Embedded Files
This part 3 is more for filling in the gaps that I may have missed with the other pages on the 360 camera. These are:
- I have had some issues with using imagemagick and feel that the process should be better explained.
- I have had a hard time recreating the original dewarped images.
- Better measurements for dewarping have resulted in better 360 images.
- Correct these issues before the machine is built so maybe the LSD-SLAM will work (better).
First I used imagemagick for windows to convert the images. I downloaded imagemagick and ran the convert program on the images; it seemed to work pretty good and gave what I thought were acceptable results. I tried again a few days later and could not get the program to run on files outside the directory the program was in or run the program in the directory the image was in (the images were also moved to that directory...Still didn't work). I decided that linux should be used to dewarp the images. In linux, the command resulted in dissimilar results from windows. I messed up somewhere and didn't document my actions properly. Because of this I started over with the imagemagick command.
I used the command originally used and checked the results. These were not what I wanted. I then searched how to use the dewarp feature. I tried different ways using dewarp: +DeWarp, -DeWarp, 0, -1 and then the 3 sets of pixel measurements. DeWarp 0 and -1 gave results close to what I was looking for, but in certain areas of the image things looked wrong. These areas were the portion of the image within the reflection that were cut off due to the camera being too close or the minimum zoom being too high. The parts of the circle that were not within the limits of the image were filled in with lines of color in the dewarped images.
To correct the strange artifacts I used the -virtual-pixel Black option. This colored in the missing areas with black. It was after I used this command that I learned that the cut off areas were causing this problem. It will be more obvious once you see the image.
The way that the command works for dewarping images is like this: -DeWarp 'outside radius in pixels, inside radius in pixels, center x, center y, starting angle, and ending angle'. After using this command, I realized that the original images did not match reality due to having the world be backwards. To correct this the -flip command was used. From all of this testing Images now look like they should and have more of the dewarped image visible.
This is the command I used in order to produce the better dewarped image:
convert IMG_1227.JPG -filter Cubic -virtual-pixel Black -distort DePolar '1257,183 1308,960 0,360' -flip -resize 100x20% again1.jpgOriginal image:
Results from new measurements and understanding:
The rounded areas are the portions of the mirror that were cut off the picture. More distance may be necessary between the camera and the parabolic mirror. This image shows that the original dewarped attempt (below) was mirrored and not capturing as much of the environment due to poor measurements, lack of understanding, and included too much of the red ring as a lower border.
After all of this, I have also learned that for good performance with outdoor SLAM, you need to have a camera with a global shutter. This means that instead of shooting pictures as most cameras do: rolling shutter--one pixel at a time, all of the pixels capture an image at the same time. This will provide less blur in images and most likely no jello effect in video. The problem is the cost of many global shutter cameras. The cheapest camera that I could find at mouser was $75.00 without a controller board $300.00 with one. I am almost certain that I can get something to work for under $100.00. If any one has any ideas, please send me an e-mail.
9/24/17
Google Sites
Report abuse