Personal reflection
I looked at all the feedback this week for our team and wrote down some ideas that Colin, Zach, and I were brainstorming already. Generally, I feel like I should not have interrupted Zach in the presentation, he knew his material and I felt embarrassed to see the feedback that peers/faculty thought otherwise. It is a teammate attitude change that I will uphold moving forward. I do also think I got a little carried away with predictions on the statistics of our application. I feel like I’m used to classes and assignments where the stakes are quite different, and even compared to the other CV projects within the class, ours stands out in its real user base. I am excited to work for what we’re doing, but worried about how much we can achieve. It is not only a change of number (say from 1% to 0.05%), but we have to follow through on actually doing it.
Some design ideas
We do plan to keep the camera mounted on the shoulder using some sort of strap-on that comes for GoPros. I have a GoPro and there seems to be some vague documentation on connecting and streaming video to a Jetson. (An example is here.)
Another idea would be to keep the Jetson and the camera in a self-contained unit (maybe 3D-printed), but I do not know how much power/energy it would give off, and whether that would be harmful to the user. I wouldn’t want to wear something hot to my chest or near my body.
Algorithm work
I also research the background of crosswalk detection itself. I wrote down some potential steps based on my research of edge detection algorithms I think I would use. More research is needed on how to actually track objects in video, which is not something I’ve delved too much into before. Here are the steps I came up with:
- I would like to start with processing a frame and filtering out all the white I notice in that frame.
- Erode frames by structuring the frame. This is necessary to remove the noise and identify some finer points in the image. (A link on why this is necessary)
- Identify contours using cv.findContours, and draw tracking lines on each of the contours
- Find the median points in the contour. I have some ideas for the actual linear regression, which could be a simpler model, but also found this specific application for image alignment called RANSACÂ https://learnopencv.com/tag/ransac/
- Basically, after finding the median points, we want to use those along with the contours found to create a middle intersection point. This will be used to calculate the direction vector from the camera compared to that points.
- This alignment vector will help decide the walking direction for the user.
Next week?
I am a little behind on project work due to attending career fairs and interviews this past week. I feel good about my algorithm idea and want to start a github repo in the coming week and sketch out some outline code for it. I also want to, if not complete the pseudocode, find images and videos of crosswalks to test my algorithm on, as well as do research on how video tracking would work. I have a direction vector and a camera’s center point, but how do these two connect? I also hope to help Zach with collecting pictures of walk signs and I’m trying to take pictures of as many as I can while walking to and from campus! I also let the professors know I would need AWS, but I’ve always worked locally to train my ML algorithms even for classes, so I would like to play around with how that works!