[Config Support]: Help Reducing Alerts

[matthieu-kr]

Describe the problem you are having

Hello all - I'm looking for help reducing false positives for cars passing in front of my camera. The desired behavior would be any car that loiters for 5s or longer in front of the camera in the "roadway" zone triggers an alert, but I'm getting many false positives for cars that are just passing through. I've tried adjusting the inertia/loitering time, but I still false positives even using very long times (10s-15s or more). It almost appears as if any vehicles that are in-frame longer than 5 or multiple cars pass through consecutively are triggering alerts.

Because I've used the same required zones for both gen AI and alerts, I would expect that any alert has a gen AI entry, but that hasn't been the case. Looking at the object lifecycle, the only events with gen AI entries appear to be those where vehicles truly loiter in the "roadway" zone for longer than 5s, leading me to think the zones are set up correctly. If that's true, then the alerts must be configured incorrectly, but I'm not sure I'm doing wrong there.

Any insight into what I might be missing? Thanks.

Version

0.16.2-4d58206

Frigate config file

street:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/street
          roles: [record, detect]
    detect:
      enabled: true
      width: 960
      height: 720
      fps: 4
    objects:
      track: [car, person, face, license_plate, package, cat, dog, amazon, 
            bicycle, fedex, motorcycle, school_bus, ups, usps]
    genai:
      enabled: true
      required_zones: [roadway, front_yard]
    motion:
      mask: 0.006,0.96,0.007,0.989,0.297,0.987,0.297,0.96
      threshold: 35
      contour_area: 15
      improve_contrast: true
    review:
      alerts:
        enabled: true
        required_zones: [roadway, front_yard]
      detections:
        enabled: true
    zones:
      roadway:
        coordinates: 
          0,0.301,0.078,0.274,0.159,0.254,0.286,0.227,0.409,0.203,0.514,0.193,0.606,0.191,0.696,0.189,0.659,0.137,0.505,0.133,0.443,0.124,0.405,0.124,0.337,0.102,0.225,0.115,0.176,0.126,0.111,0.141,0,0.179
        loitering_time: 5
        inertia: 8
      front_yard:
        coordinates: 
          0,0.347,0,1,1,1,1,0.335,0.821,0.235,0.657,0.236,0.409,0.254,0.197,0.284
        loitering_time: 1
        inertia: 4

Relevant Frigate log output

.

Relevant go2rtc log output

.

Frigate stats

No response

Operating system

UNRAID

Install method

Docker Compose

docker-compose file or Docker CLI command

.

Object Detector

OpenVino

Screenshots of the Frigate UI's System metrics pages

No response

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Based on your configuration, there are a few key issues that may be causing false positives:

Zone Presence Evaluation
Zone presence is evaluated based on the bottom center of the bounding box for the object(1). This means a car only needs its bottom-center point to touch the roadway zone to be considered "in" that zone, even if most of the vehicle is outside it.

Loitering Time Behavior
The loitering_time: 5 setting means an object must remain in the zone for 5 seconds before being considered present(1). However, this doesn't prevent alerts from triggering - it only delays when the object is marked as being in the zone. When using loitering zones, a review item will remain active until the object leaves(1).

Key Configuration Issue
Your review.alerts.required_zones includes both roadway and front_yard(2). This means an alert will trigger if a car enters either zone, not just the roadway. Cars passing through your front_yard zone (which covers a large area from coordinates you've defined) would trigger alerts even if they never loiter in the roadway.

Recommended Changes

1. Separate your alert zones - if you only want loitering alerts from the roadway:

review:
  alerts:
    enabled: true
    required_zones: [roadway]

(2)

2. Increase zone inertia - your inertia: 8 requires the object's bounding box to be within the zone for 8 consecutive frames before being considered present(1). You may need to increase this further if cars are briefly clipping the zone edge.

3. Consider object filters - add a min_area filter for cars to ignore distant/small vehicles that are just passing through:

objects:
  filters:
    car:
      min_area: 5000

(2)

The mismatch between alerts and GenAI descriptions occurs because GenAI uses the same required_zones but only generates descriptions at specific tracking events(3), while alerts can trigger throughout the object's lifecycle.

**

📚 Sources:

1. Zones
2. Full Reference Config
3. Generative AI

---

Have another question? Just tag @inkeep.

[NickM-27]

To be clear, a false positive is when an object was detected that was not actually that object type.

Regarding this issue, can you show an example of an alert that triggered? You can view alert details by tapping on the gray chip icon for a given review item. This will show what has happened.

[matthieu-kr]

Are these the alert details you're looking for? I don't see that anything stops in-zone for longer than 5s. The alerts here don't seem to align with what HA pushes either, though it's set to event type "Alert." I can always find the events in the "Explore" menu in Frigate though.

[NickM-27]

No, it is per vehicle. This simply happened due to what appears to be an ID swap.

[matthieu-kr]

The ID swap would be occurring because the point at which a vehicle enters/exit the frame from the right coincidentally happens to be aligned with the parked vehicle? What's the best approach to reducing the alerts triggered this way?

[NickM-27]

I might suggest making the loitering zone smaller and perhaps increasing the time a few seconds

[matthieu-kr]

I'm assuming object lifecycle tracks a single object as in the example below where vehicles passing opposite directions along the edge of the frame show are reflected as a single object doubling back?

[NickM-27]

Yes