Preoperative and Intraoperative Cross‐Source Point Cloud Registration Based on Attention Mechanism Enhancement

Tianbao Liu; Jiahui Guo; Yibin Song; Wei Wang; Bo Wu; Nan Zhang

doi:10.3788/CJL241199

Chinese Journal of Lasers, Volume. 52, Issue 3, 0307103(2025)

Preoperative and Intraoperative Cross‐Source Point Cloud Registration Based on Attention Mechanism Enhancement

Tianbao Liu^1,2,3, Jiahui Guo^1,2,3, Yibin Song^1,2,3, Wei Wang⁴, Bo Wu^1,2,3、*, and Nan Zhang^1,2,3

Author Affiliations

¹School of Biomedical Engineering, Capital Medical University, Beijing 100069, China

²Beijing Key Laboratory of Fundamental Research on Biomechanics in Clinical Application, Capital Medical University, Beijing 100069, China

³Laboratory for Clinical Medicine, Capital Medical University, Beijing 100069, China

⁴Department of Orthopedics, Xuanwu Hospital of Capital Medical University, Beijing 100053, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Registration method for farthest point sampling cross-source point cloud based on attention mechanism

Download full size

Fig. 2. Pre-operative local-intra-operative point cloud registration diagram based on attention mechanism

Download full size

Fig. 3. Feature aggregation module

Download full size

Fig. 4. Preoperative FPS point cloud in local regions and the preoperative and intraoperative point cloud pairs. (a) Preoperative point cloud in local regions; (b) preoperative and intraoperative point cloud pairs

Download full size

Fig. 5. Registration results of 001-L2, 002-L3, 003-L1, 004-L1, 005-L3, 006-L1, and 007-L1

Download full size

Fig. 6. Intraoperative point clouds with different levels of Gaussian noise. (a) No noise added; (b)adding noise with a standard deviation of 0.25; (c) adding noise with a standard deviation of 0.50

Download full size

Table 1. Test data sources and initial attitudes

View table

Table 1. Test data sources and initial attitudes

Exposed site	Source	Rotation /（°）			Translation /mm
Exposed site	Source	X	Y	Z	X	Y	Z
001-L2	Ⅰ	148.53	48.47	138.47	114.08	-64.99	-103.52
002-L3	Ⅰ	109.48	28.19	79.30	-55.67	7.23	40.84
003-L1	Ⅱ	-76.59	68.13	-88.32	-98.28	-59.36	26.07
004-L1	Ⅱ	104.67	70.86	97.17	398.10	34.14	19.08
005-L3	Ⅱ	96.70	42.06	92.16	216.39	-44.52	76.20
006-L1	Ⅱ	96.94	22.41	82.55	-89.11	45.34	42.65
007-L1	Ⅱ	89.39	58.15	82.35	-57.78	-42.73	38.71

Table 2. Preoperative‒intraoperative point cloud registration error and running time

View table

Table 2. Preoperative‒intraoperative point cloud registration error and running time

Exposed site	Registration algorithm	$R^{e}$ /（°）	$T^{e}$ /mm	Time /s
001-L2	FPS+FPFH^［17］	14.81	9.31	65.61
	FPS+FastReg^［23］	6.17	7.62	22.58
	FPS+ours	4.52	3.51	18.57
002-L3	FPS+FPFH^［17］	8.12	4.84	166.35
	FPS+FastReg^［23］	5.52	6.74	23.41
	FPS+ours	3.58	3.90	19.24
003-L1	FPS+FPFH^［17］	8.52	4.88	207.69
	FPS+FastReg^［23］	3.34	5.10	23.98
	FPS+ours	2.58	3.20	19.53
004-L1	FPS+FPFH^［17］	7.75	17.81	52.87
	FPS+FastReg^［23］	3.14	3.20	22.85
	FPS+Ours	2.88	2.95	17.61
005-L3	FPS+FPFH^［17］	3.43	10.25	66.98
	FPS+FastReg^［23］	4.33	4.94	22.14
	FPS+ours	1.79	2.62	19.29
006-L1	FPS+FPFH^［17］	4.28	7.68	164.09
	FPS+FastReg^［23］	5.44	7.78	21.22
	FPS+ours	2.51	4.19	17.77
007-L1	FPS+FPFH^［17］	1.75	3.50	44.66
	FPS+FastReg^［23］	2.71	3.36	22.71
	FPS+ours	2.29	2.09	17.43

Table 3. Maximum, median, mean of matched feature pairs and the number of key points after encoding

View table

Table 3. Maximum, median, mean of matched feature pairs and the number of key points after encoding

Exposed site	Registration algorithm	Median	Mean	Max	Key points after encoding
001-L2	FPS+FPFH^［17］			266	30000
	FPS+FastReg^［23］	50	64	119	512
	FPS+ours	20	27	343	512
002-L3	FPS+FPFH^［17］			111	30000
	FPS+FastReg^［23］	47	46	128	512
	FPS+ours	23	22	114	512
003-L1	FPS+FPFH^［17］			585	30000
	FPS+FastReg^［23］	50	57	128	512
	FPS+ours	29	37	186	512
004-L1	FPS+FPFH^［17］			428	30000
	FPS+FastReg^［23］	45	59	162	512
	FPS+Ours	28/43		180	512
005-L3	FPS+FPFH^［17］			1026	30000
	FPS+FastReg^［23］	31	38	84	512
	FPS+ours	34	45	116	512
006-L1	FPS+FPFH^［17］			315	30000
	FPS+FastReg^［23］	48	56	141	512
	FPS+ours	29	45	176	512
007-L1	FPS+FPFH^［17］			1121	30000
	FPS+FastReg^［23］	47	56	182	512
	FPS+ours	35	48	305	512

Table 4. Ablation experiment
View table
Table 4. Ablation experiment
Network ${\bar{R}}^{e}$ /（°） ${\bar{T}}^{e}$ /mm
Initial 2.87 3.22
Removal of cross-attention mechanism 4.81 6.62
Removal of all attention mechanism 6.52 10.57

Table 5. Noise experimental results
View table
Table 5. Noise experimental results
Noise standard deviation magnitude ${\bar{R}}^{e}$ /（°） ${\bar{T}}^{e}$ /mm
δ=0 2.87 3.22
δ=0.25 7.48 6.18
δ=0.50 9.55 7.11

Tools

Get Citation

Copy Citation Text

Tianbao Liu, Jiahui Guo, Yibin Song, Wei Wang, Bo Wu, Nan Zhang. Preoperative and Intraoperative Cross‐Source Point Cloud Registration Based on Attention Mechanism Enhancement[J]. Chinese Journal of Lasers, 2025, 52(3): 0307103

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites