I am attempting a similar feat and asked a similar question.The issue is that counting is a much more complicated problem than recognition. It appears it is possible, but a lot of work because you will need lots of labeled data. The fact that you dont need a precise number helps. You will need to train a custom model with as many examples as possible, and see how it works. I think there is no known rule of thumb, just have to try it. Create labels of the ranges of people you want it to detect. Eg 100-200 ppl, 200-300 ppl and give lots of examples, as many as possible. Im also interested to see how well it works but i dont have the data yet to test.