Research

Image Feature Extraction (2011-)

Building an Automatic Grading System for Programming Contests (2008-)

Short-term Forecast of Precipitation Using an Optical Flow Algorithm (2011)

We developed and implemented an algorithm that estimates the current motion of rain areas by matching multiple past images and predicts their future motion from the estimated motion. As a result, we improved the prediction accuracy of whether there is rain of 1 mm or more by about 25 percent, compared with Nowcast, the forecast announced by the Japan Meteorological Agency. In addition, because the computation is fast enough, we became able to forecast on a finer mesh minute by minute.

Figure 1: Radar map
at 2011-01-30 00:00

Figure 2: Radar map
at 2011-01-30 00:55

Figure 3: Forecast of Nowcast
for 2011-01-30 00:55

Figure 4: Forecast of the proposed method for 2011-01-30 00:55
(Gray indicates that the rain intensity is referenced from points that lay outside the radar image at 00:00.)

Figure 4 shows the forecast for 55 minutes later, produced by the proposed method based on Figure 1. Compared with the Nowcast forecast by the Japan Meteorological Agency for the same time (Figure 3), we can see that Figure 4 matches the actual rain areas (Figure 2) more closely. This is likely because Nowcast does not take the local motion of rain areas into account, and therefore cannot detect phenomena such as rain areas being unable to cross mountains, or the large difference in rain-area motion between the Sea of Japan side and the Pacific side.

Because conventional optical flow detection algorithms did not achieve sufficient accuracy for the optical flow detection used in this research, we devised and implemented a new algorithm. From two or more images taken with a stereo camera, this algorithm detects the optical flow and, by interpolating and extrapolating it, can also handle the images as if 3D information were given.

Figure 5: Left-side view

Figure 6: Right-side view

Figure 7: Image extrapolated to the right of Figure 6 by three times
the distance from Figure 5 to Figure 6

Figure 5 is the image taken with the left lens of the stereo camera, and Figure 6 is the image taken with the right lens. Using these images, we detect the optical flow on Figure 6, and extrapolating the obtained optical flow by a further factor of three yields the image in Figure 7. Moreover, by interpolating and extrapolating at various distances, we can also obtain videos such as [a GIF animation generated using optical flow].

Related Paper

• Kentaro Imajo, Go Hasegawa, Yoshiaki Taniguchi, Hirotaka Nakano, "Short-term Precipitation Forecasting Using an Optical Flow Algorithm," IEICE Technical Report 110(455), 93-98, 2011.

Fast Computation of the Continuous Wavelet Transform (2009-)

Last Modified: January 1, 2012