ALPR can't keep up with video

[boarder2k7]

I loaded Sighthound ALPR in a docker on Ubuntu to run a single camera evaluation before deciding on what direction to take.

The interface is successfully set up, but I constantly get messages like "Error: An error occurred trying to access the video or stream" when clicking on any of the recent vehicle detections. This is occurring even with the detection stream paused, so it shouldn't be processor or drive related.

I'm running this single camera on an i7-2670QM with 8 GB memory. When turned on it pegs all 4 cores / 8 threads to 100% and immediately starts falling behind on processing.

Detections also seem wildly inaccurate, as in the 5 minutes I had it turned on last night, it detected 40 cars on an empty road with some leaf motion in the side of the frame. I ran it for a couple minutes this morning as well to see if I would have better luck in daylight, and it has detected a parked car 5 times as separate detections within just a couple minutes before I switched it back off due to it falling woefully behind on processing.

Any thoughts on what I should do to proceed here?

[mr.Magoo]

I would be interesting to know the minimum spec to be able to keep up with a single stream.

Mine did the same and I ran it on Windows direct (no docker) and it kept lagging behind.
If I compare my rig v.s. sending the same image through the API, I get a response from the API in ~1-2 seconds whereas ALPR takes about 8-10.

Similar issues with parked cars which is kind of odd, almost as if it's looking for any type of motion (or potentially the same "car" algorithm as they're using in SH Video, but at that point when it triggers they analyze the whole frame and not just the section where the motion occurred.

[boarder2k7]

Yeah, I assumed that it would be more than plenty for one 4 MP camera to run on it's own system with a quad core i7 and 8 GB memory, but it isn't for some reason.

I have a Camect system on site as well. The box Camect sells will do 24 MP of full AI recognition, and it's only running a Celeron J3455 with 4 GB of memory. They're clearly doing a significantly better job at optimization somehow. According to UserBenchmark the processor I'm using is 83% faster and apparently handles less than 20% the camera load when comparing Camect and Sighthound ALPR.

Sighthound seems to be behaving a bit better now (for no apparent reason I'll add). It's now cycling the processor rather than pegging to 100% all the time, and seems to be holding at about 30 seconds behind. Still lots of false detections though. It thinks a stroller is a motorcycle, that leaves are cars, and that people from Finland are driving around Connecticut. It's also fond of detecting the same car multiple times as it drives by, yielding different plates each time while touting a 99% confidence in the detection.

Clicking on an alert also gives a "An error occurred trying to access the video or stream" message before crashing the web page. If the stream is paused first, the error still appears, but doesn't crash the page, so it is able to be closed to allow viewing of a still image, though the video clip will never play.

I want to like this and I want it to work, but boy is it frustrating me.

[mr.Magoo]

As much as I wish I had a Ferrari in my driveway it's a bit comical that ALPR repeatedly mistakes our portico for a high end sports car.
(usually Ferrari or Audi R8)

I mean, it doesn't get more stationary than this !?

Which is why, just like with SightHound Video, I'd love some form of calibration.
A car is roughly this size and facing THAT way (horizontal rather than vertical).
Black-out areas that are completely ignored would be great too... there's no need to analyze my house for number plates
This should also help reduce processing power a lot.