stop trusting cloud cameras!! (here's what I use instead)

Self-hosted Frigate is positioned as a privacy-first alternative to cloud-connected security cameras—because it keeps video processing local and avoids exposing camera feeds to third parties or dark-web resale. The pitch is blunt: if surveillance cameras connect to the cloud, someone could potentially watch and monetize those streams. Frigate counters that risk with an “in-house” setup that runs an open-source AI surveillance stack on your own hardware, adding facial recognition, license plate recognition, object detection, and semantic search—while integrating with Home Assistant for automation.

The practical walkthrough starts with the minimum building blocks: a local server (anything that can run Docker, such as a Raspberry Pi or a spare laptop), IP cameras that support RTSP (Real-Time Streaming Protocol), and a local network connection. The creator demonstrates a “quick and dirty” build on a Raspberry Pi using Docker Compose to run Frigate, then verifies the system is healthy via the Frigate web UI. The key technical requirement is RTSP streaming; once RTSP is enabled on the camera, Frigate can ingest the stream, record events, and run AI detections.

A major implementation detail is using dual RTSP streams when available. The setup distinguishes between a lower-quality “detect” stream (optimized for AI inference) and a higher-quality “record” stream (optimized for evidence and review). Motion-based recording is enabled so the system stores footage only when movement is detected, reducing storage and bandwidth pressure.

As the camera count grows, performance bottlenecks appear—especially on CPU-only setups. With multiple cameras, detector inference speed and CPU usage climb sharply, and the system becomes harder to sustain on a small device. To address this, the walkthrough adds AI accelerators. A Raspberry Pi AI hat (Halo) dramatically reduces CPU load and improves inference speed, while a Google Coral USB AI accelerator offloads AI inference so the main machine can focus on video decoding and streaming. With Coral, the system scales to around 10 cameras, running face detection and other AI features while keeping GPU utilization focused on decoding.

The most consequential part of the experience isn’t the AI—it’s the network. After roughly 12 hours, the Wi‑Fi network degrades: pages stop loading, cameras behave erratically, and the Frigate container eventually needs restarting. Troubleshooting rules out raw bandwidth as the primary cause by checking the unified network controller metrics. Instead, the culprit is “airtime” and retransmissions: a high TX retry rate on one access point indicates packets frequently fail on the first attempt, creating escalating retransmit traffic that fills buffers over time. The creator also notes that Rio Link E one Pro cameras can have RTSP stream degradation after long uptime.

Three fixes stabilize the system: (1) tune camera stream behavior by enabling constant bit rate (CBR/“fluency”) and using TCP instead of UDP, (2) use Go2RTC so multiple devices don’t each open separate camera connections—centralizing stream handling through Frigate, and (3) reboot problematic cameras on a staggered schedule (about once per 24 hours) to avoid RTSP degradation. Finally, adding an extra access point reduces airtime contention by distributing cameras across APs. After these changes, Frigate runs smoothly again, and the system remains local—integrated with Home Assistant—without relying on cloud camera feeds.