infinicode 2.8.23 → 2.8.25

package/packages/robopark/vision/app_pi_clean.py

@@ -50,6 +50,15 @@ webhook_url = None
 last_webhook_send_time = 0
 webhook_send_interval = 0.5
 
+# Vision-confirm (Ollama Cloud) config — stage 2 of the motion pipeline.
+# Motion detection (frame-diff) is a cheap pre-filter; before we fire the
+# session webhook we ask a vision model to confirm a person is actually in
+# frame, to cut down on false triggers from pets/shadows/wind.
+OLLAMA_CLOUD_API_KEY = os.getenv("OLLAMA_CLOUD_API_KEY", "")
+OLLAMA_CLOUD_VISION_MODEL = os.getenv("OLLAMA_CLOUD_VISION_MODEL", "gemma3:27b")
+OLLAMA_CLOUD_VISION_URL = "https://ollama.com/v1/chat/completions"
+VISION_CONFIRM_TIMEOUT = 8
+
 # Simple object detection using OpenCV DNN (MobileNet SSD)
 try:
     # Load pre-trained MobileNet SSD model
@@ -111,6 +120,66 @@ def detect_objects(frame, conf_threshold=0.5):
 
     return frame, detected_objects
 
+def confirm_person_present(frame):
+    """Ask an Ollama Cloud vision model whether a person is visible in `frame`.
+
+    This is stage 2 of the motion pipeline: motion detection (frame-diff) is a
+    cheap pre-filter, and this confirms a person is actually present before we
+    fire the session webhook — cuts down on false triggers from pets, shadows,
+    wind, etc.
+
+    Fails OPEN (returns True) on any error — missing key, network failure,
+    timeout, bad response — since the pre-existing motion-only trigger is the
+    fallback behavior and a vision-API outage shouldn't silently disable the
+    whole trigger system.
+    """
+    if not OLLAMA_CLOUD_API_KEY:
+        # No key configured: skip the check entirely, preserve motion-only
+        # behavior as the zero-config default. Caller logs this case.
+        return True
+
+    try:
+        _, buffer = cv2.imencode('.jpg', frame, [cv2.IMWRITE_JPEG_QUALITY, 80])
+        img_base64 = base64.b64encode(buffer).decode('utf-8')
+
+        payload = {
+            "model": OLLAMA_CLOUD_VISION_MODEL,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "text",
+                            "text": "Is there a person clearly visible in this image? Answer with only YES or NO."
+                        },
+                        {
+                            "type": "image_url",
+                            "image_url": {"url": f"data:image/jpeg;base64,{img_base64}"}
+                        }
+                    ]
+                }
+            ],
+            "stream": False
+        }
+
+        response = requests.post(
+            OLLAMA_CLOUD_VISION_URL,
+            json=payload,
+            headers={
+                "Authorization": f"Bearer {OLLAMA_CLOUD_API_KEY}",
+                "Content-Type": "application/json"
+            },
+            timeout=VISION_CONFIRM_TIMEOUT
+        )
+        response.raise_for_status()
+
+        answer = response.json()["choices"][0]["message"]["content"].strip()
+        return answer.upper().startswith("YES")
+    except Exception as e:
+        print(f"WARNING: vision-confirm error, failing open (treating as person present): {e}")
+        return True
+
+
 def send_webhook(frame_data):
     """Send frame to webhook URL"""
     global webhook_url, last_webhook_send_time
@@ -194,7 +263,18 @@ def generate_frames():
                     last_motion_time = time.time()
                     motion_frame_buffer = processed_frame.copy()
                     print(f"Motion detected!")
-                    send_webhook(processed_frame)
+
+                    # Only run the (network-bound) vision-confirm + webhook
+                    # once per motion "event" — reuse the same debounce timer
+                    # send_webhook() itself uses, rather than calling the
+                    # vision API on every single frame while motion continues.
+                    if webhook_url and (time.time() - last_webhook_send_time >= webhook_send_interval):
+                        if not OLLAMA_CLOUD_API_KEY:
+                            send_webhook(processed_frame)
+                        elif confirm_person_present(processed_frame):
+                            send_webhook(processed_frame)
+                        else:
+                            print("Motion event suppressed: vision-confirm found no person present")
                 else:
                     motion_detected_state = False