Local realtime person detection for RTSP cameras

Share your Projects!
122

@blakeblackshear Thanks for your effort, these are exciting times !
I really like the idea of realtime detection and therefore I will order a coral

123

Who has the best prices for the coral?

124

I havenā€™t seen anywhere to buy them except the links directly from the Google website.

1 Like
125

I setup a Coral for a 5 camera system that has 36 detection regions this morning. I am definitely maxing out the Coral, because the processing queue is actually filling up. 36 regions at 5 fps means the Coral is looking at 180 images per second.

3 Likes
126

Hope this works on dedicated rpi too.

127

It does, just at 1/10th the speed of a machine with usb 3.0.

128

Google Coral for europeans.

https://www.mouser.se/ProductDetail/Coral/G950-01456-01?qs=sGAEpiMZZMtw0nEwywcFgJjuZv55GFNmI6%2BQrIB9AdMsI%2BDXofN1SA==

129

As I understand it Frigate will continuosly scan a feed for objetcts. Would it be possible to just send stills och maybe 10 second clips (from new stream component) and have that analyzed and then acted upon? Or does one have to continously monitor a stream?

1 Like
130

I just got my coral deliveredā€¦ tried the demos from the site on a work machine, looking forward to trying this out with my cameras a home soon.

131

how do I set up a single region to just cover the whole frame? my cam is putting out 1600x900.

thanks.

Also this is what Iā€™m getting on the console - does this mean itā€™s not able to connect to the rtsp stream?

On connect called
On connect called
On connect called
On connect called
On connect called
Traceback (most recent call last):
  File "detect_objects.py", line 90, in <module>
    main()
  File "detect_objects.py", line 44, in main
    cameras[name] = Camera(name, config, prepped_frame_queue, client, MQTT_TOPIC_PREFIX)
  File "/opt/frigate/frigate/video.py", line 119, in __init__
    self.frame_shape = get_frame_shape(self.rtsp_url)
  File "/opt/frigate/frigate/video.py", line 100, in get_frame_shape
    frame_shape = frame.shape
AttributeError: 'NoneType' object has no attribute 'shape'
132

I think the regions have to be square so you couldnā€™t do one for full frame. My guess in the region setup in your config.yml file is causing those errors.

133

thanks, yea think that was part of it. though I think one of my cameras is having issues connecting as well. I got one working with a 720^2 region which picks up most of the area, and it is working well!

134

I have my frigate container on a switch in HA so I can turn it on/off from frontend and in automations. Sometimes I get a bad feed and a restart of the container normally fixes it after a few tries.

135

I lowered the camera stream to 1280x720 and now have 2 cams working. they are running at 25 fps. is this why the mjpeg stream out of frigate has the heavy overlapping of the bounding boxes?

43%20PM
43%20PM1774Ɨ1486 474 KB

136

Pretty much. It draws boxes for all frames in the past second rather than just the current frame. So 25fps could result in more than 25 overlapping boxes drawn. The MJPEG stream is really intended for debugging purposes.

1 Like
137

new updates to the Coral, now allows running multiple corals (with different models if you wanted).

1 Like
138

Nice. Hopefully this will allow me to run the container without privileged mode.

139

So, thinking of buying a coral and put it in my NUC. What is the power consumption of having the coral analyzing lets say two streams continously 24/7 with some events every hour? Will it add single digit watt consumption or more?

140

I havenā€™t measured it, but the datasheet says it requires at least 500mA. At 5v, that would be 2.5 watts.

141

Thats nothing, I will just have to order a Coral I guess.