Point Cloud Feature Extraction Network Based on Multiscale Feature Dynamic Fusion

Jing Liu; Yuan Zhang; Le Zhang; Bo Li; Xiaowen Yang

doi:10.3788/LOP241237

Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0415001(2025)

Point Cloud Feature Extraction Network Based on Multiscale Feature Dynamic Fusion

Jing Liu^1,2、*, Yuan Zhang^1,2, Le Zhang³, Bo Li^1,2, and Xiaowen Yang^1,2

Author Affiliations

¹School of Data Science and Technology, North University of China, Taiyuan 030051, Shanxi , China

²Shanxi Province Key Laboratory of Machine Vision and Virtual Reality, Taiyuan 030051, Shanxi , China

³Department of Simulation Equipment, North Automatic Control Technology Institute, Taiyuan 030006, Shanxi , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Point cloud registration model based on multi-scale feature dynamic fusion

Download full size

Fig. 2. Feature extraction module

Download full size

Fig. 3. Multi-scale feature dynamic extraction module

Download full size

Fig. 4. Residual module

Download full size

Fig. 5. GC-Res module

Download full size

Fig. 6. Visualization of point clouds to be registered. (a) Redkitchen; (b) Lab; (c) Hotel; (d) Studyroom

Download full size

Fig. 7. Feature visualization. (a) Redkitchen; (b) Lab; (c) Hotel; (d) Studyroom

Download full size

Fig. 8. Robustness comparison

Download full size

Fig. 9. Registration visualization experiment. (a)(b) Low overlap to be configured with punctuality cloud; (c)-(h) on-time cloud to be configured under influence of noise and high complexity

Download full size

Fig. 10. Registration experiment verification

Download full size

Table 1. Comparison of feature matching recall rates for different models
View table
Table 1. Comparison of feature matching recall rates for different models
Model FMR Feat. Dim. Time /s
FPFH 0.359 33 0.032
3DMatch 0.596 512 3.210
PerfectMatch 0.947 32 5.515
FCGF 0.952 32 0.011
D3feat 0.958 64 0.008
Proposed 0.962 32 0.009

Table 2. Registration recall rate for different scenarios
View table
Table 2. Registration recall rate for different scenarios
FPFH 3DMatch FCGF /% Proposed /%
Kitchen 36 85 93 93
Home 1 56 78 91 97
Home 2 43 61 71 87
Hotel 1 29 79 91 97
Hotel 2 36 59 87 87
Hotel 3 61 58 69 81
Study 31 63 75 80
Lab 31 51 80 77
Average 40 67 82 87

Table 3. Ablation experiment
View table
Table 3. Ablation experiment
Model FMR /% STD
FCGF 95.2 0.029
GC-Res+FCGF 95.5 0.027
MSFDF+FCGF 95.9 0.029
Proposed 96.2 0.025

Tools

Get Citation

Copy Citation Text

Jing Liu, Yuan Zhang, Le Zhang, Bo Li, Xiaowen Yang. Point Cloud Feature Extraction Network Based on Multiscale Feature Dynamic Fusion[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0415001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites