Lake Travis live cam: Frigate + Frigate+ for boats/birds, PTZ autotracking, YouTube chat-controlled presets

[pgregg88]

Hey all — wanted to share a Frigate-based build that's been running 24/7 for a few months now. Lots of decisions that took trial-and-error to get right, so posting in case any of it saves someone time.

What it is

A EmpireTech (Dahua) PTZ4E25X-T180 pointed at Lake Travis (Texas), live on YouTube 24/7 with viewers able to type preset names in chat to move the camera.

• Live stream: Lake Travis Weather & Interactive Boat Cam

Frigate is the central nervous system — detection, autotracking, recording, and the go2rtc restream that everything else hangs off of.

Hardware

• NVR: Dell laptop with Intel i7-1165G7 (Iris Xe iGPU for VAAPI decode)
• Detector: Coral USB Edge TPU
• Cameras : PTZ4E25X-T180 (detail camera)
• Storage: 2TB ext4 USB drive (/media/2tb)
• Frigate: 0.17.1, Frigate+ subscription

Frigate config highlights

Custom Frigate+ model for the lake

COCO out of the box was unusable — sun glare, jet skis, and shoreline rocks all kept getting called "boats." Submitted a few thousand tagged samples to Frigate+ and the lake-specific model is dramatically better.

model:
  path: plus://your-model-hash
  width: 320
  height: 320

cameras:
  lake_cam_ptz:
    objects:
      track: [bird, boat, person]
      filters:
        boat:
          min_score: 0.2
          threshold: 0.4
          min_area: 100
          max_area: 500000

Coral inference at ~9ms — never hits a bottleneck.

Killing motion noise from water ripples

This was the biggest tuning fight. Default motion settings light up constantly when wind hits the lake. What landed:

motion:
  threshold: 30
  contour_area: 46     # was 15
  improve_contrast: false   # was true — was amplifying ripple noise
  lightning_threshold: 0.6
  frame_alpha: 0.04
  frame_height: 80     # was 100 — halving was the biggest single win

Reducing frame_height from 100 to 80 pushed ripples below the contour threshold. improve_contrast: false was counter-intuitive but key.

Autotracker on a far-away scene

The lake is 200m+ out and the PTZ has 25x optical zoom, which makes calibration brittle. Small detections near the horizon try to drag the camera around.

autotracking:
  enabled: true
  zooming: relative
  zoom_factor: 0.3
  track: [bird, boat, person]
  required_zones: [boat_zone]
  return_preset: BoatsNorthEastZoom
  timeout: 120
  movement_weights: 0.04, 1.0, 1.632584571838379, 1.4581075381207214,
    0.17448916127604824, 1.2832341194152832

stationary:
  classifier: true
  interval: 15
  threshold: 100

Movement weights were arrived at empirically over weeks. Stationary classifier + interval 15 stops the detector from hammering on parked or anchored boats.

Detect resolution must match the sub-stream

Spent a while wondering why lake_cam_ptz detect was burning ~42% CPU. Turned out the detect block requested 1920x1080 @ 10fps but the actual sub-stream is 1280x720 @ 5fps — ffmpeg was upscaling and duplicating frames for nothing.

# Verify your own:
docker exec frigate /usr/lib/ffmpeg/7.0/bin/ffprobe -v error \
  -select_streams v:0 -show_entries stream=width,height,avg_frame_rate \
  -of csv=p=0 "rtsp://127.0.0.1:8554/your_cam_sub"

After matching, CPU dropped from 42% to ~10%. Worth checking on every camera. The iGPU (scale_vaapi) makes the rescale "free" but the CPU work to download, format-convert, and pipe the larger frames isn't.

Recording strategy

With alerts (30d) and detections (14d) capturing every tracked event with 5s pre/post, continuous 24/7 recording is mostly redundant for cameras where you only care about events. After the cleanup:

Camera	Continuous	Motion	Alerts	Detections
driveway, backyard	1d	3d	30d	14d
basement, kitchen	0	0	30d	14d
lake_cam_ptz	recording disabled	—	—	—

Lake cam doesn't need recording — the only useful artifacts are the snapshots that go to Frigate+ for training. Snapshot retention bumped to 36500 days ("forever until manual delete").

Net: ~172 GB/day → ~73 GB/day, no loss of useful coverage.

YouTube chat → PTZ via MQTT

The fun bit. Architecture:

[YouTube Live Chat] ──pytchat──▶ [chat_bot.py] ──MQTT──▶ [Frigate] ──▶ [PTZ]
                    ◀─Data API (greetings)─┘                ▲
                                                            │
                                                  Autotracker ON/OFF

When a viewer types a preset name (or !preset / /preset):

1. Bot publishes to frigate/lake_cam_ptz/ptz (Frigate's standard MQTT topic — no Frigate fork required).
2. Disables autotracker via API so it doesn't fight the viewer.
3. Camera moves to the preset.
4. After 5 min of chat inactivity, autotracker re-enables itself.

Also greets first-time chatters once per session. YouTube Data API quota ledger with kill-switch and lockout state, because I learned the hard way when a runaway loop ate the daily 10k-unit quota at 3am.

pytchat for reads (no quota), Data API for writes (greetings) only.

Stack summary

Layer	Tech
NVR	Frigate 0.17.1 + Coral USB
Detection model	Frigate+ (custom)
Live view	go2rtc → WebRTC
Public control panel	React + Vite + Caddy reverse proxy
YouTube broadcaster	Node.js service, ffmpeg with VAAPI overlay banner
Chat bot	Python service, pytchat + MQTT
Process supervision	systemd

Things that took longer than expected

• Movement weight calibration on a long-range PTZ. No real shortcut. Recorded a clip, watched the autotracker chase the wrong things, adjusted, repeated. Frigate's docs are a starting point but every scene is its own thing.
• Knowing when to use sub-stream vs main stream. Detect on sub, record on main, restream both via go2rtc — but only if the sub-stream is at a resolution/fps your model can use.
• Realizing motion settings have to fight the dominant noise source. For me: water. For someone else it might be foliage or headlights. The defaults are tuned for the average case, not the hard ones.

Not yet solved

• Boat detection still occasionally misses small distant fishing boats (likely below the model's effective resolution after 320x320 resize). Considering a second pass at higher zoom.
• Autotracker sometimes loses track when a boat crosses zone boundaries.
• Coral runs hot when boats are heavy on summer weekends — eyeing a small fan.

Thanks

To Blake and the Frigate team — six months ago I had no idea what an edge TPU was, and now I'm running a small surveillance + livestream system that I trust. The Frigate+ feedback loop is genuinely impressive.

Happy to answer questions or share more config. Repo isn't public yet (cleaning up paths and secrets), but planning to open-source the chat bot specifically since it's reusable for any Frigate+YouTube setup.

[hawkeye217]

Thanks for sharing! As a fellow Texan, it's great to see code you've written used in such a unique way.