Local realtime person detection for RTSP cameras

but maybe nothing much to worry as not typical situation, this traffic was in the distance of 600 feet however 12mm lens Hikvision cam daylight sunny southern desert location

I need a video clip of the same FPS and resolution that frigate analyzes. It needs to include about 30 seconds of video prior to the event for testing. A photo alone won’t tell me much.

alright, once i notice such behavior with the frigate again I will share the clip in dropbox with you

@seanb-uk here is an attempt at a roadmap: https://github.com/blakeblackshear/frigate/projects/1

4 Likes

:open_mouth:

Blake, you are an animal! Amazing.

Thank you for all your hard work!

2 Likes

Aha, that explains it

Thanks, I’ve taken a look and commented on the MQTT to disable detection issue.

Guys, I just got two Reolink cameras, but it’s hangs on ffprobe.

ffprobe -v panic -show_error -show_streams -of json 
"rtsp://admin:[email protected]:554/h264Preview_01_sub"

If I execute this command in the docker, it works fine, but it will take 5 seconds to execute.

Does anyone have any tips?

On my Reolink, I have better luck with RTMP. Here is some more discussion.
https://forums.zoneminder.com/viewtopic.php?t=25874&start=15

2 Likes

Thank you for answering do you use mainstream or sub?

I use the main stream. The sub stream resolution is not high enough for reliable detection.

1 Like

I use sub-stream since my pi cannot handle mainstream (5MP camera). Detection reliability is good. The tensor flow object detection mandates only 300x300 anyway doesn’t it?

1 Like

It will depend on how big the objects are that you want to detect. As they get further away (smaller percentage of the imaging area) they have less pixels.

Is your camera a Reolink? My RLC-410 only supports a sub stream of 640x360 which is woefully too small.

Can you share your config please?

I have RLC-410 too but if I use man stream the ffmpeg process takes 40% of cpu I have a nuc i3

Thanks

I have a Reolink 511W and I use a 640x480 substream, it can easily detect cats and dogs with approx 1000 pixels area…

here the bottom line is, there will be always users who wants high detection accuracy and stick to full HD main stream. SD detection may amaze you but you clearly had not done many or any tests to see the difference so thats why you are happy with SD too. Actually we are even at the next stage, to improve the full HD stream detection.

It also depends greatly on the range from fhe camera and size of the objects. Basically anything in the 30000 size range will be as well matched as possible, and things above 10000 will generally work fine as it’s not having to fill in much detail. If you are just looking for people on your front porch from your doorbell cam, you will probably be fine with a low res stream.

Accuracy falls off as you reach smaller sizes, but it will also continue to work in some capacity.

I seem only to have two choices: Main stream at 5MP or substream at 640x480, is there any easy way for force the Reolink to say 1080p? It does not give me a choice in the camera config, could it be done in Frigate?

Try to change the resolution from phone or desktop app

yes tried that