Avoiding the ludicrousness of actually doing that beyond an acceptable interruption to road traffic, like queuing theory and optimals… a pedestrian or cyclist prepared to wait 1 or 3 minutes for lights…
does not take an a.i. to recognise a bike or a cat or a wheelchair, those should be quite well defined algorithms… it is ludicrous to leave that shit to a black-box A.I. you want the A.I. to give you the code of the algorithm… or as the programmer of the A.I. you can reverse engineer the algorithm from the A.I. parameters/states. Thus negating need for A.I. to actually perform the production use of the algorithm… that is just code. Like you want to detect cats, you run the cat-detect code… not some black box a.i. that theoretically id’s a cat all the time… so you may go order of avoidance heirarchy for your car… hedgehog, human, cat, badger, child-in-pram… and actually that ordered list of absolutely required detection’s should be defined… human.. child-in-pram, hedgehog, other-vehicle etc. how many layers can you realistically do per frame with cpu power you have eh! 5g eh! theoretically you ship your unknown a.i. algorithm into the unknown “cloud” at the advertised speeds of 5g eh! LMFTO!
Actually can probably detect a bike quite simply in that fixed scenario by general object shape and the fact they travel faster than a pedestrian, depending on resolution the actual contact area of the object to the ground also a tell, cycle only contacts the ground by two thin things…. wonder if it can currently detect a recumbant cycle? Side recognition probably easier, looking for two round things one behind the other within a small range of distances… and a blob in the middle raised off the ground. The fixed position will make it easier because you can get images of the background without any objects….
BUT! as the article below alludes to, there is a whole other layer, an issue raised by one commentator is, “Is it actually safe to cross?” the light may be red, but did a driver see the light change in time to stop. Plus they reckon the cyclist cannot see if the light is actually green as it faces away from them. Like the missing manual winders on car windows nowadays, they probably did away with the old press to cross control panel that gave colour and audio feedback to the user, all that has just gone, no feedback to user, and you are supposed to blindly trust the a.i. eh! Nope I was wrong, they still have the old boxes, but it is them they have mounted at a stupid angle, you can see it in the picture below, they need to be rotated -45 to -90 degrees in the Z plane so they can bee seen better.
Just occurred to me the old push-to-cross crossing have a built in safety mechanism that will work with most drivers, they can see someone standing waiting to cross when the lights change so as good road-users they will stop.
plus in this case probably better to explore the algorithm for better batching people up to cross the road, involving average statistics of crossings at times of day. You could set a published schedule even. I think the thing with cyclists is there is the empty road at the point of crossing scenario where you would not even need to press the button to cross. But by actually triggering a traffic light change anyway you may cause an interruption of a vehicle/s nearish to that empty crossing window.
Possibly where our adoption of A.I. is failing… A.I. dictates the result to us and we do not understand the algorithm… sounds like a cunning plan.. call it democracy and bound to succeed eh!
