Update app.py

app.py

@@ -1,252 +1,252 @@
-import gradio as gr
-import cv2
-import numpy as np
-from ultralytics import YOLO
-from collections import deque
-import math
-
-# --- CONFIGURATION ---
-# Pastikan file model sudah diupload di satu folder yang sama dengan app.py
-MODEL_DET_PATH = "best.pt" 
-MODEL_POSE_PATH = "yolo11x-pose.pt" # Pertimbangkan ganti ke yolo11n-pose.pt jika lag
-
-# --- LOGIC CONFIGURATION ---
-CONF_THRESHOLD = 0.4
-HEAD_DROP_THRESHOLD = 110
-MOVEMENT_THRESHOLD = 50.0
-WRIST_CONF_THRESHOLD = 0.3
-EXTENSION_RATIO = 0.6
-PALM_BEND_ANGLE = 0
-WINDOW_SIZE = 10
-WINDOW_THRESHOLD = 6
-HAND_STREAK_REQUIRED = 2
-
-# Output Size untuk Web
-OUTPUT_WIDTH = 640
-OUTPUT_HEIGHT = 480
-OUTPUT_SIZE = (OUTPUT_WIDTH, OUTPUT_HEIGHT)
-
-# --- LOAD MODELS ---
-print("Loading Models...")
-try:
-    model_det = YOLO(MODEL_DET_PATH)
-    model_pose = YOLO(MODEL_POSE_PATH)
-    print("✅ Models Loaded Successfully!")
-except Exception as e:
-    print(f"❌ Error loading model: {e}")
-    # Kita tidak exit() agar Space tidak crash, tapi akan error saat run
-    pass
-
-# --- HELPERS (Sama seperti kode Anda) ---
-def is_point_in_box(point, box):
-    if box is None: return False
-    px, py = point
-    x1, y1, x2, y2 = box
-    return x1 < px < x2 and y1 < py < y2
-
-def rotate_vector(vx, vy, angle_degrees):
-    rad = math.radians(angle_degrees)
-    cos_a = math.cos(rad)
-    sin_a = math.sin(rad)
-    new_vx = vx * cos_a - vy * sin_a
-    new_vy = vx * sin_a + vy * cos_a
-    return new_vx, new_vy
-
-# --- LOGIC CLASS ---
-class DrowsinessDetector:
-    def __init__(self):
-        self.history = deque(maxlen=WINDOW_SIZE)
-        self.l_wrist_hist = deque(maxlen=WINDOW_SIZE)
-        self.r_wrist_hist = deque(maxlen=WINDOW_SIZE)
-        self.pose_hand_streak = 0
-        self.det_hand_streak = 0
-
-    def analyze_activity_level(self, wrist_history):
-        if len(wrist_history) < 5: return False
-        points = np.array([p for p in wrist_history if p is not None])
-        if len(points) < 5: return False
-        std_x = np.std(points[:, 0])
-        std_y = np.std(points[:, 1])
-        return ((std_x + std_y) / 2) > MOVEMENT_THRESHOLD
-
-    def get_pose_features(self, frame):
-        # ... (Logika pose Anda, disederhanakan untuk Gradio) ...
-        # Catatan: Karena kompleksitas kode asli, saya menyalin inti logikanya
-        # namun membuang bagian force_det untuk efisiensi di web.
-        
-        annotated_frame = frame.copy()
-        is_head_drop = False
-        is_hand_near_mouth = False
-        l_wrist_pt = None; r_wrist_pt = None; face_box = None
-        l_wrist_conf = None; r_wrist_conf = None
-        
-        H, W = frame.shape[:2]
-        
-        try:
-            res = model_pose.predict(frame, conf=0.5, verbose=False)[0]
-            # annotated_frame = res.plot() # Opsional: plot bawaan yolo
-            
-            # Ambil orang pertama saja
-            if hasattr(res.keypoints, "data") and len(res.keypoints.data) > 0:
-                kpts = res.keypoints.data[0].cpu().numpy()
-                
-                # Index keypoints COCO standard
-                # 0:Nose, 5:L-Shoulder, 6:R-Shoulder, 7:L-Elbow, 8:R-Elbow, 9:L-Wrist, 10:R-Wrist
-                nose = kpts[0][:2]
-                l_shoulder = kpts[5][:2]; r_shoulder = kpts[6][:2]
-                l_wrist = kpts[9][:2]; r_wrist = kpts[10][:2]
-                l_elbow = kpts[7][:2]; r_elbow = kpts[8][:2]
-                
-                l_wrist_conf = float(kpts[9][2]); r_wrist_conf = float(kpts[10][2])
-
-                # 1. Logic Perpanjangan Tangan (Sama seperti kode Anda)
-                if l_wrist[0] != 0 and l_elbow[0] != 0:
-                    vx = l_wrist[0] - l_elbow[0]; vy = l_wrist[1] - l_elbow[1]
-                    rvx, rvy = rotate_vector(vx, vy, PALM_BEND_ANGLE)
-                    l_wrist[0] += rvx * EXTENSION_RATIO; l_wrist[1] += rvy * EXTENSION_RATIO
-                
-                if r_wrist[0] != 0 and r_elbow[0] != 0:
-                    vx = r_wrist[0] - r_elbow[0]; vy = r_wrist[1] - r_elbow[1]
-                    rvx, rvy = rotate_vector(vx, vy, -PALM_BEND_ANGLE)
-                    r_wrist[0] += rvx * EXTENSION_RATIO; r_wrist[1] += rvy * EXTENSION_RATIO
-
-                l_wrist_pt = (float(l_wrist[0]), float(l_wrist[1]))
-                r_wrist_pt = (float(r_wrist[0]), float(r_wrist[1]))
-
-                # 2. Head Drop Logic
-                if nose[0] != 0 and (l_shoulder[0] != 0 or r_shoulder[0] != 0):
-                    sy = (l_shoulder[1] + r_shoulder[1]) / 2 if (l_shoulder[0]*r_shoulder[0])!=0 else max(l_shoulder[1], r_shoulder[1])
-                    dist = sy - nose[1]
-                    cv2.putText(annotated_frame, f"Head Dist: {int(dist)}", (10, 100), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255,255,0), 1)
-                    if dist < HEAD_DROP_THRESHOLD: is_head_drop = True
-
-                # 3. Face Box Est (Simple from keypoints)
-                # Ambil min/max dari mata, hidung, telinga untuk kotak wajah
-                face_pts = kpts[0:5, :2] # nose, eyes, ears
-                valid_pts = face_pts[face_pts[:,0] > 0]
-                if len(valid_pts) > 0:
-                    x1, y1 = valid_pts.min(axis=0); x2, y2 = valid_pts.max(axis=0)
-                    pad = 20
-                    face_box = [max(0, x1-pad), max(0, y1-pad), min(W, x2+pad), min(H, y2+pad + 50)] # +50 ke bawah dagu
-                    cv2.rectangle(annotated_frame, (int(face_box[0]), int(face_box[1])), (int(face_box[2]), int(face_box[3])), (255,255,0), 2)
-
-                # 4. Wrist near Mouth
-                if face_box:
-                    if l_wrist_conf > WRIST_CONF_THRESHOLD and is_point_in_box(l_wrist_pt, face_box): is_hand_near_mouth = True
-                    if r_wrist_conf > WRIST_CONF_THRESHOLD and is_point_in_box(r_wrist_pt, face_box): is_hand_near_mouth = True
-                    
-        except Exception as e:
-            print(f"Pose Error: {e}")
-            pass
-            
-        return is_head_drop, is_hand_near_mouth, annotated_frame, l_wrist_pt, r_wrist_pt, face_box, (l_wrist_conf, r_wrist_conf)
-
-    def process_frame(self, frame):
-        # Gradio pass RGB, opencv needs BGR
-        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
-        H, W = frame.shape[:2]
-
-        # 1. Detection
-        res_det = model_det.predict(frame, conf=CONF_THRESHOLD, verbose=False)[0]
-        
-        # 2. Pose
-        is_head_drop, is_hand_in_box, pose_frame, lw, rw, face_box, wrist_confs = self.get_pose_features(frame)
-
-        # 3. Det Logic (Simple Overlap check)
-        det_hand_mouth_near = False
-        try:
-            for box, cls in zip(res_det.boxes.xyxy, res_det.boxes.cls):
-                if model_det.names[int(cls)] == "Hand_Mouth" and face_box:
-                     # Simple check: box center in face box
-                     bx1, by1, bx2, by2 = box.cpu().numpy()
-                     cx, cy = (bx1+bx2)/2, (by1+by2)/2
-                     if is_point_in_box((cx, cy), face_box):
-                         det_hand_mouth_near = True
-        except: pass
-
-        # 4. Streak
-        if is_hand_in_box: self.pose_hand_streak += 1
-        else: self.pose_hand_streak = 0
-        
-        if det_hand_mouth_near: self.det_hand_streak += 1
-        else: self.det_hand_streak = 0
-
-        flag_hand_mouth = (self.pose_hand_streak >= HAND_STREAK_REQUIRED) or (self.det_hand_streak >= HAND_STREAK_REQUIRED)
-
-        # 5. Activity
-        self.l_wrist_hist.append(lw if (wrist_confs[0] or 0) >= WRIST_CONF_THRESHOLD else None)
-        self.r_wrist_hist.append(rw if (wrist_confs[1] or 0) >= WRIST_CONF_THRESHOLD else None)
-        
-        is_active = self.analyze_activity_level(self.l_wrist_hist) or self.analyze_activity_level(self.r_wrist_hist)
-
-        # 6. Combined Logic (Simplified)
-        labels_set = set([model_det.names[int(c)] for c in res_det.boxes.cls.cpu().numpy()]) if hasattr(res_det.boxes, 'cls') else set()
-        
-        # Instant Drowsy Check
-        is_drowsy_frame = False
-        if not flag_hand_mouth:
-            if "Closed_Eyes" in labels_set and "Open_Mouth" in labels_set: is_drowsy_frame = True
-            elif "Closed_Eyes" in labels_set and "Closed_Mouth" in labels_set: is_drowsy_frame = True
-            elif is_head_drop: is_drowsy_frame = True
-
-        # 7. Sliding Window Logic
-        if is_active or flag_hand_mouth:
-             self.history.append(0)
-        else:
-             self.history.append(1 if is_drowsy_frame else 0)
-
-        drowsy_score = sum(self.history)
-        is_drowsy_window = (drowsy_score >= WINDOW_THRESHOLD)
-
-        # 8. Visualization
-        status_text = "AMAN"
-        status_color = (0, 255, 0)
-        
-        if flag_hand_mouth:
-            status_text = "BAHAYA! (Hand on Face)"
-            status_color = (0, 0, 255)
-        elif is_drowsy_window:
-            status_text = "BAHAYA! (Mengantuk)"
-            status_color = (0, 0, 255)
-        elif is_active:
-            status_text = "AMAN (Aktif)"
-            status_color = (0, 255, 0)
-
-        # Draw HUD
-        annotated = res_det.plot() # Gambar kotak deteksi
-        
-        # Overlay Pose Lines
-        if face_box:
-            cv2.rectangle(annotated, (int(face_box[0]), int(face_box[1])), (int(face_box[2]), int(face_box[3])), (255,255,0), 2)
-        
-        if lw: cv2.circle(annotated, (int(lw[0]), int(lw[1])), 5, (0,255,255), -1)
-        if rw: cv2.circle(annotated, (int(rw[0]), int(rw[1])), 5, (0,255,255), -1)
-
-        # Top Bar
-        cv2.rectangle(annotated, (0, 0), (annotated.shape[1], 80), (0, 0, 0), -1)
-        cv2.putText(annotated, f"STATUS: {status_text}", (20, 35), 0, 0.8, status_color, 2)
-        cv2.putText(annotated, f"Score: {drowsy_score}/{WINDOW_SIZE} | Act: {is_active}", (20, 65), 0, 0.6, (200, 200, 200), 1)
-
-        return cv2.cvtColor(annotated, cv2.COLOR_BGR2RGB)
-
-# --- GRADIO INTERFACE ---
-detector = DrowsinessDetector()
-
-def inference(frame):
-    if frame is None: return None
-    return detector.process_frame(frame)
-
-if __name__ == "__main__":
-    with gr.Blocks() as demo:
-        gr.Markdown("# 😴 AI Drowsiness Detector (YOLOv11)")
-        gr.Markdown("Pastikan webcam aktif. Deteksi mengantuk berbasis Pose & Object Detection.")
-        
-        with gr.Row():
-            # Input webcam, streaming=True agar real-time
-            input_video = gr.Image(source="webcam", streaming=True, label="Webcam Input", tool=None)
-            output_video = gr.Image(label="AI Output")
-            
-        # Saat webcam berubah (tiap frame), panggil fungsi inference
-        input_video.change(fn=inference, inputs=input_video, outputs=output_video)
-
+import gradio as gr
+import cv2
+import numpy as np
+from ultralytics import YOLO
+from collections import deque
+import math
+
+# --- CONFIGURATION ---
+# Pastikan file model sudah diupload di satu folder yang sama dengan app.py
+MODEL_DET_PATH = "best.pt" 
+MODEL_POSE_PATH = "yolo11x-pose.pt" # Pertimbangkan ganti ke yolo11n-pose.pt jika lag
+
+# --- LOGIC CONFIGURATION ---
+CONF_THRESHOLD = 0.4
+HEAD_DROP_THRESHOLD = 110
+MOVEMENT_THRESHOLD = 50.0
+WRIST_CONF_THRESHOLD = 0.3
+EXTENSION_RATIO = 0.6
+PALM_BEND_ANGLE = 0
+WINDOW_SIZE = 10
+WINDOW_THRESHOLD = 6
+HAND_STREAK_REQUIRED = 2
+
+# Output Size untuk Web
+OUTPUT_WIDTH = 640
+OUTPUT_HEIGHT = 480
+OUTPUT_SIZE = (OUTPUT_WIDTH, OUTPUT_HEIGHT)
+
+# --- LOAD MODELS ---
+print("Loading Models...")
+try:
+    model_det = YOLO(MODEL_DET_PATH)
+    model_pose = YOLO(MODEL_POSE_PATH)
+    print("✅ Models Loaded Successfully!")
+except Exception as e:
+    print(f"❌ Error loading model: {e}")
+    # Kita tidak exit() agar Space tidak crash, tapi akan error saat run
+    pass
+
+# --- HELPERS (Sama seperti kode Anda) ---
+def is_point_in_box(point, box):
+    if box is None: return False
+    px, py = point
+    x1, y1, x2, y2 = box
+    return x1 < px < x2 and y1 < py < y2
+
+def rotate_vector(vx, vy, angle_degrees):
+    rad = math.radians(angle_degrees)
+    cos_a = math.cos(rad)
+    sin_a = math.sin(rad)
+    new_vx = vx * cos_a - vy * sin_a
+    new_vy = vx * sin_a + vy * cos_a
+    return new_vx, new_vy
+
+# --- LOGIC CLASS ---
+class DrowsinessDetector:
+    def __init__(self):
+        self.history = deque(maxlen=WINDOW_SIZE)
+        self.l_wrist_hist = deque(maxlen=WINDOW_SIZE)
+        self.r_wrist_hist = deque(maxlen=WINDOW_SIZE)
+        self.pose_hand_streak = 0
+        self.det_hand_streak = 0
+
+    def analyze_activity_level(self, wrist_history):
+        if len(wrist_history) < 5: return False
+        points = np.array([p for p in wrist_history if p is not None])
+        if len(points) < 5: return False
+        std_x = np.std(points[:, 0])
+        std_y = np.std(points[:, 1])
+        return ((std_x + std_y) / 2) > MOVEMENT_THRESHOLD
+
+    def get_pose_features(self, frame):
+        # ... (Logika pose Anda, disederhanakan untuk Gradio) ...
+        # Catatan: Karena kompleksitas kode asli, saya menyalin inti logikanya
+        # namun membuang bagian force_det untuk efisiensi di web.
+        
+        annotated_frame = frame.copy()
+        is_head_drop = False
+        is_hand_near_mouth = False
+        l_wrist_pt = None; r_wrist_pt = None; face_box = None
+        l_wrist_conf = None; r_wrist_conf = None
+        
+        H, W = frame.shape[:2]
+        
+        try:
+            res = model_pose.predict(frame, conf=0.5, verbose=False)[0]
+            # annotated_frame = res.plot() # Opsional: plot bawaan yolo
+            
+            # Ambil orang pertama saja
+            if hasattr(res.keypoints, "data") and len(res.keypoints.data) > 0:
+                kpts = res.keypoints.data[0].cpu().numpy()
+                
+                # Index keypoints COCO standard
+                # 0:Nose, 5:L-Shoulder, 6:R-Shoulder, 7:L-Elbow, 8:R-Elbow, 9:L-Wrist, 10:R-Wrist
+                nose = kpts[0][:2]
+                l_shoulder = kpts[5][:2]; r_shoulder = kpts[6][:2]
+                l_wrist = kpts[9][:2]; r_wrist = kpts[10][:2]
+                l_elbow = kpts[7][:2]; r_elbow = kpts[8][:2]
+                
+                l_wrist_conf = float(kpts[9][2]); r_wrist_conf = float(kpts[10][2])
+
+                # 1. Logic Perpanjangan Tangan (Sama seperti kode Anda)
+                if l_wrist[0] != 0 and l_elbow[0] != 0:
+                    vx = l_wrist[0] - l_elbow[0]; vy = l_wrist[1] - l_elbow[1]
+                    rvx, rvy = rotate_vector(vx, vy, PALM_BEND_ANGLE)
+                    l_wrist[0] += rvx * EXTENSION_RATIO; l_wrist[1] += rvy * EXTENSION_RATIO
+                
+                if r_wrist[0] != 0 and r_elbow[0] != 0:
+                    vx = r_wrist[0] - r_elbow[0]; vy = r_wrist[1] - r_elbow[1]
+                    rvx, rvy = rotate_vector(vx, vy, -PALM_BEND_ANGLE)
+                    r_wrist[0] += rvx * EXTENSION_RATIO; r_wrist[1] += rvy * EXTENSION_RATIO
+
+                l_wrist_pt = (float(l_wrist[0]), float(l_wrist[1]))
+                r_wrist_pt = (float(r_wrist[0]), float(r_wrist[1]))
+
+                # 2. Head Drop Logic
+                if nose[0] != 0 and (l_shoulder[0] != 0 or r_shoulder[0] != 0):
+                    sy = (l_shoulder[1] + r_shoulder[1]) / 2 if (l_shoulder[0]*r_shoulder[0])!=0 else max(l_shoulder[1], r_shoulder[1])
+                    dist = sy - nose[1]
+                    cv2.putText(annotated_frame, f"Head Dist: {int(dist)}", (10, 100), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255,255,0), 1)
+                    if dist < HEAD_DROP_THRESHOLD: is_head_drop = True
+
+                # 3. Face Box Est (Simple from keypoints)
+                # Ambil min/max dari mata, hidung, telinga untuk kotak wajah
+                face_pts = kpts[0:5, :2] # nose, eyes, ears
+                valid_pts = face_pts[face_pts[:,0] > 0]
+                if len(valid_pts) > 0:
+                    x1, y1 = valid_pts.min(axis=0); x2, y2 = valid_pts.max(axis=0)
+                    pad = 20
+                    face_box = [max(0, x1-pad), max(0, y1-pad), min(W, x2+pad), min(H, y2+pad + 50)] # +50 ke bawah dagu
+                    cv2.rectangle(annotated_frame, (int(face_box[0]), int(face_box[1])), (int(face_box[2]), int(face_box[3])), (255,255,0), 2)
+
+                # 4. Wrist near Mouth
+                if face_box:
+                    if l_wrist_conf > WRIST_CONF_THRESHOLD and is_point_in_box(l_wrist_pt, face_box): is_hand_near_mouth = True
+                    if r_wrist_conf > WRIST_CONF_THRESHOLD and is_point_in_box(r_wrist_pt, face_box): is_hand_near_mouth = True
+                    
+        except Exception as e:
+            print(f"Pose Error: {e}")
+            pass
+            
+        return is_head_drop, is_hand_near_mouth, annotated_frame, l_wrist_pt, r_wrist_pt, face_box, (l_wrist_conf, r_wrist_conf)
+
+    def process_frame(self, frame):
+        # Gradio pass RGB, opencv needs BGR
+        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+        H, W = frame.shape[:2]
+
+        # 1. Detection
+        res_det = model_det.predict(frame, conf=CONF_THRESHOLD, verbose=False)[0]
+        
+        # 2. Pose
+        is_head_drop, is_hand_in_box, pose_frame, lw, rw, face_box, wrist_confs = self.get_pose_features(frame)
+
+        # 3. Det Logic (Simple Overlap check)
+        det_hand_mouth_near = False
+        try:
+            for box, cls in zip(res_det.boxes.xyxy, res_det.boxes.cls):
+                if model_det.names[int(cls)] == "Hand_Mouth" and face_box:
+                     # Simple check: box center in face box
+                     bx1, by1, bx2, by2 = box.cpu().numpy()
+                     cx, cy = (bx1+bx2)/2, (by1+by2)/2
+                     if is_point_in_box((cx, cy), face_box):
+                         det_hand_mouth_near = True
+        except: pass
+
+        # 4. Streak
+        if is_hand_in_box: self.pose_hand_streak += 1
+        else: self.pose_hand_streak = 0
+        
+        if det_hand_mouth_near: self.det_hand_streak += 1
+        else: self.det_hand_streak = 0
+
+        flag_hand_mouth = (self.pose_hand_streak >= HAND_STREAK_REQUIRED) or (self.det_hand_streak >= HAND_STREAK_REQUIRED)
+
+        # 5. Activity
+        self.l_wrist_hist.append(lw if (wrist_confs[0] or 0) >= WRIST_CONF_THRESHOLD else None)
+        self.r_wrist_hist.append(rw if (wrist_confs[1] or 0) >= WRIST_CONF_THRESHOLD else None)
+        
+        is_active = self.analyze_activity_level(self.l_wrist_hist) or self.analyze_activity_level(self.r_wrist_hist)
+
+        # 6. Combined Logic (Simplified)
+        labels_set = set([model_det.names[int(c)] for c in res_det.boxes.cls.cpu().numpy()]) if hasattr(res_det.boxes, 'cls') else set()
+        
+        # Instant Drowsy Check
+        is_drowsy_frame = False
+        if not flag_hand_mouth:
+            if "Closed_Eyes" in labels_set and "Open_Mouth" in labels_set: is_drowsy_frame = True
+            elif "Closed_Eyes" in labels_set and "Closed_Mouth" in labels_set: is_drowsy_frame = True
+            elif is_head_drop: is_drowsy_frame = True
+
+        # 7. Sliding Window Logic
+        if is_active or flag_hand_mouth:
+             self.history.append(0)
+        else:
+             self.history.append(1 if is_drowsy_frame else 0)
+
+        drowsy_score = sum(self.history)
+        is_drowsy_window = (drowsy_score >= WINDOW_THRESHOLD)
+
+        # 8. Visualization
+        status_text = "AMAN"
+        status_color = (0, 255, 0)
+        
+        if flag_hand_mouth:
+            status_text = "BAHAYA! (Hand on Face)"
+            status_color = (0, 0, 255)
+        elif is_drowsy_window:
+            status_text = "BAHAYA! (Mengantuk)"
+            status_color = (0, 0, 255)
+        elif is_active:
+            status_text = "AMAN (Aktif)"
+            status_color = (0, 255, 0)
+
+        # Draw HUD
+        annotated = res_det.plot() # Gambar kotak deteksi
+        
+        # Overlay Pose Lines
+        if face_box:
+            cv2.rectangle(annotated, (int(face_box[0]), int(face_box[1])), (int(face_box[2]), int(face_box[3])), (255,255,0), 2)
+        
+        if lw: cv2.circle(annotated, (int(lw[0]), int(lw[1])), 5, (0,255,255), -1)
+        if rw: cv2.circle(annotated, (int(rw[0]), int(rw[1])), 5, (0,255,255), -1)
+
+        # Top Bar
+        cv2.rectangle(annotated, (0, 0), (annotated.shape[1], 80), (0, 0, 0), -1)
+        cv2.putText(annotated, f"STATUS: {status_text}", (20, 35), 0, 0.8, status_color, 2)
+        cv2.putText(annotated, f"Score: {drowsy_score}/{WINDOW_SIZE} | Act: {is_active}", (20, 65), 0, 0.6, (200, 200, 200), 1)
+
+        return cv2.cvtColor(annotated, cv2.COLOR_BGR2RGB)
+
+# --- GRADIO INTERFACE ---
+detector = DrowsinessDetector()
+
+def inference(frame):
+    if frame is None: return None
+    return detector.process_frame(frame)
+
+if __name__ == "__main__":
+    with gr.Blocks() as demo:
+        gr.Markdown("# 😴 AI Drowsiness Detector (YOLOv11)")
+        gr.Markdown("Pastikan webcam aktif. Deteksi mengantuk berbasis Pose & Object Detection.")
+        
+        with gr.Row():
+            # Input webcam, streaming=True agar real-time
+            input_video = gr.Image(sources=["webcam"], streaming=True, label="Webcam Input")
+            output_video = gr.Image(label="AI Output")
+            
+        # Saat webcam berubah (tiap frame), panggil fungsi inference
+        input_video.change(fn=inference, inputs=input_video, outputs=output_video)
+
     demo.launch()