Mini Project 1

Was your social distance detector effective at detecting potential violations?

My social distance detector was somewhat effective at detecting potential violations. The video I chose depicted lots of people walking in a public place. It is hard for me to tell if the people are walking farther enough away from each other because the video does not allow for me to accuratel gauge distance between people. However, I think that this issue with depth perception affects the social distance detector. The detector classifies most people as safe; however, a lot of people are violating social distancing. The detector does identify some violations.

Are you able to describe how the distance detector is applying its calculations of either being safe or noting a violation?

The social distance detector uses computer vision to detect violations. Based on the code, it seems that there is a way for the computer to check the distance between people based on the distance between the rectangles it forms around them. There appears to be a function for checking the distance, a function for applying a deep neural network and weights, and a function for processing the image.

Do you think this approach would be effective for estimating new infections in real time?

I think that this approach could be effective, depending on the situation. If the camera recording the footage which the detector was applied to was positioned correctly, the detector may be quite effective. For example, the detector was fairly effective at detecting social distancing violations in the gymnastics video. However, it was less effective at detecting violations in the video of lots of people walking. I think that if the camera for this detector was located at eye level, it would work well in a setting where there are not many people (for example, a classroom or library). However, it is clearly not as effective at detecting violations in a busy public place (even at eye level), and I would imagine that it may have a more difficult time if the camera was located significantly far above or below eye level because of the way it seems to identify people.

How would you implement such an approach in response to the COVID-19 pandemic we are currently experiencing?

It may be helpful to implement this approach in public spaces where it would be effective (i.e. less crowded public spaces). I believe it would work well in campus spaces such as classrooms, dorm common rooms, labs, and libraries. In most public spaces, it would be difficult to enforce strict social distancing even if violations were able to be detected. However, students using campus spaces are can more easily be bound by these restrictions. Perhaps violation of social distancing protocol would cause some sort of beeping in the space which would alert people to the violation. However, this could end up being more harmful than helpful, especially if the detector was not 100% effective.

What limitations or improvements might you include in order to improve your proposed design?

The biggest limitation of this detector is its slow speed. If I were to implement the detector in real life, I would want to modify it so that it could run in real time. I wonder if using convolutions and pooling would help accomplish this - perhaps there would be a way to extract the features identifying a person. Another limitation is the perspective of the camera; ideally, the detector would be improved so that it could take as input footage from various angles.

Watch my video here!