Risks and Mitigation
We have not tried to get our whole project functioning on an RPi yet. So far the RPi seems to be working as intended, but we have only tried it separately from the project, and have not worked to integrate it yet. This will be attempted starting this coming week, ASAP. If we run into issues we cannot solve, we will reach out, and worst case scenario the ‘final’ project will be run on a laptop. We will also need to redo some tests once the code is running on the RPi, which could lead previously successful tests to fail. We will handle any such cases the way we would handle a failed test on any platform – seek to alter/refine the code to improve efficiency.
Project Changes
There are no changes at this time.
Schedule Changes
A few of our remaining features are a bit behind schedule, as the last week of classes was quite busy – we’d like to improve the live video (very low frame rate), log-in is in progress, and we haven’t had a chance to integrate with the RPi. It is our goal to mostly wrap up these loose ends by the end of this weekend/very beginning of this week, so we can move more exclusively into testing for the week leading up to the live demo.
Testing
We tested notification speed from the pet entering the forbidden zone on the camera to the CV detecting that the pet has entered the forbidden zone. This was done via slow mo video – to mark the points in time when a ‘pet’ entered a forbidden square in real life, versus when the CV detected this (marked by the camera window closing). We wanted this to be under a second, and our results showed that the detection speed was ~0.625 seconds.
Then, we tested the notification speed from the pet entering the forbidden zone to the Web App displaying a notification to the user that a pet has entered the forbidden zone. Once again, this was done via slow mo video, marking when the pet was detected by the CV and when the notification first shows on the web app. The goal was for this to occur in under 10 seconds, and the resulting data showed that this process occurred in ~1.125 seconds. We found that the speed from pet entering the forbidden zone to the web app displaying the notification is most limited by the rate of the frontend of the web app making requests to the backend to check if a notification should be created or not.
Next, we tested the accuracy of CV tracking by sampling certain frames and marking where a ‘human’ thinks the pet is versus where the CV thinks the pet is. We wanted this to be with 1ft, and using the the difference in centers of the bounding boxes as our metric, we were usually within 3-6 inches (measured based on knowing what the distance of two spots in the frame would be, i.e. the distance from the cat’s head to its back).
Lastly, we’ve measured the time between when an animal enters the room and when it is picked up by the CV motion detection (also with slow mo video). This took about ~0.75 seconds on average, well less than our goal of 5 seconds.
Note: Many of these tests (any pertaining to tracking speed/frame rate) will be repeated when we switch platforms from a laptop to an RPi. It is our hope that we have sufficient wiggle room in the laptop numbers that any slow downs on the RPi will not bring us outside of our targets. If needed, we may seek out ways to make our design more efficient.