Building a Low-Power, High-Performance Frigate NVR System with ArmSoM Sige7

List of Features and Technical Specifications

Category	Specification
System Type	Home Security System + Network Video Recorder (NVR)
Processor	ArmSoM Sige7 (RK3588), 8-core ARM CPU
NPU	6 TOPS for YOLO object detection, license plate recognition (OCR), face recognition
GPU/VPU	Hardware VPU for H.264/H.265, decoding up to 32 channels of 1080p@30fps
Memory	8GB LPDDR4
Storage	M.2 SSD slot (1TB local storage installed)
Networking	Dual 2.5GbE ports (camera LAN isolation from home LAN)
Power	USB-C, idle 2.5W / full load ~10W (15W budget)
Cooling	Passive, target thermal stability at 65°C (ambient 45°C)
Software Stack	Ubuntu + Docker + Frigate NVR + Home Assistant + MQTT + ffmpeg
Camera Support	8 cameras × 2 streams (1080p@25fps recording + 640×480@5fps detection)
Automation	Home Assistant (WhatsApp, AI agent, smart device control)
Remote Access	Mobile video/snapshot review, local + cloud archive

Project Background

This system is deployed in a large family home in Sydney, Australia. The goal was to build a low-power, high-performance NVR running 24/7. Beyond basic security monitoring and logging, the system needs to:

• Recognize people, cars, motorcycles, and cats, and trigger different automations based on the scenario

• Detect garage door open/closed status and lights on/off via image recognition (no traditional sensors available in the old house)

• Send snapshots of people/cars/motorcycles to an external AI agent to identify mail or package delivery, then push WhatsApp notifications

• Feed cat detection images to an external Home Automation system to catalog and store to enable cat finder utility

The system must integrate with Home Assistant to access AI agents, WhatsApp gateway, and smart device controllers. Local and cloud archive and mobile review are also required.

 

Challenges Faced

1. Reliable image recognition under extreme lighting changes Garage door status recognition faces full daylight sun to nighttime darkness. Multiple rounds of image training were required.

2. License plate recognition (LPR) under harsh conditions LPR must run in real-time across day, night, and rainstorms using NPU-based OCR. When the homeowner drives up the driveway, the time from positive LPR to garage door/lights activation must be under 2 seconds with very high accuracy.

3. Thermal constraints Sydney summer temperatures often exceed 40°C, reaching up to 45°C. With a thermal budget of ambient +20°C, the system must be stable at 65°C with a safety buffer at 70°C.

4. Performance vs. power balance

• 8 cameras continuously streaming recording (1080p@25fps) + detection (640×480@5fps)

• Dual network ports for camera/home LAN isolation

• Total power budget ≤15W

• Silent operation, low heat output, high stability

5. Open-source software ecosystem Must use industry-standard processor with full open-source support (Linux, Docker, Frigate, etc.)

Screenshot of RKTOP performance monitor for RK3588.  CPU, GPU, NPU just idling and temperature at 49 degrees C.

The Sige7 Solution (Why ArmSoM Sige7?)

The Sige7 board (RK3588) is one of the few platforms that meets all key requirements simultaneously:

Requirement	How Sige7 Meets It
Low power	Idle 2.5W, full load ~10W (fits 15W budget)
High-performance CPU	8-core ARM, ~2× performance of Raspberry Pi 5
Video processing	Hardware VPU decodes 32×1080p@30fps; ffmpeg hardware transcoding
NPU acceleration	6 TOPS for YOLO, LPR (OCR), face recognition – offloads CPU/GPU
Sufficient memory	8GB LPDDR4 fits OS, cache, Frigate system, and video frame buffers
Dual 2.5GbE	Physical isolation between camera network and home LAN
M.2 SSD	1TB local storage for smooth video scrubbing
Open-source support	Optimized Ubuntu with stable Rockchip drivers; runs Frigate + Home Assistant perfectly
Thermal stability	CPU temperature ~49°C (ambient 45°C), well below 65°C target

Implementation Process

Software Stack – Frigate NVR Leverages Sige7’s heterogeneous computing:

• Video decoding/transcoding – ffmpeg with GPU/VPU hardware acceleration (H.264/H.265)

• Storage optimization – Direct-to-disk recording + RAM cache + segment writing to reduce SSD wear

• Motion detection – CPU handles lightweight motion detection and audio transcoding

• Object detection & recognition – NPU runs YOLO for detection and bounding boxes; if recognition is needed (face, LPR, subclass like "delivery truck"), NPU handles that as well

• License plate recognition – Both plate detection and OCR run on NPU (saving GPU/CPU cycles). Driveway camera runs at 1080p@5fps for better accuracy and range.

Automation Integration Frigate sends events via MQTT to Home Assistant, which handles:

• External AI analysis (e.g., is the person a mail carrier or a passerby?)

• WhatsApp messaging (with AI summary + snapshot)

• Smart device control (garage door, lights)

  

  Frigate Metrics view

Network Deployment

Dual 2.5GbE ports – one for home LAN, one dedicated to camera subnet for security and traffic isolation.

Thermal Validation

Tested in Sydney summer conditions. Passive cooling keeps CPU temperature steady at ~49°C, well within the 65°C target.

Outcomes and Benefits

Measured Performance

Metric	Value
Uptime	33 days (and growing)
Average CPU load	~25% across 8 cores
CPU temperature	~49°C
Object detection speed	~58ms average
NPU load	<20% (near zero during low motion)
SSD usage (33 days)	298GB / 1024GB (includes OS, Docker, Frigate, tools)
Power draw	Idle 2–3W / peak 8W

Achieved Outcomes

✅ LPR → garage door/lights in <2 seconds, high accuracy across day/night/rain

✅ Real-time detection and classification of people, cars, motorcycles, cats

✅ Courier identification + WhatsApp push notification (with AI summary + snapshot)

✅ Visual recognition of garage door and light status (full sun to nighttime)

✅ 8 cameras × 2 streams stable recording and retrieval

✅ Silent, passively cooled, stable under extreme heat

✅ Mobile video and snapshot review

Screenshot of Frigate NVR bird's-eye view showing four outdoor cameras on a rainy winter day. Note the recent activity bar at the top with a bounding box around the person detected by the NPU, and the small car icon in the top-left of the garage camera indicating vehicle recognition.

Additional Resources

A detailed build guide (hardware BOM, software config, driver tuning, thermal validation) is available on Medium for anyone who wants to build their own system.

The ArmSoM Sige7 delivers industrial-grade smart home NVR at 10W in 45°C Sydney heat – something x86 and Raspberry Pi cannot match.