Below is an animated GIF showing Sequence 4 its entirety. The original images from the camera are at 1024x1024 12-bit monochrome, and the dewarping and processing done takes the images to 360x135 8-bit monochrome. I am presenting the animation in this small format due to bandwidth and memory limitations. A larger version is also available. There is a slight calibration problem here which is manifested by the slight sinusoidal appearance of the camera rim along the bottom of the dewarped images. That should be fixed some time soon.
Each pixel in the dewarped image corresponds to a particular azimuth and elevation angle in the camera frame of reference. The direction in which I was pulling the sled here actually corresponds to about column #270 (out of 360) in these images. That means the center column points to the left of the motion, the left edge is to the right of the motion, etc. It's more clear in the following diagram:
When the sled translates forward, we expect points between column 90 and column 270 to move left, and points between 270 and the 360 (right edge), as well as between 0 (left edge) and 90 to move right. Points that are lower in the image tend to be closer to the camera (flat ground plane assumption) so the points at the bottom of the image tend to move further between frames, and points which are in front move downward and points behind upward. Again, I appeal to a diagram to explain this better:
The sequence here consists of only translation forward. If the camera stayed in one place and simply rotated, then all points on the dewarped image plane should appear to move in the same direction by the same amount. That sort of behavior can be seen in other data sets which will be posted shortly.
Because the camera pitches and rolls as the sled slides across the blue ice, the image warps in an interesting way. If you think about the camera imaging a truly flat plane, then if the camera pitches forward, the horizon will move up on the image plane in front of the camera, and down behind the camera. To the left and right, we will see an apparent local rotation of the horizon. What results is a sinusoidal appearance to the horizon. This effect can be seen in the sequence above as well.
Right now, I am working on computing optical flow across image sequences to get an idea of how to track points in the dewarped images. From the optical flow, it will be possible to compute feature point locations from panoramic camera motion.
Back to Panoramic