LDASH: A Local Feature Descriptor of Point Cloud with High Discrimination and Strong Robustness

Lei Zhou; Bao Zhao; Dong Liang; Zihan Wang; Qiang Liu

doi:10.3788/LOP231825

Laser & Optoelectronics Progress, Volume. 61, Issue 12, 1215007(2024)

LDASH: A Local Feature Descriptor of Point Cloud with High Discrimination and Strong Robustness

Lei Zhou¹, Bao Zhao^1,2、*, Dong Liang^1,2, Zihan Wang¹, and Qiang Liu¹

Author Affiliations

¹School of Internet, Anhui University, Hefei 230039, Anhui , China

²National Joint Local Engineering Research Center for Agroecological Big Data Analysis and Application Technology, Anhui University, Hefei 230601, Anhui , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(14)

Fig. 1. Constructive illustration of the surface change angle θ_j

Download full size

Fig. 2. Generating process of LDASH descriptor. (a) Extraction of local surface around the keypoint p on the chicken model from the U3M dataset; (b) construction of LRA on the local surface; (c) uniform division of local space in the radial direction; (d) calculation of five attribute values for each neighboring point; (e) statistical analysis of the distribution of the five attributes; (f) statistical histograms of five geometric attributes; (g) generation of five sub-histograms; (h) final histogram H_LDASH

Download full size

Fig. 3. Distribution statistics of six attribute values (r, h, α, β, γ, and $D_{D W A V}$ ) in a large number of local neighborhoods on the B3R dataset. (a) Radial distance r; (b) height distance h; (c) angle α; (d) angle β; (e) angle γ; (f) $D_{D W A V}$

Download full size

Fig. 4. RPC performance evaluation results of five histograms on the B3R dataset

Download full size

Fig. 5. Three typical point clouds in B3R, QULD, U3M, U3OR and S3R dataset. (a) B3R dataset; (b) QuLD dataset; (c) U3M dataset; (d) U3OR dataset; (e) S3R dataset; (f) K3R dataset

Download full size

Fig. 6. Parameter settings for the LDASH descriptor (the large solid markers indicate the selected parameter values). (a) N_r、N_α、N_β、N_γ、N_h and $N_{D_{D W A V}}$ ; (b) w₁、w₂、w₃、w₄ and w₅; (c) R_s

Download full size

Fig. 7. RPC performance evaluation results of nine descriptors on six datasets (the number in parenthese is the AUC_prvalue of the corresponding descriptor, arranged in descending order). (a) B3R dataset; (b) QuLD dataset; (c) U3M dataset; (d) U3OR dataset; (e) S3R dataset; (f) K3R dataset

Download full size

Fig. 8. Robustness assessments of nine descriptors at different levels of six nuisance (the number in parenthese represents the average AUC_pr value over the whole curve, arranged in descending order). (a) Gaussian noise; (b) mesh decimation; (c) Gaussian noise combined with mesh decimation; (d) distance between boundary and keypoint; (e) keypoint localization error; (f) occlusion

Download full size

Fig. 9. Time required to generate nine descriptors under different support radii (the number in parenthese is the average time consumption of the corresponding descriptor, arranged in descending order)

Download full size

Fig. 10. Correct registration rates of five transformation estimation algorithms combined with nine descriptors on U3M dataset (the number in parenthese represents the average correct registration rate over the whole curve, arranged in descending order)

Download full size

Fig. 11. Two registration cases of CG-SAC transformation estimation algorithm combined with nine descriptors on U3M dataset

Download full size

Table 1. Characteristics of the six baseline datasets

View table

Table 1. Characteristics of the six baseline datasets

Dataset	Scenario	Challenge	Scanner	Quality	#Model	#Scene	#Matching pair
B3R	Retrieval	Gaussian noise，mesh decimation	Cyberware 3030 MS	High	6	6	6
U3M	Registration	Occlusion	Minolta vivid 910	High	‒	75	433
U3OR	Recognition	Occlusion，clutter	Minolta vivid 910	High	5	80	188
QuLD	Recognition	Occlusion，clutter，real noise	NextEnigne	Low	5	80	240
S3R	Registration	Occlusion，real noise，outliers，missing regions	SpaceTime Stereo	Low	‒	57	240
K3R	Registration	Occlusion，real noise，missing regions	Microsoft Kinect	Low	‒	69	284

Table 2. Parameter setting process for LDASH descriptor

View table

Table 2. Parameter setting process for LDASH descriptor

	N_r	N_α	N_β	N_γ	N_h	$N_{D_{D W A V}}$	w₁	w₂	w₃	w₄	w₅
Setting N_r	2‒20	15	15	15	5	5	1.0	1.0	1.0	1.0	1.0
Setting N_α	5	2‒20	15	15	5	5	1.0	1.0	1.0	1.0	1.0
Setting N_β	5	18	2‒20	15	5	5	1.0	1.0	1.0	1.0	1.0
Setting N_γ	5	18	15	2‒20	5	5	1.0	1.0	1.0	1.0	1.0
Setting N_h	5	18	15	17	2‒20	5	1.0	1.0	1.0	1.0	1.0
Setting $N_{D_{D W A V}}$	5	18	15	17	13	2‒20	1.0	1.0	1.0	1.0	1.0
Setting w₁	5	18	15	17	13	8	0.1‒2.0	1.0	1.0	1.0	1.0
Setting w₂	5	18	15	17	13	8	1.0	0.1‒2.0	1.0	1.0	1.0
Setting w₃	5	18	15	17	13	8	1.0	1.6	0.1‒2.0	1.0	1.0
Setting w₄	5	18	15	17	13	8	1.0	1.6	1.0	0.1‒2.0	1.0
Setting w₅	5	18	15	17	13	8	1.0	1.6	1.0	0.8	0.1‒2.0

Table 3. Parameter settings for nine descriptors

View table

Table 3. Parameter settings for nine descriptors

Descriptor	Support radius	Dimensionality	Length
SI ^［17］	15r_m	15×15	225
RoPS^［31］	15r_m	3×3×3×5	135
SHOT^［20］	15r_m	8×2×2×11	352
TriSI^［4］	15r_m	15×15×3	675
TOLDI^［21］	15r_m	3×20×20	1200
MaSH^［30］	15r_m	5×（10+8+10）	140
SDASS^［22］	15r_m	15×5×5‒30	345
DLFS^［23］	15r_m	5×（9+12+12+15）	240
LDASH	15r_m	5×（18+15+17+13+8）	355

Tools

Get Citation

Copy Citation Text

Lei Zhou, Bao Zhao, Dong Liang, Zihan Wang, Qiang Liu. LDASH: A Local Feature Descriptor of Point Cloud with High Discrimination and Strong Robustness[J]. Laser & Optoelectronics Progress, 2024, 61(12): 1215007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Machine Vision

Received: Jul. 31, 2023

Accepted: Oct. 13, 2023

Published Online: May. 20, 2024

The Author Email: Bao Zhao (zhaobao625@163.com)

DOI:10.3788/LOP231825

CSTR:32186.14.LOP231825

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology