A/B Street: A simulation game to fix Seattle's traffic
abstreet.org
abstreet.org
Happy to answer questions! And if you're serious about running this in your city, the importing process has improved greatly since December, so get in touch.
Does this only simulate what would be considered a typical traffic load, or can it simulate outlier events (sport games, concerts, COVID quarantine, etc)?
Thanks for reporting! I'll try fixing this blind (no Windows machine handy) and send a new build; it'd be great if someone could confirm it.
I believe MS provides premade free virtualbox images for IE testing purposes too.
Now that I'm looking more closely, this looks like some sort of initial aspect ratio issue, possibly related to a resize.
The game launches and the window either starts maximized or it maximizes itself. If I restore (un-maximize) and re-maximize, the aspect ratio is clearly different because the logo in the center of the screen gets slightly taller (while staying the same width).
In other words, the initial aspect ratio of the graphics seems incorrect (squished vertically slightly for me, running at 1920x1080).
Do you have your windows taskbar in non-default position? Mine is on top of screen and some games in some display modes give offset mouse position
I tried the Windows version and some of the UI controls don't respond to mouse clicks, e.g. home and gear button in bottom left, speed controls and dropdown. I noticed the target areas for them are about 90px below where the buttons are painted. So when I click the checkbox to show/hide Bus, it actually turns on and off the one for Car. I'm on v0.2.0b and the client area of my game window (excluding titlebar, edges) is about 2560x1520.
Some elements, like the dropdowns at the top (Sandbox map and traffic setting) work ok.
EDIT: Looks like schemescape beat me to reporting this. I'd love to know if this was just a dumb "programmer-oops" or if there's an opportunity for the language/libraries/framework to better funnel code toward correctness.
2) How well does it validate? I couldn't find any reference to observed traffic count data anywhere
2) I wrote a bit here: https://old.reddit.com/r/urbanplanning/comments/hdylmo/ab_st... In short, validation is hard, and I'd say I'm a bit understaffed.
He demoed his actor system with 100 000 cars in the middle of this talk: https://www.youtube.com/watch?v=qr9GTTST_Dk
Maybe you guys should team up :) (if not already...)
It completely ruined my fancy subway system's timetable. :(
https://www.aicrowd.com/challenges/neurips-2020-flatland-cha...
(disclaimer: I'm an organizer)
Ideally, a government dashboard to see traffic congestion changes per year, month, day of week, but also hour. If we can create metrics for cities like p99 door-door trip time (i.e. latency) policy makers can better measure changes, bottlenecks, etc and plan better.
https://www.digitaltrends.com/cars/waze-for-cities-data-puts...
The more infrastructure that is built downtown, the more people want to be downtown.
Chocolate is not free at the point of use but roads usually are. Without price signals or other limits people over-consume resulting in a shortage and people paying in time.
I walked to work (when I lived in Seattle recently) but the amount of traffic jams I saw caused by both regular people in their cars, and the mass transit bus drivers, driving right into an intersection as the light turned red is crazy. I don't know how you can plan traffic flow around it, because as green as the light might be, no one is moving with a bus sitting in the middle of the intersection.
If any company tries to limit what I do in my free time, it's a quick and hard "no" on working for them.
So the question is, given a bit of ML/etc driving it, and some actual commute time data, what are the changes something like this actually can be used to improve traffic?
I actually got started in traffic sim while I was in Austin! If you know anybody there who'd be interested in putting in some work to get the area running smoothly, I'd love to include it as a second city in the game.
I can't wait till tonight when I'll have enough time to get it installed and try it out!
Big kudos to the author for this nice project ! I hope it gives further ideas to people playing it
Thanks for pointing out that it’s in rust.
Biggest barrier is lack of trip demand data. If you come up with how many people leave from building1 to building2 at time T using some mode of transport, then this can work.
I no longer work there either, but feel free to reach out if you're interested in chatting.
Hopefully I can have the poc going in the next month. A city wide integration with something like AB street would be really cool and was phase 2 for this project. Phase 3 being an optimised citywide traffic control.
In between these steps, (all based in python) opencv takes the video, splits it, sticks some of the frames on a kafka cluster.
Another node picks up approx 2 fps, hands it to a object detection algo based on tensorflow which then puts the results on another kafka topic.
A heavily customised tracking algo pulls all of the frames available and the tensorflow results for initialisation and reverification. Once a vehicle is tracked through the field of view the vehicle is counted.
For the privacy issues, as this is only tracking for a short amount of time, not tracking between intersections and not tracking persistent uuids or numberplates, there shouldn't be too much trouble. But it will depend on the local law requirements. As this is a statistical problem, tracking the actual path of a car between point a and b isn't completely necessary if you can get a sufficient amount of coverage over a city. (You probably already understand this.)
As for opening the datasets up, I plan to do this with the poc. I've been using the Biloxi Mississippi camera live stream from youtube for all of my testing so far. My poc will be a website with this live stream having an overlayed view and the ability to download the stats for the day/month/year etc. As for other cities, I'm sure a FOIA will be able to get you the data. But I will be pushing to publish it openly as they will have to pay for it either way.
As you can probably tell, I'm trying to turn this into a business and the costs of actually collecting and processing the data isn't something that is reasonable for single person to bear. After talking to the traffic dept for my country, the largest city of 1.5m has about 3000 cameras. But I am aiming for a system that's cheap enough to enable 24/7 monitoring as that's when this all gets pretty exciting.
Is it possible to create a map with one click for a new city?
I read through https://github.com/dabreegster/abstreet/blob/master/docs/art... but I can't tell if this must be manually followed.
Not quite one-click, but you can try the one-shot importer. You'll probably hit some issues; I have strict assertions to prevent regressions in Seattle. Please file a bug if you can't get a new place to work.
I assume that there are software systems that take in schedules and various inputs (cameras, magnets, crosswalk buttons) but I'm curious if anyone works in the field and would talk about their experience.
https://www.kimley-horn.com/service/kits-advanced-traffic-ma...
A user manual might give you some sense of the complications (bus priority, special events scheduling, etc.):
https://dpw.lacounty.gov/TNL/ITS/KITS/site_files/LACO%20KITS...
For example, autonomous vehicles would be interesting. You potentially don't need intersection control with all autonomous traffic. Or what is the benefit of certain streets or lanes being designated autonomous only?
And what is the impact of education? Is there a way to change behavior that's cheaper than changing a street for example? You can change the street, but you can also change how people use the street.
I thought this myth had been widely debunked. In a world where every commuter drives a 100% autonomous car, you still have hobbyists driving old vintage cars, cyclists riding non-autonomous bikes, pedestrians, birds, runaway shopping carts, etc. disrupting the flow of traffic. Without traffic control you make the intersection in-crossible for a significant subset of the traffic.
Lots of traffic lights already have cameras on them. Hook 'em up to an AI and a goal of maximizing throughput.
Keep an eye out on the hn home page I'll be posting my poc in the next month
Besides, if you have an intersection that deals with 500 cars and 2 bikes, what would make the most sense to optimize for?
First, the vast majority of signals aren't networked, they're on simple timers. And you can't just put them on the internet for hopefully obvious reasons.
Second, there's a bit of a butterfly effect. Change one thing and the whole system reacts in sometimes unpredictable ways.
Third, you need to account for pedestrians too. There's ADA requirements around how long signals are green for based on the width of the road.
Fourth, you need to account for pedestrian load. There's no cameras trained on the areas where pedestrians gather. If you're going to say they just need to press the button, no:
https://cal.streetsblog.org/2020/04/01/stop-touching-pedestr...
Basically, it's really not that simple.
> First, the vast majority of signals aren't networked, they're on simple timers.
Doesn't matter. Even doing the ones that are networked would be a big improvement.
> And you can't just put them on the internet for hopefully obvious reasons.
But you can put every neighboring building on the internet with their security cameras? It's not like you need hires or high frame rates to detect large moving blobs, count them, and estimate their speed & distance.
The problem is, the city has no incentive to do this, as the cost falls on other people.
To the first point, the networked lights are already in this mode (it's the reason why they added it in the first place). I can't find the report right now but upgrading all of the signals on 2nd from Denny to Yesler cost something like 30 million bucks.
The upgrade in the Mercer project area isn't exactly a glowing success:
https://www.theurbanist.org/2018/09/06/eleven-ways-adaptive-...
I didn't consider it because the solution is obvious. An interlock that makes it electrically impossible to have 2 greens, like an airlock has a piece of hardware that prevents both doors opening at the same time. I suppose I should also mention the obvious that if the internet connection goes down, or the camera lens gets iced up, etc., it should automatically revert to egg-timer mode.
See https://en.wikipedia.org/wiki/Therac-25
> I can't find the report right now but upgrading all of the signals on 2nd from Denny to Yesler cost something like 30 million bucks.
Somehow the money is always found for internet connected cameras to spy on us.
> The upgrade in the Mercer project area isn't exactly a glowing success:
These are solvable problems. Except point 11, which is ironically complaining about it working successfully at what it was supposed to do.
> These buttons have long been decried and criticized by advocates for walking, anyway. The buttons’ purpose is less to keep people safe than to reinforce the primacy of cars on the street by forcing people who want to cross a street to “beg” for a walk signal.
This makes it sound like buttons are some kind of implement of the bourgeoise, used to "reinforce the primacy of cars" and keep the pedestrians powerless. In reality, whether pedestrians like it or not, cars are primary. The roads literally exist for cars. As a pedestrian, I would do the very difficult task of pressing a button, if it meant that as a driver I could have less traffic and hit less red lights.
> Typically, a person expects the signal to show walk, then flashing don’t walk, then solid don’t walk. Signals follow a familiar pattern and people embed that pattern in their muscle memory. Adaptive signals, however, can flip from walk to don’t walk to walk again
That simply doesn't happen, adaptive or not.
https://en.wikipedia.org/wiki/Sydney_Coordinated_Adaptive_Tr...
We have some trams, mostly fairly new routes, but hardly any share the roads with traffic. There used to be a lot that did, until the 1960s I believe, but they got rid of them in favour of the supremacy of cars. We have good buses and trains (and ferries) here though.
Off hand, I'd say lights are often timed such that once you get one green you get a lot of greens, then you're asking for more stop and go traffic simply because someone showed up and threw off the sequence.
Take away that and you gain some benefits, but you also lose in other areas.
Not an easy problem.
That's what AI is good at. Besides, isn't a little incongruous to argue that the optimal solution for a complex optimization problem is a 1950's egg timer?
https://www.researchgate.net/publication/281356849_On_Traffi... where you have nice (or almost) nice regular blocks with a constant speed. When you start deviating from this, the math all breaks down.
The ideal case would be to pulse the traffic such that given a certain amount of traffic flow and everyone following the proper speed limit that the time it takes for the traffic pulse to get to the next traffic light is green (and the pulse can continue).
This can be seen when you switch from one traffic light synchronization district to another - and hitting a traffic light.
There's a stretch on my commute where when I follow the speed limit (its 25 in this segment), I can go from one end of the traffic district to the next without ever hitting a red light (about 15 blocks).
The thing is that if the traffic isn't pulsed, then the spot where it changes from one traffic district to the next backs up significantly and causes other backups further down traffic.