Mayor Olivia Chow's Toronto

[zang]

Cell service works on other subway systems just fine.
Nobody should be pulling out a knife on a train. Security cameras systems can use ML/AI for object detection and there are models available for weapons (knives, guns). This isn't impossible.

I think you're unaware of the current capabilities of ML at the moment and the cost and computing power required for what you're expecting. We're nowhere near that level yet. Now, if you have a few spare billion you could offer up…

Additionally, "automatic" alarms triggered by that ML network would send an insane amount of false negatives. Every glancing angle on a shiny phone is likely to trigger an alarm, as well as any doofus who decides to bring their own cutlery and eat breakfast on the train (something I've seen a surprising number of times on the TTC). Such false alarms would inevitably mean much more train stoppage and a huge slowdown of the system.

It's a noble idea, but we're not anywhere close. Give it 20 years.

 

[000]

The cameras would transmit over the same data network, if you have cell/WiFi service you can broadcast camera feeds and have those images processed remotely. There could always be a human on standby to verify those alerts.

 

[ShonTron]

I put an end to the latest thread derailment. If you’re promoting white supremacists here, it’s not going to go well for you.

 

[smably]

Security cameras systems can use ML/AI for object detection and there are models available for weapons (knives, guns). This isn't impossible.

Which transit systems are currently doing this?

 

[zang]

The cameras transmit would data over the same data network, if you have cell/WiFi service you can broadcast camera feeds and have those images processed remotely. There could always be a human on standby to verify those alerts.

Okay, so current systems that use AI for weapons detection are hardwired. They require a lot of high-resolution video static camera data over high-speed ethernet connections. As well, they're usually used at the entrances of buildings for instance, where you have a small, limited flow of people.

Installing that kind of system in the TTC would be the exact opposite of that. If you figure there are what? Probably 24 cameras (four per car) on any given 6-car train right now? I'm likely wrong and that number is higher, but I'll err on the side of caution. Several of those are globe–non-static, rotating cameras. You'd need to replace those with static cameras. Probably at least four for each globe. So let's just say a conservative minimum of 50 cameras on a train. (Theoretically, you could replace those globes with 360° cameras, but then you'd need a whole lot of lens distortion correction happening before you process an image; or have an entirely separate ML network to deal with that technical challenge.)

During rush hour, I believe the number was up to 80 trains running at a time.

50 cameras per train x 80 = 400 cameras.

400 cameras sending consistent wireless data at the rate of 8000kbps (the recommended HEVC/H.265 bitrate for 1080p) = 3.2gbps of consistent wireless video with low-latency from 80 different moving trains. Okay, so that may not seem like that much, but remember it's not in bursts, that's consistently running, and that's using conservative numbers. It would saturate whatever wireless service exists down there. The TTC would likely have to build out their own anyway, so they're not reliant on 3rd-party cell service (imagine last year's Rogers outage happening again).

But even then, those current AI weapons-detection systems don't deal with nearly a million people a day. They maybe deal with up to a couple of thousand people from a few static cameras within a building streaming in at reasonably manageable numbers. They aren't meant to deal with humans in close proximity, sudden shifts in body movements, or up to a thousand people suddenly entering/exiting an area (train) at a time—which can happen at busy stations like Bloor-Yonge.

And while current AI monitoring systems claim "99%" accuracy, that's again with static cameras, low numbers of humans, etc. But even using 99% accuracy, what's the likelihood that 1 in 100 people on a given train at any given second are holding an item that might be detected as weapon? I'd say it's pretty high. Multiply that by 80 trains holding 1200-1500 people each (see; rush hour) and we're very quickly dealing with a huge number of false alarms that need to be checked by a human viewer at any given second. Then what happens when those false alarms get marked as such? Do you need to mark that rider (otherwise, they may continue triggering false alarms)? In which case, you now need a system capable of recognizing and following the individual themselves, and not just what an adversarial network has deemed to be a weapon.

Then presumably the cameras would not be limited to trains only, and you include the stations, which could at least be hard-wired. But it's still a lot of data to process. The size of the data centre required would currently bankrupt the TTC. Then you need to figure out where does that data centre go, where do the human "false alarm gates" go? where does the now-massive operating cost for such a system come from? Where does the massive capital cost for such a system come from?

In short, this is just not currently possible at the TTC. And even if you were to pour billions into it, it's not to the level of consistency and reliability that would be required.

 

[afransen]

I think you're unaware of the current capabilities of ML at the moment and the cost and computing power required for what you're expecting. We're nowhere near that level yet. Now, if you have a few spare billion you could offer up…

Additionally, "automatic" alarms triggered by that ML network would send an insane amount of false negatives. Every glancing angle on a shiny phone is likely to trigger an alarm, as well as any doofus who decides to bring their own cutlery and eat breakfast on the train (something I've seen a surprising number of times on the TTC). Such false alarms would inevitably mean much more train stoppage and a huge slowdown of the system.

It's a noble idea, but we're not anywhere close. Give it 20 years.

20 years is an absurdly long timeline. I'd actually argue that such AI security monitoring is not all that difficult and widely used in industry. It may not be perfectly reliable, so most deployments would flag it for a human operator to review. This is not at all difficult. AI monitoring works much better because humans are terrible at maintaining attention on video feeds of nothing happening

 

[zang]

20 years is an absurdly long timeline. I'd actually argue that such AI security monitoring is not all that difficult and widely used in industry. It may not be perfectly reliable, so most deployments would flag it for a human operator to review. This is not at all difficult. AI monitoring works much better because humans are terrible at maintaining attention on video feeds of nothing happening

20 years may have been a bit facetious on my part, but see my above post about the requirements needed for running such a system at the TTC. It's nowhere close to being a near-term solution.

EDIT: And the above is strictly for the Subways. Add busses, streetcars, and their stops and stations into the mix and we're dealing with probably the largest unified surveillance system in the world that would require absolute bleeding edge technology. Just for sheer cost-effectiveness, it's at a minimum of 10 years away. That's if the TTC weren't already cash-starved. Any budgetary increase to cover the costs of such a system would still be starving the system of the needed funding it's lost over the years, while also increasing overall TTC operating costs by who knows how much.

 

[zang]

I mean really, technology is great and all, but we could just try to go back to having guardsmen on the train and more people in the stations themselves. Eyes are the biggest deterrent to crime. Aside from being cheaper, it's also jobs for humans.

 

[turini2]

I am not so sure you are right. I read the following to mean a Speaker ("Head of Council" ) can be appointed 'by the City with the Mayor's approval. The City of Toronto Act reads:

Role of the mayor as head of council​

133 (1) It is the role of the mayor of the City, as the head of council,

(a) to act as chief executive officer of the City;

(b) to preside over meetings of council so that its business can be carried out efficiently and effectively;

(c) to provide leadership to council;

(d) to represent the City at official functions; and

(e) to carry out the duties of the head of council under this or any other Act. 2006, c. 11, Sched. A, s. 133 (1).

Same​

(2) Without limiting clause (1) (c), the mayor’s role includes providing information and making recommendations to council with respect to council’s role under clauses 131 (d) and (e). 2006, c. 11, Sched. A, s. 133 (2).

Substitution​

(3) The City may, with the consent of the head of council, appoint a member of council to act in the place of the head of council on any body, of which the head of council is a member by virtue of being head of council. 2006, c. 11, Sched. A, s. 133 (3).

I live in hope of day 1 actions of a) bye bye Rick Leary b) hello new Speaker of Council

 

[Islington Express Bus]

Nah if they remove Nunziata as speaker then we gotta hear her opinions.

 

[zang]

I live in hope of day 1 actions of a) bye bye Rick Leary b) hello new Speaker of Council

I'm reminded that the infamous Neil Flagg was trying to use whatever clout he felt he had a few years back to get all the city councillors turfed *except* Frances Nunziata. It was titled something like "43 must go (Frances can stay)".

 

[zang]

Nah if they remove Nunziata as speaker then we gotta hear her opinions.

It won't be for long. She only won her ward last fall by 0.4% difference over Chiara Padovani (who herself was up over 25% from the previous election). Frances' slid over the years; this is likely to be her last run at it before she's left packing for greener pastures.

 

[afransen]

20 years may have been a bit facetious on my part, but see my above post about the requirements needed for running such a system at the TTC. It's nowhere close to being a near-term solution.

EDIT: And the above is strictly for the Subways. Add busses, streetcars, and their stops and stations into the mix and we're dealing with probably the largest unified surveillance system in the world that would require absolute bleeding edge technology. Just for sheer cost-effectiveness, it's at a minimum of 10 years away. That's if the TTC weren't already cash-starved. Any budgetary increase to cover the costs of such a system would still be starving the system of the needed funding it's lost over the years, while also increasing overall TTC operating costs by who knows how much.

The camera feeds on the subway are going somewhere currently, surely. The processing doesn't need to be done all centrally either, you could have edge compute on the vehicles to analyze the video inputs. False positives are a matter of tuning sensitivity. Briefly mistaking a shiny phone for a knife is not likely to be a problem. I think you would be looking more for patterns of aggressive behaviour or passenger panic as a trigger for security attention.

I mean really, technology is great and all, but we could just try to go back to having guardsmen on the train and more people in the stations themselves. Eyes are the biggest deterrent to crime. Aside from being cheaper, it's also jobs for humans.

I think the technology would work well hand-in-hand with additional security/police presence. I think part of this is not merely dealing with the aftermath of violent events, but to deal with lower level misbehaviour to reduce the feeling of impunity trouble-makers have currently. There is a great deal of public good to be created in returning a sense of personal security when using transit. In due course, it may actually save money once the norms are re-established that misbehaviour on transit results in consequences.

 

[Videodrome]

Sorry for the off topic post, but I was just at Taste Of Lawrence and saw John Tory there. Doug Ford was apparently also there earlier. I guess Chow will show up next year.

 

[zang]

The camera feeds on the subway are going somewhere currently, surely. The processing doesn't need to be done all centrally either, you could have edge compute on the vehicles to analyze the video inputs. False positives are a matter of tuning sensitivity. Briefly mistaking a shiny phone for a knife is not likely to be a problem. I think you would be looking more for patterns of aggressive behaviour or passenger panic as a trigger for security attention.

Currently the cameras are accessible to the drivers cabins at either end of the train, and recordings are collected when needed (https://www.ttc.ca/en/transparency-and-accountability/policies/video-recording-policy). It does not appear there's any live transmission off the train.

False positives in any kind of AI system are a problem, and tuning can only do so much. It’s the sheer quantity of data, the number of people and their proximity that needs to be analyzed at any given time that is the main problem with putting them on a train. AI is good, but not great at detecting a single human form (see; Tesla/pedestrian collisions for how well current systems actually work - https://arstechnica.com/cars/2023/0...ing-sees-pedestrian-chooses-not-to-slow-down/). It’s absolutely terrible at knowing what a mass of human beings in close proximity actually is. A packed train full of 1200 people is either going to generate a steady number of anomalous positives or you’re going to have to tune so many out that you’re going to miss real problems. And that’s a much bigger issue.

Even if you could determine intent by physical action via AI (you really can’t right now), there’s the issue of “what is aggressive behaviour”? Rowdy drunkards, sports fans, playing kids would likely trigger systems. Even weapons detection itself (which is most of what AI is capable of at the moment) is gonna go berserk on Toronto Comicon or FanExpo weekends.

And what I haven’t brought up is that current weapons-detection AI isn’t living up to the claims of developers. Most of it isn't 100% camera based, and a lot of it involves passing through body scanners. (https://theintercept.com/2023/05/07/ai-gun-weapons-detection-schools-evolv/).

Outside of detection, what do you do when something real is happening? The driver can’t leave the cabin. So do you just stop the train at the next station and wait? Do you open the doors allowing a perpetrator to escape (which is generally what happens now), or do you keep them closed and allow more stuff to happen? Detecting violent acts isn't going to stop them from happening. Someone pulling out a knife and stabbing someone will always be faster than AI passed onto a human for verification and the time taken before potential intervention. It will almost always be reactionary and then, what purpose has an AI system served?.

We have passenger alarms that when pressed alert the driver's cabs on which car the alarm was pressed. They were (apparently) pressed during the recent stabbing. This is why having a guardsman on the train is a better overall idea for right now. We shouldn’t have gotten rid of them anyway. Workers and experts said quite loudly that getting rid of them would compromise safety, and look where we are now. https://www.blogto.com/city/2021/05/ttc-planning-eliminate-subway-guards-and-people-arent-happy/

I think the technology would work well hand-in-hand with additional security/police presence.

I agree the technology would work well, but when it's up to snuff and cost effective. Which is not now.

If the city is going to spend vast gobs of money, it should be addressing the root causes of things like this and not reacting to the symptoms.

I think part of this is not merely dealing with the aftermath of violent events, but to deal with lower level misbehaviour to reduce the feeling of impunity trouble-makers have currently.

A lot of this problem is stemming from just a lack of people within the subway system who were there even 5 years ago. The automation promised with the Toronto Rocket trains and Presto card entry became a way for bean counters to assume that humans were no longer as necessary on the system. The TTC ditched the 2nd person on the train (guardsman), and severely cut back on booth staff.

It used to be common to see staff all over the place in a station, now you're lucky if you see one on your way to the train. As I've mentioned elsewhere, I've been to Bloor-Yonge station (the busiest in the network) and seen the only booth they use now, empty many, many times.

I remember days not so long ago with both booths at that main entrance were always manned. TTC cuts have (like just about every industry) come primarily from labour.

Automation shouldn't be a way to get rid of the humans. By comparison, Tokyo has five lines of fully-automated trains and electronic entry, but even aside from the passenger pushers, there are still gobs of transit workers on platforms, on the train itself and in stations. Most of them are customer service, but a human presence is always felt.