I work on technology that's parallel to the one you're describing - for urban rail. Many of the issues and the methods for getting data are different, but there are some fundamental problems that apply to all travel delay predictions. (Also, I've learned something about the road side of this stuff in school and at industry conferences.)
Predicting the future is hard, and all ETAs are inherently predictions of the future, with a lower bound set by various hard constraints but essentially no upper bound, because you never know what might happen between now and whatever even you're predicting the time of. Disruptions, such as an accident that blocks up the road, are especially troublesome because a) the time to clear the blockage may be dependent on too many exogenous and initially unknown factors to be reliably predictable with any useful precision, while b) that's when users especially want to see a prediction with useful precision, so they can make effective routing decisions. Whether there's a human in the loop or not, taking into account that there's a disruption and applying that knowledge to ETAs is a hard problem.
I would guess that the predicted ETAs on the variable message signs (VMS) are driven by automatic systems and not dependent on human input. Most likely, they are based on real-time measurements of actual speeds, flows, and possibly travel times (possible technologies discussed below). Those measurements are going to vary continuously, and changes in them will tend to lag the more discontinuous events marking the start and end of a disruption. So, an automatic system calculating ETAs based on them, especially if there's some smoothing algorithm to prevent momentary outliers from messing with the final numbers, will produce ETAs that creep up continuously after a disruption starts and then down continuously after a disruption ends, almost definitely more slowly than an educated human observer's guesses might change in light of knowledge of the disruption itself.
Technologies I'm aware of for automatic detection of traffic conditions:
- Induction loop detectors: These have been around for a long time. There's a rectangular loop of metal embedded in the road. Vehicles going over it induce a current, which is picked up by a sensor connected to the loop. It's possible to infer from that current turning on and off how many vehicles pass by in a given time and how fast they are going.
- Traffic flow cameras: Visual pattern recognition is good enough these days, I think, that computers can look at a traffic camera feed and determine flow and speed.
- License plate tracking for travel times: A camera at point A picks up when a particular license plate passed by, and then a camera at point B picks up the same license plate some minutes later, providing for a calculation of that vehicle's travel time. Do that across all the vehicles that go by, and you have live, real travel time data. There are, of course, civil liberties concerns with recording everyone's license plates, but I think they tend to implement such technologies with automatic dumping of the ID data. The thing about this is that while it gives you the exact units you'd use as an ETA, the measurement lags changes in conditions even more than point flow and point speed do, since you don't get a measurement of travel time from A to B until a vehicle completes that whole trip.
- EZPass tracking for travel times: Same idea as license plate tracking, except that you don't have to do image processing to identify individual vehicles; they identify themselves using RFID tags. Of course, this can only be used on EZPass-enabled toll roads, and the data may be biased by the fact that it only captures EZPass users.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Predicting the future is hard
Date: 2017-07-27 10:00 am (UTC)Predicting the future is hard, and all ETAs are inherently predictions of the future, with a lower bound set by various hard constraints but essentially no upper bound, because you never know what might happen between now and whatever even you're predicting the time of. Disruptions, such as an accident that blocks up the road, are especially troublesome because a) the time to clear the blockage may be dependent on too many exogenous and initially unknown factors to be reliably predictable with any useful precision, while b) that's when users especially want to see a prediction with useful precision, so they can make effective routing decisions. Whether there's a human in the loop or not, taking into account that there's a disruption and applying that knowledge to ETAs is a hard problem.
I would guess that the predicted ETAs on the variable message signs (VMS) are driven by automatic systems and not dependent on human input. Most likely, they are based on real-time measurements of actual speeds, flows, and possibly travel times (possible technologies discussed below). Those measurements are going to vary continuously, and changes in them will tend to lag the more discontinuous events marking the start and end of a disruption. So, an automatic system calculating ETAs based on them, especially if there's some smoothing algorithm to prevent momentary outliers from messing with the final numbers, will produce ETAs that creep up continuously after a disruption starts and then down continuously after a disruption ends, almost definitely more slowly than an educated human observer's guesses might change in light of knowledge of the disruption itself.
Technologies I'm aware of for automatic detection of traffic conditions:
- Induction loop detectors: These have been around for a long time. There's a rectangular loop of metal embedded in the road. Vehicles going over it induce a current, which is picked up by a sensor connected to the loop. It's possible to infer from that current turning on and off how many vehicles pass by in a given time and how fast they are going.
- Traffic flow cameras: Visual pattern recognition is good enough these days, I think, that computers can look at a traffic camera feed and determine flow and speed.
- License plate tracking for travel times: A camera at point A picks up when a particular license plate passed by, and then a camera at point B picks up the same license plate some minutes later, providing for a calculation of that vehicle's travel time. Do that across all the vehicles that go by, and you have live, real travel time data. There are, of course, civil liberties concerns with recording everyone's license plates, but I think they tend to implement such technologies with automatic dumping of the ID data. The thing about this is that while it gives you the exact units you'd use as an ETA, the measurement lags changes in conditions even more than point flow and point speed do, since you don't get a measurement of travel time from A to B until a vehicle completes that whole trip.
- EZPass tracking for travel times: Same idea as license plate tracking, except that you don't have to do image processing to identify individual vehicles; they identify themselves using RFID tags. Of course, this can only be used on EZPass-enabled toll roads, and the data may be biased by the fact that it only captures EZPass users.