GN2-bottleneck 48 24 2
GN2-bottleneck 72 24 1
GN2-bottleneck 72 40 1 2
GN2-bottleneck 120 40 1 1
GN2-bottleneck 240 80 1

Table 2. Comparison of enhancement effects of different algorithms

View table

Table 2. Comparison of enhancement effects of different algorithms

Image	MG
Image	Origin image	CLAHE	Single scale Retinex	Zero Dec++	Dark channel	Proposed algorithm
IMG_01	0.0063	0.0111	0.0145	0.0178	0.0057	0.0221
IMG_02	0.0054	0.0111	0.0077	0.0283	0.0128	0.0122
IMG_03	0.0077	0.0171	0.0278	0.028	0.0128	0.0579
IMG_04	0.0024	0.0284	0.0122	0.0184	0.0049	0.0354
IMG_05	0.0022	0.0414	0.0145	0.0174	0.0127	0.0332
Image	IE
Image	Origin image	CLAHE	Single scale Retinex	Zero Dec++	Dark channel	Proposed algorithm
IMG_01	5.4204	6.7988	6.6287	6.8776	4.6348	6.9092
IMG_02	3.3372	4.1168	6.2618	5.0369	3.6426	6.4038
IMG_03	6.2675	5.3494	2.8774	4.9475	5.2164	6.4354
IMG_04	5.6147	2.6518	2.6187	4.5407	1.5445	6.4038
IMG_05	4.0427	1.8092	1.9049	3.7571	3.5512	4.1852
Image	Cost time /s
Image	Origin image	CLAHE	Single scale Retinex	Zero Dec++	Dark channel	Proposed algorithm
IMG_01	0	0.0034	0.5432	0.1546	0.3452	0.1345
IMG_02	0	0.0041	0.6232	0.4234	0.5434	0.3486
IMG_03	0	0.0123	0.7365	0.6031	0.4264	0.3540
IMG_04	0	0.0031	0.6756	0.1975	0.4324	0.1645
IMG_05	0	0.0024	0.5376	0.1343	0.3001	0.1265
Image	Number of features
Image	Origin image	CLAHE	Single scale Retinex	Zero Dec++	Dark channel	Proposed algorithm
IMG_01	185	348	242	341	326	636
IMG_02	126	222	756	508	401	806
IMG_03	330	785	1232	806	254	1397
IMG_04	55	1004	1428	68	55	1025
IMG_05	94	216	942	143	151	1396

Table 3. Robustness test of feature point extraction algorithms against illumination
View table
Table 3. Robustness test of feature point extraction algorithms against illumination
Algorithm Image dateset 01-57
NMS is 3 NMS is 4 NMS is 8
Harris 0.632 0.603 0.542
Shi_tomasi 0.641 0.613 0.523
SuperPoint 0.695 0.673 0.645
GN2_SuperPoint 0.691 0.678 0.642

Table 4. ATE comparison of GS-VINS and VINS-Mono

View table

Table 4. ATE comparison of GS-VINS and VINS-Mono

Dataset	VINS-Mono	GS-VINS
MH_01	0.090987	0.074370
MH_02	0.118628	0.052764
MH_03	0.077581	0.066330
MH_04	0.156468	0.128692
MH_05	0.219968	0.135611
V1_01	0.057514	0.043245
V1_02	0.063660	0.047077
V1_03	0.238229	0.130703
V2_01	0.060146	0.051538
V2_02	0.104841	0.082816
V2_03	0.246471	0.205975

Tools

Get Citation

Copy Citation Text

Xudong Zeng, Shaosheng Fan, Shangzhi Xu, Yuting Zhou. Monocular VI-SLAM Algorithm Based on Lightweight SuperPoint Network in Low-Light Environment[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1815001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites