# Frigate NVR & CCTV Setup **Last Updated:** May 2026 **Host:** thinkstation-p710 — `192.168.88.41` --- ## Hardware | Component | Details | |-----------|---------| | Machine | Lenovo ThinkStation P710 | | CPU | Intel Xeon E5-2623 v4 @ 2.60 GHz (4 cores / 8 threads, boost to 3.2 GHz) | | RAM | 24 GB | | GPU | NVIDIA GeForce GTX 1050 Ti (4 GB VRAM) | | OS Disk | 234 GB SSD (`/dev/sda2`) | | Media Storage | 14 TB HDD — mounted at `/mnt/14TB-HD/frigate/media` | --- ## Frigate Docker Setup | Item | Value | |------|-------| | Image | `ghcr.io/blakeblackshear/frigate:stable-tensorrt` | | Container name | `frigate` | | Config path | `/srv/docker/frigate/config/` | | Media path | `/mnt/14TB-HD/frigate/media` | | Compose file | `thinkstation/frigate/docker-compose.yml` (this repo) | | Web UI | `http://192.168.88.41:5000` → `https://cameras.jgitta.com` | | RTSP port | `8554` | | WebRTC port | `8555` | **GPU acceleration:** Container runs with `runtime: nvidia` and `NVIDIA_VISIBLE_DEVICES=all`. The GTX 1050 Ti handles: - LPR (licence plate recognition) via GPU - Birdseye re-stream encoding via `h264_nvenc` - ONNX object detection via the `onnx` detector **Shared memory:** `shm_size: 1024mb` — required for frame buffers across 6 cameras. --- ## Object Detection | Item | Value | |------|-------| | Detector | ONNX (GPU-accelerated) | | Model | YOLOv9-t-320 (`yolov9-t-320.onnx`) | | Input size | 320 × 320 | | Label map | COCO-80 | | Model path | `/config/model_cache/onnx/yolov9-t-320.onnx` | --- ## Licence Plate Recognition (LPR) | Item | Value | |------|-------| | Enabled | Yes | | Device | GPU | | Model size | Small | **Known plates:** | Name | Plate | |------|-------| | Joe | ABC1234 | | Cynthia | ABC4567 | | Wesley | XYZ9876 | --- ## Camera Inventory All cameras are **Reolink** PoE units on the `192.168.88.0/24` subnet. | Camera | IP | Model notes | Mount location | |--------|----|-------------|----------------| | front_driveway | 192.168.88.105 | Dual-lens panoramic (7680×2160 = two 4K sensors stitched) | Front driveway | | front_yard | 192.168.88.104 | 4K 811A | Front yard | | shop | 192.168.88.103 | 4K 811A (uses HTTPS for API) | Shop / outbuilding | | deck | 192.168.88.101 | 4K 811A | Deck | | back_yard | 192.168.88.102 | 4K 811A | Back yard | | front_right | 192.168.88.100 | 4512×2512 sensor (non-standard resolution) | Front right | --- ## Camera Stream Settings All cameras confirmed via Reolink API (May 2026): | Camera | Main Resolution | Main Codec | Bitrate | FPS | I-frame | Sub Resolution | Sub Codec | |--------|----------------|------------|---------|-----|---------|----------------|-----------| | front_driveway (.105) | 7680×2160 | H.265 | 10240 Kbps | 20 | 1x | 1536×432 | H.264 | | front_yard (.104) | 3840×2160 | H.265 | 6144 Kbps | 25 | 1x | 640×360 | H.264 | | shop (.103) | 3840×2160 | H.265 | 6144 Kbps | 25 | 1x | 640×360 | H.264 | | deck (.101) | 3840×2160 | H.265 | 6144 Kbps | 25 | 1x | 640×360 | H.264 | | back_yard (.102) | 3840×2160 | H.265 | 6144 Kbps | 25 | 1x | 640×360 | H.264 | | front_right (.100) | 4512×2512 | H.265 | 8192 Kbps | 20 | 1x | 896×512 | H.264 | **I-frame interval** set to 1x on all cameras so Frigate can start/end recording clips cleanly without waiting for a keyframe. --- ## go2rtc Stream Architecture Chrome/Firefox cannot play H.265 (HEVC) via MSE (Media Source Extensions), so each camera requires two go2rtc streams: ``` camera_h265 → native H.265 RTSP from camera └─ consumed by Frigate ffmpeg for RECORDING (stream copy, no re-encode) camera → ffmpeg:camera_h265#video=h264 (software H.265 → H.264 transcode) └─ served to browser via MSE WebSocket for LIVE VIEW camera_sub → native low-res H.264 RTSP sub-stream └─ consumed by Frigate for OBJECT DETECTION only ``` **Key behaviour notes:** - go2rtc uses **lazy initialization** — transcoders only start when a browser opens the live view, and shut down ~30 seconds after the last viewer leaves. This causes a 2–3 second gray screen on first view, which is expected and unavoidable. - The `_h265` streams are kept alive continuously by Frigate's recording ffmpeg processes, so recording starts instantly. - The `#hardware` flag for NVENC transcoding in go2rtc causes EOF crashes with Reolink cameras and **must not be used**. Software transcoding (`libx264`) is used instead. - Reolink cameras drop RTSP connections every 5–20 minutes (firmware behaviour). go2rtc reconnects automatically; this produces a brief gray screen, not static. ### shop camera note The shop camera (`192.168.88.103`) redirects HTTP → HTTPS. All API calls to shop must use `https://` with `-k` (skip cert verification). --- ## Frigate Camera Detection Zones | Camera | Detect Resolution | Tracked Objects | Notes | |--------|------------------|-----------------|-------| | front_driveway | 1536×432 | person, car, dog, cat, license_plate | Has zone + LPR; alerts require `front_driveway_zone` | | front_yard | 640×360 | person, car, dog, cat | — | | shop | 640×360 | person, car, dog, cat | Motion mask on upper-right and mid-frame | | deck | 640×360 | person, car, dog, cat | Two motion masks | | back_yard | 640×360 | person, car, dog, cat | Two motion masks | | front_right | 896×512 | person, car, dog, cat | Motion mask on upper-centre | --- ## Recording & Retention | Type | Retention | |------|-----------| | Alert clips | 30 days | | Detection clips | 30 days | | Motion clips | 30 days | | Continuous | Disabled | | Snapshots | 30 days | --- ## Known Issues & Quirks - **Reolink H.265 sub-stream firmware bug:** On some Reolink firmware versions, setting the sub-stream to H.265 produces a corrupted bitstream that go2rtc rejects with 404. Sub-streams should always remain H.264. - **front_driveway panoramic:** The 7680×2160 resolution is two 4K sensors stitched side by side. The wide 1536×432 detect resolution in the Frigate config accounts for this. - **front_right non-standard resolution:** Camera reports 4512×2512 rather than standard 4K. Model unknown — may be a 5MP variant or older sensor. - **Swap pressure:** ThinkStation runs ~5.5 GB swap used under normal load. Monitor if adding more cameras or enabling Birdseye.