Structure from Motion vs. the Kinect: Comparisons of River Field Measurements at the 10-2 to 102 meter Scales

Thursday, 18 December 2014
Mark A Fonstad, University of Oregon - 1299, Eugene, OR, United States and James T. Dietrich, Dartmouth College, Hanover, NH, United States
At the very smallest spatial scales of fluvial field analysis, measurements made historically in situ are often now supplemented, or even replaced by, remote sensing methods. This is particularly true in the case of topographic and particle size measurement. In the field, the scales of in situ observation usually range from millimeters up to hundreds of meters. Two recent approaches for remote mapping of river environments at the scales of historical in situ observations are (1) camera-based structure from motion (SfM), and (2) active patterned-light measurement with devices such as the Kinect. Even if only carried by hand, these two approaches can produce topographic datasets over three to four orders of magnitude of spatial scale. Which approach is most useful?

Previous studies have demonstrated that both SfM and the Kinect are precise and accurate over in situ field measurement scales; we instead turn to alternate comparative metrics to help determine which tools might be best for our river measurement tasks. These metrics might include (1) the ease of field use, (2) which general environments are or are not amenable to measurement, (3) robustness to changing environmental conditions, (4) ease of data processing, and (5) cost. We test these metrics in a variety of bar-scale fluvial field environments, including a large-river cobble bar, a sand-bedded river point bar, and a complex mountain stream bar.

The structure from motion approach is field-equipment inexpensive, is viable over a wide range of environmental conditions, and is highly spatially scalable. The approach requires some type of spatial referencing to make the data useful. The Kinect has the advantages of an almost real-time display of collected data, so problems can be detected quickly, being fast and easy to use, and the data are collected with arbitrary but metric coordinates, so absolute referencing isn’t needed to use the data for many problems. It has the disadvantages of its light field generally being unable to penetrate water surfaces, becoming unusable in strong sunlight, and providing so much data as to be sometimes unwieldy in the data processing stage.