For our demand distribution comparison to work, we want to remove days with crashes that put us in weird places on the demand curve for that sensor (keeping in mind that each sensor will have a different demand curve - for instance, a downstream bottleneck may cause a sensor to saturate below 2400 veh/lane/hr). We plan to do this by creating a monthly or moving-window 75th percentile distribution of occupancy, and computing the Fréchet distance for each day from that distribution. I suspect results will be bimodal - a peak for normal days and a peak for days with crashes. Probably need to use only the same day of the week.
For our demand distribution comparison to work, we want to remove days with crashes that put us in weird places on the demand curve for that sensor (keeping in mind that each sensor will have a different demand curve - for instance, a downstream bottleneck may cause a sensor to saturate below 2400 veh/lane/hr). We plan to do this by creating a monthly or moving-window 75th percentile distribution of occupancy, and computing the Fréchet distance for each day from that distribution. I suspect results will be bimodal - a peak for normal days and a peak for days with crashes. Probably need to use only the same day of the week.