-
-
Notifications
You must be signed in to change notification settings - Fork 1.5k
Night time IR (back and white) detection #1519
Replies: 2 comments · 14 replies
-
You will need to watch the stream at night with motion/bounding boxes turned on to see if it is the motion detection that is not sensitive enough or if the object detection scores are too low. |
Beta Was this translation helpful? Give feedback.
All reactions
-
Sure
|
Beta Was this translation helpful? Give feedback.
All reactions
-
No regions defined, it works perfectly during the daytime on the same spot. |
Beta Was this translation helpful? Give feedback.
All reactions
-
You don't define regions in the config. They show up as green boxes when viewing the debug stream when you have the regions switch on. The scores may just be too low at night. Do you have the option to use a low resolution stream? |
Beta Was this translation helpful? Give feedback.
All reactions
-
I ran into a similar issue when testing last night. When on IR mode, motion detection drops to almost nothing. I tested on an indoor and outdoor camera. Outdoor i couldn't trigger it at all, though the point of interest is somewhat distant from the camera. Triggers fine in the day though. Indoors, when I took up a reasonable ammount of the frame, it also didn't trigger motion, except when i walked in front of an open door with a light on. In both cases, the issue wasn't detection, it was that there were no motion events detected (no red boxes). Any suggestions? Side note: It would be cool if we could set two sets of motion trigger thresholds, and either switch between on a schedule or via mqtt. |
Beta Was this translation helpful? Give feedback.
All reactions
-
👍 3
-
I'm wondering if ramping up sensitivity all the time and just paying the detection cost is a good solution (my coral is VERY underutilized as i understand it) |
Beta Was this translation helpful? Give feedback.
All reactions
-
Bit of a bump but I experienced the same problem. There is a Tensorflow model here which has been trained on greyscale images, it works a little better for night time detection but unfortunately it now detects cats as people, lol. https://gitlab.com/Curid/TF-CCTV I believe, but I am not sure at all that currently Frigate just uses the default sample for object detection? |
Beta Was this translation helpful? Give feedback.
All reactions
-
😄 1
-
Yes, I have that on too but great suggestion and for sure is the best place to start. The night vision on my camera isn't the best so I did have to up the motion sensitivity and use that improve_contrast you suggested. I think I need to up the motion detection a bit more too but for the first time in pretty much forever I saw a few greyscale snapshots in the events listing since uploading that model. I'm still testing it anyway, I will probably switch between the standard one and this for more testing. I am just trying to find the best way to setup a test camera as I am finding more and more that I need to be able to test footage to see what happened. I wish there was the option to see boundary boxes etc during event reviews. |
Beta Was this translation helpful? Give feedback.
All reactions
-
Reviewing footage can be done manually for now using: https://docs.frigate.video/contributing#2-create-a-local-config-file-for-testing |
Beta Was this translation helpful? Give feedback.
All reactions
-
is the The same scene during the day time with colors and light in Debug mode are always full of motion boxes, while the tree branches move in the wind. In the night view I can clearly see the branches moving heavily but looks like the motion is not detected. |
Beta Was this translation helpful? Give feedback.
All reactions
-
it is a frigate setting: motion:
improve_contrast: true |
Beta Was this translation helpful? Give feedback.
All reactions
-
👍 1
-
Thanks, much better now.. at least some motion is being detected. Will see about the detection. Will test if humans are detected once I will have a chance to step in front of the camera again. |
Beta Was this translation helpful? Give feedback.
-
Managed to get this system working with home assistant with RP4 pretty well - great and massive kudos for the team behind it!
During the day time it works really well - detects people very well and occasionally considers my narrow walking path as giraffe and other fun :)
But during the night time when Hikvision camera will switch to black and white with infrared light enabled.. there is zero detection even if the person is really close. Is there any extra conf required or a bit of more training for AI needs to be done? Because night time is probably most of the serious use of this camera in case of security.
Beta Was this translation helpful? Give feedback.
All reactions