Point Cloud Registration Algorithm for Augmented Reality

Fig. 3. Registration results. (a) Buddha for registration; (b) result of Super4PCS; (c) result of FPFH; (d) result of TrimmedICP; (e) result of our method

Download full size

Fig. 4. Registration results of real data. (a) (f) Point clouds for registration; (b)(g) result of Super4PCS; (c)(h) result of FPFH; (d) result of TrimmedICP; (e)(j) result of our method

Download full size

Table 1. Comparison of noise recognition accuracy
View table
Table 1. Comparison of noise recognition accuracy
Method COIR /% FPR /%
RANSAC 94.4 64.2
RPCA 88.1 8.2
Our method 92.8 8.6

Table 2. Comparison of point cloud registration algorithms

View table

Table 2. Comparison of point cloud registration algorithms

Method	Time /s			e_R			e_t
Method	Bunny	Dragon	Buddha	Bunny	Dragon	Buddha	Bunny	Dragon	Buddha
TrimmedICP^[8]	38.5	24.1	80.6	0.012	0.014	0.015	0.248	0.254	0.183
FPFH^[10]	57.6	32.4	176.3	0.019	0.025	0.021	0.756	0.561	0.282
Super4PCS^[14]	29.5	21.3	63.5	0.018	0.021	0.016	0.425	0.364	0.315
Our method	2.3	0.8	3.2	0.013	0.019	0.016	0.312	0.319	0.297

Table 3. Comparison of registration algorithms on Kinect point cloud
View table
Table 3. Comparison of registration algorithms on Kinect point cloud
Method Time /s e_R e_t
Sofa Can Sofa Can Sofa Can
TrimmedICP^[8] 56.5 40.2 0.011 0.012 0.157 0.098
FPFH^[10] 75.3 50.2 0.016 0.086 0.373 0.325
Super4PCS^[14] 23.5 17.2 0.016 0.035 0.215 0.217
Our method 2.8 1.6 0.012 0.021 0.213 0.129