Security Pipeline (Multi-camera, Anonymising)

import os
import numpy as np
from cyberwave.data import FILTERED_FRAME_CHANNEL
from cyberwave.vision import anonymize_frame

CAMERA_1 = os.environ["CAMERA_1_TWIN"]
CAMERA_2 = os.environ["CAMERA_2_TWIN"]

model = cw.models.load("yolov8n")

def _anonymise(frame, ctx):
    result = model.predict(frame, classes=["person"], confidence=0.4)
    # Privacy fail-closed: ``anonymize_frame`` only obscures bbox
    # regions for matching detections. On a frame where the model
    # missed everyone (sub-threshold, occluded, partial body) it
    # would otherwise return the raw frame unchanged and the driver
    # would substitute that into WebRTC. Publish a black frame
    # instead — matches the driver's fail-closed behaviour for
    # stale and shape-mismatched filtered frames.
    if any(d.label == "person" for d in result.detections):
        out = anonymize_frame(
            frame,
            result.detections,
            mode="pixelate",
            draw_skeleton=False,
        )
    else:
        out = np.zeros_like(frame)
    cw.data.publish(FILTERED_FRAME_CHANNEL, out, twin_uuid=ctx.twin_uuid)
    for det in result.detections:
        if det.area_ratio > 0.3:
            cw.publish_event(ctx.twin_uuid, "person_too_close", {
                "area_ratio": round(det.area_ratio, 3),
                "frame_ts": ctx.timestamp,
            })

@cw.on_frame(CAMERA_1, sensor="default")
def on_camera_1_frame(frame, ctx): _anonymise(frame, ctx)

@cw.on_frame(CAMERA_2, sensor="default")
def on_camera_2_frame(frame, ctx): _anonymise(frame, ctx)

CYBERWAVE_METADATA_FRAME_FILTER_ENABLED=true
# Optional: widen the freshness window for CPU-only workers
# (default 200 ms is tuned for >= 5 Hz GPU workers).
CYBERWAVE_METADATA_FRAME_FILTER_FRESHNESS_MS=400