Next week, a portion of my work will be submitted into our Code's SPIE conference paper. Although I have decided to explore a SIFT-based method for soda can detection, my more immediate goal is to make the most of Dlagnekov's LPR detection algorithm when applied to soda can detection. The Robotics Code at SPAWAR would like to place my soda can detection algorithm directly on Robart III to test the algorithm's efficiency and accuracy. In order to make my algorithm "Robot Friendly", I am going to refine its focus to detect soda cans at a single scale against a static background. This way, my chances of success in reducing the false positive rate and increasing the true positive rate within the next week are much greater than they are for detecting soda cans at multiple scales. Furthermore, I will limit the number of orientations to only two --upright and horizontal. This will require me to train two separate strong classifiers. By keeping my detection method simple, I will be able to place my classifier directly on Robart III without having to recompile anything as well. If this is successful and there is enough time, I will add additional functionality to my algorithm and possibly get Robart III to shoot at the soda cans he sees (using all six barrels of course).
Setting Up The Soda Cans
As illustrated in the photo above, I set up Diet Pepsi Cans on and around a wooden box. This wooden box has been featured in previous SPIE conference papers submitted by the Robotics Code at SPAWAR because it is safe for Robart III to shoot at it. Without modifying the position of the wooden box or the background, I placed the soda cans at different locations and Greg recorded each setup for 5-10 seconds using Robart III's head camera. I then rotated each soda can by about 60 degrees and Greg recorded this new setup as a separate MPEG file. This was repeated 44 times to create 44 different MPEG videos. When setting up soda cans, I tried to make sure that there was just about the same number of horizontally and vertically-oriented soda cans with which I can build my training set.
Extracting JPEG Images from MPEG VideosI used ffmpeg to extract 1-2 JPEG files from each of the 44 MPEG videos. I used the following ffmpeg command for each MPEG video:
ffmpeg -ss 00:00:03 -t 00:00:01 -i ./
As illustrated in the photo above, I set up Diet Pepsi Cans on and around a wooden box. This wooden box has been featured in previous SPIE conference papers submitted by the Robotics Code at SPAWAR because it is safe for Robart III to shoot at it. Without modifying the position of the wooden box or the background, I placed the soda cans at different locations and Greg recorded each setup for 5-10 seconds using Robart III's head camera. I then rotated each soda can by about 60 degrees and Greg recorded this new setup as a separate MPEG file. This was repeated 44 times to create 44 different MPEG videos. When setting up soda cans, I tried to make sure that there was just about the same number of horizontally and vertically-oriented soda cans with which I can build my training set.
1 comment:
Thank you!! I was just looking for something like this, but I never expected to be that easy!
You save me hours of fighting against Matlab or OpenCV!!
Thanks a lot!
Post a Comment