[Config Support]: Stationnary cars - fine tuning false detection #11277
-
Describe the problem you are havingGreetings to all, forgive me in advance for my lack of IT knowledge. I experience frequent false detections on stationary cars, highly taxing on storage. Could I take your time for a kind of pratical "case analysis", and see how you would think ? I wonder which combination of parameters (zones, min/max areas & ratios, object filters) would be most optimal. Do not hesitate to ask for any information I missed. I hope to benefit from your guidance and patience due to my beginner level. Thanks ! Version0.13.2-6476F8A Frigate config filemqtt:
enabled: false
ffmpeg:
hwaccel_args: preset-vaapi
detectors:
coral:
type: edgetpu
device: usb
record:
enabled: True
retain:
days: 7
mode: motion
events:
retain:
default: 30
mode: motion
snapshots:
enabled: True
retain:
default: 30
objects:
track:
- person
- cat
- dog
- car
- bike
# filters:
# person:
# min_area: 5000
# max_area: 100000
cameras:
FrontCam:
ffmpeg:
inputs:
# High Resolution Stream
- path: rtsp://admina:rootbeer123@192.168.1.79:554/H264/ch1/main/av_stream
roles:
- record
# Low Resolution Stream
- path: rtsp://admina:rootbeer123@192.168.1.79:554/H264/ch1/main/av_stream
roles:
- detect
detect:
width: 3632
height: 1632
fps: 8
motion:
mask:
- 0,749,3632,689,3632,686,3632,0,0,0
- 656,858,1288,892,1420,828,1325,753,989,660,652,739 Relevant log output2024-05-07 14:54:06.378219838 [INFO] Preparing Frigate...
2024-05-07 14:54:06.395726299 [INFO] Starting Frigate...
2024-05-07 14:54:07.490546038 [2024-05-07 14:54:07] frigate.app INFO : Starting Frigate (0.13.2-6476f8a)
2024-05-07 14:54:07.515536903 [2024-05-07 14:54:07] peewee_migrate.logs INFO : Starting migrations
2024-05-07 14:54:07.519263950 [2024-05-07 14:54:07] peewee_migrate.logs INFO : There is nothing to migrate
2024-05-07 14:54:07.526032621 [2024-05-07 14:54:07] frigate.app INFO : Recording process started: 289
2024-05-07 14:54:07.526407893 [2024-05-07 14:54:07] frigate.app INFO : go2rtc process pid: 97
2024-05-07 14:54:07.551623591 [2024-05-07 14:54:07] detector.coral INFO : Starting detection process: 298
2024-05-07 14:54:10.164453634 [2024-05-07 14:54:07] frigate.app INFO : Output process started: 300
2024-05-07 14:54:10.164561941 [2024-05-07 14:54:07] frigate.app INFO : Camera processor started for FrontCam: 314
2024-05-07 14:54:10.164650285 [2024-05-07 14:54:07] frigate.app INFO : Capture process started for FrontCam: 315
2024-05-07 14:54:10.165705522 [2024-05-07 14:54:07] frigate.detectors.plugins.edgetpu_tfl INFO : Attempting to load TPU as usb
2024-05-07 14:54:10.168735622 [2024-05-07 14:54:10] frigate.detectors.plugins.edgetpu_tfl INFO : TPU found
2024-05-07 14:59:07.600789056 [2024-05-07 14:59:07] frigate.storage INFO : Less than 1 hour of recording space left, running storage maintenance...
2024-05-07 14:59:07.627177130 [2024-05-07 14:59:07] frigate.storage INFO : Cleaned up 2787.599999999999 MB of recordings Frigate statsNo response Operating systemProxmox Install methodDocker Compose Coral versionUSB Any other information that may be helpfulPlease find snapshots which illustrate my situation. Grey car "false detect" :Black car "false detect": |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
you could try to use a max area or min_area filter but that might block out real detections too. This is one of those cases where the default model just struggles because it is not trained on images from cameras. We can see that in these cases even the "accurate" boxes are scoring really low This is where something like Frigate+ or another improved model would likely work better. Frigate+ does allow you to specifically fine tune the model with your images so you can teach it that those boxes are not cars / correct the smaller boxes that are not capturing the full car. |
Beta Was this translation helpful? Give feedback.
you could try to use a max area or min_area filter but that might block out real detections too. This is one of those cases where the default model just struggles because it is not trained on images from cameras. We can see that in these cases even the "accurate" boxes are scoring really low
This is where something like Frigate+ or another improved model would likely work better. Frigate+ does allow you to specifically fine tune the model with your images so you can teach it that those boxes are not cars / correct the smaller boxes that are not capturing the full car.