Answers to project questions
- Yes, my social distance detector seemed to work very well with my video. While going back and watching, I tried to estimate the distances between the people myself, and I agreed very closely with what the detector was assessing.
- I think if a model like this were to be implemented in various locations, it would be good at determining just how many social distancing violations there are. For example, if we implement this in one location, there may be
many more or many less violations in that area compared to the rest of the city/state/country. If this were to be implemented in many different places, it would be like taking a sample of the ‘population’. A good way to implement this
could possibly be choosing a proportional amount of different locations relative the traffic of cities (like I live in Virginia Beach so there would be many beach samples vs few outlet mall samples). This way the samples accurately
represent the population of the city. Each cities data can be compiled to estimate cases for the state, and then the same can be done with the states to estimate the national numbers.
- One limitation that immediately stood out to me was the detector’s inability to recognize when certain people are walking together, such as a husband and wife who are both out walking. I feel like situations like this should not count
as social distancing violations, as the two already live together. Maybe this could be fixed by trying to recognize when people are very close together and are maintaining the same trajectory and then exempting them from counting as a violation?
Of course this sounds like it would be difficult to implement, but if possible, could make the detector better at estimating the amount of actual violations. There does seem to sometimes be an issue with perspective, especially if the camera angle
is poor. I am not sure what could be done about this, but this appears to be a slight issue with the detector.
Output Video