This is pretty cool and a great use case for a DeepLens. Since the DeepLens can do the inference on the machine it would save a ton of bandwidth and you could even make it play a sound when a spot opens up, so it would work disconnected from the internet.
Do you live in London? It has been said, without too much digging on my part, that London has more cameras than anywhere in the world.... (or is closely behind Beijing).
Yes. And yes, it's not news that there are loads of cameras. But the reality of actually seeing them all there on that page like that, for all to see. Knowing that anyone could easily watch my commute tomorrow, that's still a bit weird. Does that make sense?
There are local Facebook groups I'm in where people ask for cam footage after their car parked on the street gets damaged and there always seems to be at least one video offered.
Cool. Can we add infrared detection to determine if the car is just standing rather than parked, so you can eliminate the cases where a car is not in a parking spot, but would appear parked to this detector?
I'm a user of Mask R-CNN. While the masks it generates are great, I find too many false positives. So in applications like this, the alerts you get may not be what you think. Actual available parking spot may not be there.
Have you experimented with having confidence level filters on your detections and further training the model with more data on the types of object you are trying to detect in the environments that they appear? Depending on what you are doing, it might just be a case where some fine-tuning of the model would solve your accuracy problem.
Sorry if this is an obvious suggestion and you've already tried all that.
This is pretty cool and a great use case for a DeepLens. Since the DeepLens can do the inference on the machine it would save a ton of bandwidth and you could even make it play a sound when a spot opens up, so it would work disconnected from the internet.
> With a modern GPU, we should be able to detect objects in high-res videos at several frames a second.
How much power does this thing use?
And which modern GPU... my 1050ti struggles with out of memory issues and pytorch.
Well a 1050 is pretty underpowered for real-time object detection...
1080ti work?
For training maybe. Inference uses far less memory.
Add all the public traffic cameras from your neighborhood to get more data.
I want to monitor the traffic cameras on my commuter bus route to see if I can better estimates than google maps.
Now I'm interested in using the public cameras to track myself around the city.
which publically available cameras are there? which city makes traffic cameras publically available?
edit: https://webcams.nyctmc.org/multiview2.php
holy crap
Check out the _video_ feed Seattle makes available.
https://web6.seattle.gov/travelers/
For Seattle and environs: https://www.seattle.gov/trafficcams/links.htm
Holy crap indeed!
I wonder if London's network is public.
With some delay, indeed it is.
https://www.tfljamcams.net/
Wow! And the delay is only a few minutes. This is awesome and also scary.
Do you live in London? It has been said, without too much digging on my part, that London has more cameras than anywhere in the world.... (or is closely behind Beijing).
Yes. And yes, it's not news that there are loads of cameras. But the reality of actually seeing them all there on that page like that, for all to see. Knowing that anyone could easily watch my commute tomorrow, that's still a bit weird. Does that make sense?
Sure. I mean, this is what privacy advocates bang on about all the time.
I'd guess that page is probably <5% of CCTV being recorded in London.
Hell, my car has an always-on camera.
There are local Facebook groups I'm in where people ask for cam footage after their car parked on the street gets damaged and there always seems to be at least one video offered.
Found Finch from Person of Interest :)
Cool. Can we add infrared detection to determine if the car is just standing rather than parked, so you can eliminate the cases where a car is not in a parking spot, but would appear parked to this detector?
But if stationary cars are considered as parked, what about cars that are momentarilly stopped due to a red traffic sign?
Wouldn't it cause false positives?
I'm a user of Mask R-CNN. While the masks it generates are great, I find too many false positives. So in applications like this, the alerts you get may not be what you think. Actual available parking spot may not be there.
Have you experimented with having confidence level filters on your detections and further training the model with more data on the types of object you are trying to detect in the environments that they appear? Depending on what you are doing, it might just be a case where some fine-tuning of the model would solve your accuracy problem.
Sorry if this is an obvious suggestion and you've already tried all that.
I'm surprised stores that offer time-limited free parking don't utilize something like this to identify/tag cars that park past the limit.
In the UK this is fairly common using number plate recognition.
On entry and exit your plates are recorded by a camera.