One Camera Brain

One Camera Brain is software that turns the cameras a site already has into smart cameras: analog cameras behind a recorder, IP cameras, cloud cameras or a phone. It runs three packs of rules (theft and intrusion, safety, and work insights), grades what each camera is good enough for, seals a signed evidence clip for every alert, and can feed its results back into the site's existing camera software as a virtual camera. It was built to a startup brief, first as a small prototype and then as a much larger platform.

Other people's cameras. A rule such as "someone is in the restricted zone" is only as good as the detector underneath it, and off-the-shelf detectors miss people who are small, distant or packed into a crowd. Running the product on real footage found 22 problems, each fixed with a test that fails without the fix. A person standing still when the camera started was never detected. Small people on high-definition cameras were missed until the frame was also analyzed in tiles. Video from older recorders could not be read at all. And the brief asked for 98 to 100% accuracy, which the measurements support on some rules and plainly do not on others.

A platform of about 21,400 lines of Python with 10,200 lines of tests. A rules engine with the three packs. Connectors for ONVIF cameras, go2rtc streams and phone cameras. Perception: YOLOX and RF-DETR detectors run through ONNX Runtime, tiled detection for small people, a tracker, and a small vision-language model that double-checks an alert. An evidence ledger in which every clip is hash-sealed and signed, so a changed or missing clip shows. Ways out through MQTT, webhooks and the virtual camera. Around that, a node runtime, a local app for desktop and phone, and the layers for a site cluster and a cloud service. No model was trained for it: every detector is used as released.

Everything was measured on recorded public datasets, through the product's own code. After-hours intrusion: every event found on two street datasets (3 of 3 and 4 of 4), and 98% precision and recall on a shopping-corridor dataset. Restricted-zone entry on the corridor: 97.2% precision, 94.6% recall. Camera tampering: 42 of 42 detected. The detector takes 9.8 ms for a 1080p frame on an Apple M3 Ultra with tiling off.

The weak spots are reported by the project itself. In dense crowds the detector finds a third of the people (32.6% recall) and line-crossing recall is 31.8%. A person lying on the ground is recognized in 53% of frames. Night and fisheye scenes are worse than daylight ones. The overall score the project gives itself is 83.9 out of 100.

A test on a real DVR or NVR, which has not happened yet. A detector fine-tuned for crowds, night and fisheye views, because those are the measured gaps. A recorded demo, from live view to alert to sealed clip. Also not done: adapters for consumer cloud cameras are untested, and web push, talk-down audio, PTZ presets and tracking across cameras are not built.