Virtual Cameras in Frigate.md

Virtual Cameras in Frigate

Create extra views (sky-only, cropped, perspective-corrected, etc.) from one physical camera. Handy, but very resource intensive.

---

Contents

• Implementation
• Performance Review
• Recommendations
• Alternative
• Conclusion

---

Implementation

1 – Go2RTC stream

go2rtc:
  streams:
    # Physical camera
    balcony:
      - rtsp://user:pass@camera_ip/stream1

    # Virtual camera with filters
    balcony_sky:
      - ffmpeg:rtsp://user:pass@camera_ip/stream1#video=h264#raw=-vf \
        perspective=x0=80:y0=400:x1=1342:y1=104:x2=360:y2=1340:x3=1554:y3=1330,scale=1280:720#hardware

2 – Camera block

cameras:
  balcony_sky:
    ffmpeg:
      output_args:
        record: preset-record-generic-audio-aac
      inputs:
        - path: rtsp://127.0.0.1:8554/balcony_sky
          roles: [record, detect]
    detect:
      width: 848
      height: 480
      fps: 5
    snapshots:
      enabled: true
      timestamp: true
      bounding_box: true
      quality: 95
    record:
      enabled: true
      retain:
        days: 30
        mode: motion
    motion:
      mask: [coords]
      threshold: 30
      contour_area: 10
      improve_contrast: true
    zones:
      Park:
        coordinates: [coords]
        loitering_time: 0
        objects: person
    objects:
      mask: [coords]

Key FFmpeg filters

Filter	Purpose
perspective=…	Rectifies angled shots
scale=1280:720	Down-scales to ease downstream load
#hardware	Uses GPU/ASIC if available

---

Performance Review

Method

1. 30-second CPU sample, virtual cam off
2. 30-second CPU sample, virtual cam on
3. docker stats --no-stream every second
4. Average the samples, extrapolate to 24 h

Environment

Item	Value
Host kernel	Linux 6.8.12-13-pve
Frigate	0.15.0-beta2
Camera feed	1080 p RTSP (persp. + scale)

Results

State	Avg CPU %	CPU-s / day	Samples
Virtual cam OFF	84.7	73 147	30
Virtual cam ON	266.2	230 016	30
Delta	+181.6 %	+156 869	30

Take-aways

• Roughly 3× CPU for one virtual cam.
• Adds ~44 CPU-hours per day.
• Filters (perspective + scale) are the main culprits.
• Expect similar overhead for each extra virtual camera.

---

Recommendations

Use a virtual camera when:

• Hardware acceleration is available.
• A corrected or specialty view is important.
• Different areas need different detection settings.

Don't use one when:

• The host is already CPU-bound or has thermal issues.
• A simple crop via zones is enough.

---

Alternative

Detection zones

zones:
  Sky_Area:
    coordinates: [coords]
    objects: []
  Ground_Area:
    coordinates: [coords]
    objects: [person, car]

---

Conclusion

Virtual cameras work reliably, but are extremely resource intensive. In my case, the cost of using the feature for over 6 months has exceeded the cost of a second camera.

---

Tested on Frigate 0.15.0-beta2. Results will vary with different filters, hardware, and acceleration support.