Asynchronous camera-to-inference relay with bounded queue backpressure. Engineered for 512MB RAM Raspberry Pi & Kneron-class embedded systems.
Explicit drop_newest (429 Too Many Requests) and drop_oldest policies prevent out-of-memory crashes on resource-constrained devices.
Decoupled capture and inference abstractions allow hot-swapping between YOLO v8 ONNX models and deterministic fixture adapters.
Buffers detection envelopes during connectivity outages and drains to cloud ingestion APIs automatically upon network restoration.