A sample quad-frame illustrates the contents of the MPEG animation below.
The upper left quadrant of each frame is the original 256x512 pixel image, taken from a 115 degree wide-angle lens, subsampled to 128x256 pixels. (In the sample frame you can see the back of the chair in the lower left corner. This chair is the closest obstacle.) The upper right quadrant shows the optical flow "needle" plot calculated from a 32x64 subsampled version of the original image, with the needle diagram itself subsampled to 16x32 for clarity. (Compare with the full-resolution 32x64 "needle" plot .)
The lower right quadrant of each frame shows a time-space plot of the 1-D divergence estimates, median-filtered in space but not in time (11x17 pixel window). In these plots brighter areas indicate higher divergence and thus closer objects. (Each individual plot is automatically contrast-adjusted to improve clarity.) Time increases upward in these plots: thus, the 1-D divergence vector for the current frame is found at the top of the subimage. The lower left quadrant of each frame shows the same divergence plot but median filtered in time as well as in space (11 frames x 11 pixels). The second filtered divergence estimate is used at each frame to control the robot's steering. (In the sample frame the divergence for the chair is clearly visible. The robot is about to make a right turn due to the chair's proximity.) Because of the temporal median's width, the filtered divergence information has a latency of 5 frames, which is about one second. This effect is offset by the system's ability to sense objects up to several meters in advance.
These sequences are in the format described above and show collision tests at 20, 30, 40, and 50 cm/s respectively beginning from a distance of approximately 4 meters. These tests were performed to characterize the collision detection performance under the best possible circumstances; therefore no steering or speed adjustments were made during these trials.