For the pedestrian problem at least, it seems even cheaper to just use a button. That also completely bypasses the problem of trying to detect if they intend to cross. Is there really much of a disadvantage? The pedestrian has to be in a specified area beside the crossing already to use it, so likely within arms reach of a pole already. But I can see it being useful for cars and bikes - as a cyclist I've hit a good number of times the magnetic sensors clearly just don't work for a bike, and had to wheel over to the pedestrian crossing to press the button to get the lights to change.
And I expect "person detection" to be much easier with an IR image, there's probably not many places where human body temperature isn't noticeably higher than the background should make it more accurate to pick out. And if you don't need accuracy of temperature reading, an IR camera isn't really any more expensive than a visual camera - it's the same CCD just with a different filter on top. And it completely bypasses any problems of darkness.
I also wouldn't be surprised if the actual camera unit was a near negligible cost of installing something like this, so even if it was multiples of the price, optimizing for that may be the wrong target if it affects accuracy at all.
Granted it's probably nothing too big in absolute terms, but does feel a little unfair, because vehicles usually get a few seconds head start for free due to being able to call a signal as soon as they reach the location of the green extension detector, which can be a few tens of metres in advance of the stop line.
(On the other hand yes, I do acknowledge the because "turning lanes" don't really work for pedestrian traffic, it's often hard to detect pedestrians much in advance because you'll have no idea which way they'll want to turn at the intersection. But it still feels a little unfair :-) )
And unless the default state of an intersection is "red for all directions", switching the lights will always take a few seconds even in the best of cases. By giving the traffic lights controller some advance warning of somebody approaching, it's possible to switch the traffic lights just in time so that that user doesn't have to stop (if traffic conditions allow it of course – with heavy traffic it's unavoidable that somebody will have to wait eventually).
With semi-flexible traffic light designs (where phase length and whether a phase is called at all depends on traffic conditions, but otherwise they still always cycle through the phases in a fixed sequence), there are also situations where a signal can only turn green (if called) until a certain point in the cycle no later than X (in which case it'll remain green until let's say for example X + 10 seconds). If you miss that time X, you'll have to wait a whole signal cycle until the signal can turn green again. If it just happens that you're reaching the intersection at a time X + 5 seconds, it means that if you could have called the signal while still six seconds away (i.e. at X - 1 second), you still could have crossed successfully (because in that case the signal would have been green between X and X + 10), whereas without that possibility you can only press the button at X + 5 seconds and now have to wait one whole signal cycle.
And I expect "person detection" to be much easier with an IR image, there's probably not many places where human body temperature isn't noticeably higher than the background should make it more accurate to pick out. And if you don't need accuracy of temperature reading, an IR camera isn't really any more expensive than a visual camera - it's the same CCD just with a different filter on top. And it completely bypasses any problems of darkness.
I also wouldn't be surprised if the actual camera unit was a near negligible cost of installing something like this, so even if it was multiples of the price, optimizing for that may be the wrong target if it affects accuracy at all.