From a missing accelerator to autonomous perception

The Pi booted, but the Hailo accelerator was missing. Firmware and driver checks established the software baseline without fixing the external PCIe link. A separate reversed PCIe ribbon was the physical fault. After shutdown, power removal, and correction, Hailo-8 enumeration and firmware communication passed.

A fresh 4608 × 2592 camera image confirmed that the camera still worked with the HAT installed. The accelerator completed 1,549 model-inference frames at 309.35 FPS. That number describes the accelerator benchmark, not the speed of the complete camera pipeline.

Live camera inference and recording worked, and a real-world person detection was verified. Object recognition also produced clear failures: a vertical bottle was repeatedly labeled a surfboard, and the horizontal bottle an airplane. A stable pipeline did not make those classifications correct.

We built a perception gate to turn detections into confirmed-presence and sustained-absence events. Its first version incorrectly let another object in the same frame reset the person-confirmation streak. Aggregating the frame before evaluating presence fixed that defect. Structured events were then written to the log.

The launcher and service passed manual operation, clean stop, camera release, and restart checks before boot enablement. After reboot, perception started without a manual launch and produced a confirmed-person event roughly two seconds after service activation. The observed boot had zero automatic service restarts; final temperature was 42.8°C with no throttling.

Open work remained. Camera angle interfered with reclined seating, spatial calibration was deferred, and general object accuracy was unresolved. The working live path also did not qualify the failing AI still-capture mode or long-duration recovery.