What's Your View For Suicide Prevention On A Foot Bridge With Analytics

We are looking at deploying thermal cameras in conjunction with an analytics package to detect loitering cross line detection and to set up some rules that would be consistent with somebody thinking about jumping before they jump . The reason for the thermal option is due to the challenging nature of the lighting and we feel that the thermal option will give us less false positives .

Is this in reference to a particular spot, or to create a general product?

Is the jump off point at or nearly at a publicly accessible area?

I think your rate of false positives will be inversely proportional to how difficult it is to get to the spot. Once you get to the spot though, I'm thinking that it will be hard to distinguish a jumper from your ordinary tresspasser.

If you are only interested in the jumpers, I think the false positive rate would be way too high assuming you want a reasonably low false negative rate.

We are looking to create a general rule as there is no one specific spot with all areas accessible to the public .

I see this as your real barrier to implementation:

"set up some rules that would be consistent with somebody thinking about jumping before they jump"

If the 'jump points' are not separate from areas where tourists normally would stop and mill about, how can you implement a 'rule' that would somehow distinguish from tourists taking selfies and those contemplating the Final Leap?

There have been reviews of video from train station platforms of jumpers. They exhibit some common features, some of the time. I recall it being said they would walk to the platform edge and back to the walls for up to 20 minutes. Then again, so do kids just excited to be there.

I would say by the time you detect the "pattern" you will never get there to intercede. It seems like it would have about the same value as Object Left Behind" in that you wouldn't send a team of suicide negotiators because someone is walking towards the edge, just like you don't empty a station when a large fast food bag is left behind.

Physical barriers would do a better job, but unsightly.

I recall it being said they would walk to the platform edge and back to the walls for up to 20 minutes.

Not from the projects I've been involved in.

There is no way to detect "suicidal" behaviors with analytics on the market today.

Some jumpers make a beeline for the edge, they just walk up, jump, and that's it. Others will hang around like a person waiting for a friend, others will climb over a ledge and then hesitate for a bit, or possibly even climb back and forth.

Depending on the bridge or location you also get thrill-seekers walking along the edge, peering over, standing on the railing, etc. These can all be good things to detect, and potentially detectable, but they don't indicate a suicidal person, and only very rarely can you respond fast enough to do any prevention anyway (though the early-warning itself can still be valuable).

I don't think thermal cameras have any particular benefit here, but it would depend on the site itself. These days with analytics you're typically better off (IMO) with a higher-res visible camera than with a thermal camera for analytics.

"There is no way to detect "suicidal" behaviors with analytics on the market today."

Let's go further: "There will never be a way to detect "suicidal" behaviors with analytics, period."


Can a human detect "suicidal" behaviors just by observing?


what, besides actually jumping from the foot bridge, would you suggest is a behavior that can be classified as suicidal?

my point being: I'm not sure that any behavior that one might classify as 'the kind of thing someone who is about to jump to their death does before they jump' will be dis-similar enough from other, expected behaviors in this location to create any kind of 'rule' to base an alert on.

I agree with U 4, analytics will never be able to detect "suicidal" behavior analytics. Someone who is impatient or fidgety could exhibit the same nervous behavior. I don't think analytics can accurately predict a suicidal person's behavior.

You may be right, but you need not put the word 'analytics' in your qualification, IMHO.

'Never' is a long time, and if it's possible for a person to accurately predict by looking at pixels, I'm confident that a machine could model it as well.

If a person can't tell, then how could an analytic be expected to?

Do you agree?

We worked on one of these once. The mission was not to detect behavior, but to detect the body falling for police and medical response. Line crossing was tried and it didn't work.

Would it be considered insensitive to ask how you figured out it didn't work?

In a similar application we've had divers jump to test system performance. Also life-size human "dolls" were used, and in cases with thermal cameras we had to leave them out in the sun for a bit so the ballistic-gel center could warm up.

Part of the difficulty with detection is that the person/object speeds up very quickly. It can be difficult to get a FOV wide enough that the person is present long enough to be detected by the analytics.

Lastly, we did some tests with actual recorded video of jumpers. I can't say where the video was recorded, but most customers that have these problems tend to have some CCTV footage that can be used for test/training purposes.

Its nice to have you back Brian ;)

CPR dummy with a rope attached to haul it back up. The thermal was set up using a "from" zone and a "to" zone. The mass had to be detected in the "from" zone and when it moved to the "to" zone, the alarm would trigger.
Sorry. Went back and checked my records. It was a plywood human cut out, dressed in clothes that were soaked on hot water before it was thrown over the bridge, then hauled back out by a rope up and down the bridge throughout the cameras field of view to the far edge.

Better solution would be to install a suicide hotline phone on the bridge itself.

I see this a lot with analytic requests when end user's expect cameras to solve the problem of lacking physical infrastructure. For example, end user requests an analytic to detect a truck pulling up to a building before it blows up. What difference will the analytic make; it won't prevent the truck from blowing up and you won't have enough time to evacuate the building. Better solution is to physically make it impossible for an unauthorized truck to get near the building.

Same thing goes for bridges and jumpers. Make a better bridge that doesn't allow for jumpers; put up a higher fence, or put a net underneath. Analytics won't prevent anything, just alerts you it's happening.

The Gap at Watsons Bay in Sydney has something similar, thermal and cross line detection as well as help points that go straight to a suicide prevention line or police. You may be able to reach out to them for advice / feedback on the success of the project.


We are currently working on this application and there are some interesting features that can be used to detect if somebody is planning to commit suicide.
I agree on the fact that you can't be 100% sure but there are some things that can be done.

We currently use this system (thermal camera based) on train platforms and alongside train tracks. As you can imagine, these are also a target location for people wanting to commit suicide.


Several years ago, we did a project in phase with an access security project on a bridge. The bridge by itself have several restricted access and nobody usually goes walks there without authorisation. Due to the fact that all the walking path must remain open, the decision was to improve our delay of response by adding analytics detection at the beginning of the bridge. Depending on the situation, the quick response of emergency can really help to avoid the worst case. We used analog thermal camera with analytic encoder. The units can detect walking human or car and send alerts through our vms apps. The loitering detection was not used but was available.

The resultats was and are still pretty accurate.

The only issue we have is when it's rainy/snowy outside the analytics generates false alarms.

I think the suicide hotline phone is a good idea too..

If you need anything else, you can pm me.


Thanks for all the fantastic feedback I will keep you posted on how we get on .