[Config Support]: Unexpected Alerts in Night vision

[jfh250]

Describe the problem you are having

For the life of me I can’t understand why I’m getting ‘Alerts’ when my zone rules are otherwise working correctly during daylight. When my camera flips to night vision, things get weird. Furthermore, what’s happening isn’t supported by what’s reported in object lifecycle - this is the most confusing bit, I don't see how alert conditions are met without this.

Essentially, I have two zones, Driveway & Road. People and Cars on my Driveway should be classified as an Alert. People and Cars in the Road should be classified as a Detection.
This works as expected during hours of daylight. But when the camera flips to night vision, Passing cars become prone to classification as an Alert rather than a Detection. This would suggest they must have interacted with my Driveway zone. However, and the bit I find most puzzling is that reviewing the object Lifecycle for these instances confirms all activity took place in the Road.
I don’t really know how to troubleshoot, due to the conflicting information in Object Lifecycle. Watching the scene in debug also doesn’t really help.

I would entirely expect object lifecycle to suggest some interaction with the Driveway, but it doesn't. Hence I don't know how alert conditions are being achieved, but they are.

I don't think it's related, but but in my screenshots below, it does appear to have tracked a stationary car, parked in the road, caused by the motion of an actual passing car. But again, it all happened in the Road zone, so this shouldn't amount to an alert.

Version

0.15-1

Frigate config file

# ==========================
# SYSTEM CONFIGURATION
# ==========================
mqtt:
  enabled: false  # Disable MQTT if not using it

detectors:
  coral:
    type: edgetpu
    device: usb

ffmpeg:
  hwaccel_args:
    - -hwaccel
    - vaapi
    - -hwaccel_device
    - /dev/dri/renderD128
    - -hwaccel_output_format
    - yuv420p

# ==========================
# CAMERA CONFIGURATION
# ==========================
cameras:
  front:
    ffmpeg:
      inputs:
        - path: rtsp://frigate:password@192.168.1.5:554/Streaming/Channels/102
          roles:
            - detect
        - path: rtsp://frigate:password@192.168.1.5:554/Streaming/Channels/101
          roles:
            - record

    detect:
      width: 640
      height: 480
      fps: 8
      stationary:
        threshold: 240 # How long before Frigate considers an object "stationary" (FPS) (reflecting headlight problem)
        interval: 100 # How often Frigate checks if a stationary object moved (FPS) (reflecting headlight problem)


    objects:
      track:
        - person
        - car

    # ==========================
    # RECORDING CONFIGURATION
    # ==========================
      filters:
        person:
          mask: 0.535,0.123,0.557,0.121,0.557,0.164,0.535,0.167
    record:
      enabled: true
      retain:
        days: 0 # Keep recordings for 35 days
        mode: active_objects
      alerts:
        retain:
          days: 35
          mode: active_objects
      detections:
        retain:
          days: 35
          mode: active_objects


    # ==========================
    # SNAPSHOT CONFIGURATION
    # ==========================
    snapshots:
      enabled: true
      timestamp: false
      bounding_box: true
      retain:
        default: 35  # Keep snapshots for 35 days
      required_zones:
        - Road
        - Driveway


    # ==========================
    # MOTION DETECTION SETTINGS
    # ==========================
    motion:
      mask:
        - 0.268,0.047,0.268,0.082,0.334,0.082,0.334,0.039
        - 0.429,1,0.137,0.618,0.004,0.727,0.001,0.993
        - -0.001,0.354,0.08,0.251,0.136,0.237,0.144,0.038,0.132,0,-0.001,0.001
        - 1,0.182,0.665,0.074,0.394,0.066,0.157,0.086,0.147,0.005,0.996,0.001
        - 0.819,0.993,1,0.357,0.999,0.994
        - 0.217,0.136,0.183,0.23,0.354,0.266,0.428,0.215,0.424,0.153,0.315,0.141
        - 0.895,0.383,0.86,0.558,0.752,0.523,0.594,0.438,0.636,0.322,0.765,0.329
      threshold: 45
      contour_area: 25
      improve_contrast: true

    # ==========================
    # ZONE CONFIGURATION
    # ==========================
    zones:
      Driveway:
        coordinates: 
          0.313,0.988,0.916,1,0.956,0.428,0.896,0.372,0.716,0.323,0.662,0.312,0.445,0.24,0.424,0.205,0.404,0.195,0.226,0.166,0.112,0.149,0.002,0.28,0.003,0.761
        filters:
          person:
            min_area: 5000  # Adjust as needed to filter out false positives
          car:
            min_area: 8000  # Adjust as needed

        inertia: 3
        loitering_time: 0
        objects:
          - car
          - person
      Road:
        coordinates: 
          0.963,0.403,1,0.107,0.657,0.009,0.521,0,0.15,0,0.152,0.144,0.406,0.186,0.448,0.233,0.675,0.307,0.766,0.333
        filters:
          person:
            min_area: 300 # Adjust as needed
          car:
            min_area: 2000 # Adjust as needed
        loitering_time: 0  # Record cars only if they linger for 7 seconds or more (temp disabled)
        inertia: 2

        objects:
          - car
          - person
    review:
      alerts:
        required_zones: Driveway
      detections:
        required_zones: Road
version: 0.15-1

Relevant Frigate log output

N/A

Relevant go2rtc log output

N/A

Frigate stats

No response

Operating system

Debian

Install method

Docker Compose

docker-compose file or Docker CLI command

N/A

Object Detector

Coral

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[NickM-27]

Are you sure it's that specific car that is triggering the alert? If I had to guess this is an object detection issue where the car in the driveway has an unstable box at night

[jfh250]

This is the timestamp that matches the alert. I would go for that idea if I could find something corresponding to activity in the Driveway zone to cause it, but so far I haven't been able to. I can try and get some more examples from around the same time tomorrow.

[NickM-27]

You'll need to check the object lifecycle of the event in explore for the car in the driveway

[jfh250]

ah interesting, If I filter my explore items to Driveway and look at Object lifecycle I can see that the car on the Driveway is:
06:35:20 Detected
06:35:20 Entered Driveway
06:35:50 Stationary
06:38:24 Active (the picture indicates a car is just about to pass in the road - it's headlights have entered the scene and are lighting up the road and to some extent the car on the driveway).
06:38:55 Stationary.
This overlaps with the passing car in the road.
So, I guess the headlights from passing cars are triggering false positives against my own car in the driveway?
You can see from the screenshot below the headlights are just starting to sweeping into the scene.

[NickM-27]

yeah unfortunately looks like that excessive motion is causing the object detection to detect the car incorrectly