Got a simple TAPO C310 camera looking out my kitchen window.
If a car stops, I usually can capture manually registration number.
But since very sideways there is a lot of frame prediction stuff going on, and a single frame seldom can tell the full number.
It's 500-1000 kbps bitrate so not much to go on.
What kind of tools are there?
- ideal would be some object tracking thing like in video editors
- manually frame a number plate, and can capture this on a range of frames
Things I search on internet is gaussian blur of image and other stuff to enhance image for this purpose.
This was some special tool made in perl or similar, maybe too much of hacker level.
If it takes hours to do does not matter that much, very rare I need to do this.
Any tips on procedure or tools is much appreciated.
+ Reply to Thread
Results 1 to 9 of 9
Last edited by larioso; 5th Aug 2022 at 10:22.
Thanks. On the other hand if I can read it no reason to analyze it either.
Here is a system for automatic analysis
Probably meant for automatic charging systems or similar.
- identify car
- identify plate
- enhance plate
- use ocr software
Way higher demands than I have regarding automation of the process.
But a graphic editor is probably good thing with all sorts of processing.
The best way from a couple of frames maybe to join them on a specified object that is tracked, is my assumption might be a toolbox for.
HDR editors has some way to identify and align multiple frames/pics so they can be enhanced from exposure point of view.
But follow an object is probably way more complicated.
Just thought it would be good to practice and see what is possible, or get a better camera.
Note how the license plate in their sample image is already legible. The article isn't about clearing up an unreadable license plate but about locating it and using OCR to read it.
Yes, tracking and stacking can sometimes work. But with highly compressed video the tracked license plate (or portions thereof) may be identical over several frames because that's how high compression video codecs work (see motion prediction in video codecs). It's not until the geometry has changed that a new image is available. By then simply stacking won't work. You will also need some kind of scaling, rotation, etc.
Might be this camera is way too simple for the task.
It's even 15 fps, and rather slow shutter speed it seems.
So predictive frames from previous frame change too much.
At least have a camera to set shutter speed to each frame is sharper.
Sideways movement is difficult.
Thanks for your input.
If you COULD determine ahead of time the area where the camera could expect to see a resting license place, and the camera had decent optical zoom capability, the cam could be placed and zoomed to stay at that setting and then it would have a much better chance of capturing something legible (also have to assume sufficient illumination/contrast).
But if that is a run of the mill security cam, it is very unlikely to fit the bill.
The dirty little secret is that the popular security cameras (Nest, Ring, Arlo, etc.) are often quite useless for both license plates and faces. Their lenses are cheap, the advertised resolutions are usually bogus (they upres from a cheap sensor), they do a terrible job at night, and they don't have sophisticated light algorithms for, as an example, getting the license plate from a car driving towards the camera at night, where the headlights screw up the exposure and make the plate unreadable.
Years ago I was asked to purchase and install a camera at the entrance to our non-gated, 206 home development. I ordered three cameras that seemed promising, including one from Samsung, but sent them all back. The Samsung couldn't even resolve the big text on the box it came in at a distance of only six feet!
However, you can get really good cameras, but they cost a LOT more than the ones I've described so far. The nearby town of Carmel installed cameras and has made quite a few arrests, based on reading license plates. A nearby gated community (with lots more money than we've got to spend) contacted Carmel and ordered the same equipment. They too have had great success. I got them to send me the complete bill of sale. They bought four cameras from Grid Surveillance in San Jose and spent $20,380 for the whole system. There was a lot of money in that for things a homeowner would not need, but the key thing is that the cameras were each about $1,500.
They bought two of these for $1,299 each:
AXIS Q1700-LE License Plate Camera
and two of these for $1,799 each:
AXIS Q1798-LE Network Camera
Last edited by johnmeyer; 6th Aug 2022 at 00:02. Reason: typo
If I stepped frame by frame, I could see various parts of car plate, not always though. Every 3-4 frames possibly a letter or a digit. So basic idea was if there were a tool to track an object, the plate, and overlay each frames object with various degree of transparency or something, one could figure out a moving cars registration number. Situation is they go really slow passing here, 10 mph or similar.
Basic issue is probably shutter speed, and slow frame rate where changes one frame to the other is too big for H264 predictive frames, especially going all sideways.
This is probably a very narrow field of interest, or there would be such a tool. All is built on good enough cameras to start with.
Best option is to do manually and by sight as far as it goes.
This was already answered above;
Better hardware is the best option
There are quite a few tools used , but not for this exact purpose. The software would be motion tracking software, such as mocha whose specialty is tracking planes, but there are others. This can be used to help stack as jagabo mentioned above. Each area of interest in each frame (the plate) will have translation, rotation, scale, +/- skew, +/- perspective parameters calculated, so that can be used to undistort the plate on each frame , so it looks like it's shot straight on, and each frame is "superimposed" . This can be mainipulated with stacking algorithms. Frames can be weighted higher or lower (e.g. if one frame is clearer), or you can do this with scripts
But if you cannot make out the partial plate number with your eyes on a few frames, and put them together with you mind, it's unlikely any software will help you especially if the recording is low quality. If you post a sample I can tell you if it's even worth trying, I have quite a bit of experience with tracking
There are some interesting developments in machine learning software that can fwd/back autotrack and propogate - but current ones that are publically available do not handle compression artifacts and low quality video very well (the trained models typically do not include an additional lossy compression degradation step). They can sometimes produce good results with low resolution, but clean video.