sonjageorgievska / arena Goto Github PK

Conferences or journals

e.g. very relevant based on other papers that use mobile data or crowd analytics : EPJ Data Science
http://link.springer.com/journal/volumesAndIssues/13688

Because the smoothing may partially distribute them inside the stadium, and because it is unlikely that outsiders would be detected (wall etc, what Jan said on Friday 10/06).

More concretely, slide #3 from the Friday 10/06 presentation. We should not exclude those detected outside the region.

In this way, we actually don't need to adjust the integrals inside to be exactly 1, because of the added effect from the outsiders. Then both the full and the dotted lines at slide 15 should "overlap".

The stadium boundary is artificial anyway.

We only detect a fraction x of the people, because not everybody has a smart phone and because some people have 2 smartphones. The number x should be found online in some reports or papers. Then our calculations should scale to take into account x, too.

One solution is:

1. make a statistics of e.g. what (average or median?) percentage p of the addresses detected in a time window is randomized.
2. exclude all randomized addresses during calculation of density
3. after all calculation is done, scale the histogram to take into account that p% of data was ignored.

After all, we only detect a fraction x of the people, because not everybody has a smart phone and because some people have 2 smartphones. The number x should be found online in some reports or papers. Then our calculations should scale to take into account x, too.

Either by Monte Carlo simulations or by analyzing video data

Our method involves the superposition of a series of Gaussian 'bumps' which are centred on top of the fitted positions. To do so, we smooth each position separately and keep it in memory as a density histogram.

However, this is mathematically very similar to kernel density estimation and the use of radial basis or weight functions, which does not require to build separate histograms. We have to check whether using one implementation or the other has any effect on the results, as this choice might influence the amount of truncation and approximation introduced in the statistical modelling, and thus the accuracy.