Landmark graphs provide a means for surface registration, based on determining subgraph isomorphism to find scene-to-scene correspondences. Surface data used herein included both range and colour imagery. Images were acquired of a static scene from a moving sensor. The continuous motion allowed the sensor position to be predicted, which helped stabilize graph formation. Landmarks were determined using the KLT corner detector. Graph structure was established using nodes (landmarks) and edges that agreed well with predicted locations. Subgraph matching was approximated using the LeRP algorithm. A 6 DOF rigid transformation including translation and rotation was found via Horn’s method. Test results on real and synthetic images indicate that a substantial speed improvement is possible, with greater determinism than ICP, while maintaining accuracy. Tests incorporated relatively large image displacements, spanning up to 30% of the sensor FOV for the image stream. Mean absolute errors remained under 0.8% FOV. Mean compute rates were ≈ 10 Hz with standard deviation ranging 6-9%, for an image size of 200x200. Tests were run on a 900 MHz PC. 141 test trials are reported, with comparisons against a fast version of ICP.


Electrical and Computer Engineering



URL: http://digitalcommons.calpoly.edu/eeng_fac/7