Aaron Wilhelm

This work uses microscopic images of natural ground texture to determine a robot’s position within a map with errors of tens of micrometers. It examines how optical magnification and image-matching uncertainty affect positioning accuracy, and addresses calibration, camera modeling, and focusing at high magnification. Across 1,007 held-out gantry queries per surface on printed wood grain, terrazzo, and concrete, the system localizes every query with 28–73 µm translation RMSE relative to a calibrated gantry reference. Its position estimates agree with separate photometric registration estimates to within approximately 9 µm RMSE on valid comparison subsets. Further experiments demonstrate localization on a low-cost mobile robot using gantry-collected maps, as well as measurement of micrometer-scale displacement caused by thermal expansion.

GT-BoW is a geometry-aware image-retrieval method for ground-texture localization. It combines an approximate k-means vocabulary and soft assignment with constraints derived from the fixed scale and consistent orientation of downward-facing images. Separate high-accuracy and high-speed variants address the different demands of global localization and loop-closure detection. On the full benchmark, the method increases global-localization mean average precision from 0.026 to 0.559 and detects nearly 3× as many loop closures as a standard DBoW baseline.

Pruning weight helps growers assess vine vigor and plan vineyard management, but conventional measurement methods are labor-intensive or require specialized equipment. This work presents an affordable computer-vision method that combines a smartphone camera with structured light. It achieves R² = 0.80 on vertical-shoot-positioned vines, outperforming prior vision-based approaches, and reports initial results (R² = 0.29) for high-cordon vines, a training system not previously evaluated with computer vision.