Metrics and Optimization Techniques for Registration of Color to Laser Range Scans
Chad Hantak, Anselmo Lastra · 2006
Web found previous intensity-based techniques for automatically registering color images to three-dimensional laser scanned scenes to be inadequate. The similarity metric used to score the registration creates a number of local minima that inhibits searching via Powell's multidimensional minimization algorithm, a gradient-descent technique. To find the best metric for general environment scanning, we examine the results of different information-theoretic metrics. Our examination leads us to the conclusion that gradient- descent based techniques are not a good choice for unsupervised automatic registration for images from environment scans. However an unsupervised process is possible through global-optimization techniques at the cost of longer processing times.