Update app.py

app.py

@@ -1,774 +1,481 @@
 import os
 import sys
-import subprocess
 import logging
-import warnings
+import time
 import cv2
-import gradio as gr
-import torch
 import numpy as np
+import torch
+import gradio as gr
 from ultralytics import YOLO
-import time
-from simple_salesforce import Salesforce
-from reportlab.lib.pagesizes import letter
-from reportlab.pdfgen import canvas
-from reportlab.lib.units import inch
-from io import BytesIO
-import base64
-from retrying import retry
-import uuid
-from multiprocessing import Pool, cpu_count
-from functools import partial
-
-# ========================== # Configuration and Setup # ==========================
-os.environ['YOLO_CONFIG_DIR'] = '/tmp/Ultralytics'
-os.makedirs('/tmp/Ultralytics', exist_ok=True)
+from collections import defaultdict
+from typing import List, Dict, Tuple, Optional
+
+# ========================== # Configuration # ==========================
+class Config:
+    # Model settings
+    MODEL_PATH = "yolov8_safety.pt"
+    FALLBACK_MODEL = "yolov8n.pt"
+    DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+    
+    # Violation settings
+    VIOLATION_LABELS = {
+        0: "no_helmet",
+        1: "no_harness",
+        2: "unsafe_posture",
+        3: "unsafe_zone",
+        4: "improper_tool_use"
+    }
+    
+    # Tracking settings
+    FACE_TRACKING = True  # Enable face tracking for helmet violations
+    TRACK_BUFFER = 30  # Number of frames to keep track of a person
+    MIN_FACE_CONFIDENCE = 0.7  # Minimum confidence for face detection
+    MIN_VIOLATION_CONFIDENCE = 0.5  # Minimum confidence for violation detection
+    
+    # Output settings
+    OUTPUT_DIR = "static/output"
+    SNAPSHOT_QUALITY = 90
+    FRAME_SKIP = 2  # Process every nth frame for efficiency
+    
+    # Violation suppression
+    VIOLATION_COOLDOWN = 5.0  # Seconds before same violation can be reported again
+    POSITION_THRESHOLD = 50  # Pixel distance to consider same location
+    
+    # Display colors
+    COLOR_MAP = {
+        "no_helmet": (0, 0, 255),
+        "no_harness": (0, 165, 255),
+        "unsafe_posture": (0, 255, 0),
+        "unsafe_zone": (255, 0, 0),
+        "improper_tool_use": (255, 255, 0)
+    }
+    
+    # Penalty scores for safety calculation
+    PENALTIES = {
+        "no_helmet": 25,
+        "no_harness": 30,
+        "unsafe_posture": 20,
+        "unsafe_zone": 35,
+        "improper_tool_use": 25
+    }
 
+# Initialize logging
 logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
 logger = logging.getLogger(__name__)
 
-# ========================== # ByteTrack Implementation # ==========================
-class BYTETracker:
-    def __init__(self, track_thresh=0.3, track_buffer=30, match_thresh=0.7, frame_rate=30):
-        self.track_thresh = track_thresh
-        self.track_buffer = track_buffer
-        self.match_thresh = match_thresh
-        self.frame_rate = frame_rate
-        self.next_id = 1
-        self.tracks = {}  # Store active tracks
-        self.worker_history = {}  # Track worker positions over time
-        self.last_positions = {}  # Last known positions of workers
-
-    def update(self, dets, scores, cls):
-        tracks = []
-        current_time = time.time()
-        
-        # Update existing tracks with new detections
-        for i, (det, score, cl) in enumerate(zip(dets, scores, cls)):
-            if score < self.track_thresh:
-                continue
-                
-            x, y, w, h = det
-            matched = False
-            best_iou = 0
-            best_track_id = None
-            
-            # Try to match with existing tracks
-            for track_id, track_info in self.tracks.items():
-                if current_time - track_info['last_seen'] > self.track_buffer / self.frame_rate:
-                    continue
-                    
-                tx, ty, tw, th = track_info['bbox']
-                iou = self._calculate_iou([x, y, w, h], [tx, ty, tw, th])
-                
-                if iou > self.match_thresh and iou > best_iou:
-                    best_iou = iou
-                    best_track_id = track_id
-                    matched = True
-            
-            if matched:
-                # Update existing track
-                self.tracks[best_track_id].update({
-                    'bbox': [x, y, w, h],
-                    'score': score,
-                    'cls': cl,
-                    'last_seen': current_time
-                })
-                
-                # Update position history
-                if best_track_id not in self.worker_history:
-                    self.worker_history[best_track_id] = []
-                self.worker_history[best_track_id].append([x, y])
-                self.last_positions[best_track_id] = [x, y]
-                
-                tracks.append({
-                    'id': best_track_id,
-                    'bbox': [x, y, w, h],
-                    'score': score,
-                    'cls': cl
-                })
-            else:
-                # Create new track
-                # Check if this detection might be the same worker from a different angle
-                same_worker = False
-                for worker_id, last_pos in self.last_positions.items():
-                    if self._is_same_worker([x, y], last_pos):
-                        self.tracks[worker_id] = {
-                            'bbox': [x, y, w, h],
-                            'score': score,
-                            'cls': cl,
-                            'last_seen': current_time
-                        }
-                        tracks.append({
-                            'id': worker_id,
-                            'bbox': [x, y, w, h],
-                            'score': score,
-                            'cls': cl
-                        })
-                        same_worker = True
-                        break
-                
-                if not same_worker:
-                    self.tracks[self.next_id] = {
-                        'bbox': [x, y, w, h],
-                        'score': score,
-                        'cls': cl,
-                        'last_seen': current_time
-                    }
-                    self.worker_history[self.next_id] = [[x, y]]
-                    self.last_positions[self.next_id] = [x, y]
-                    tracks.append({
-                        'id': self.next_id,
-                        'bbox': [x, y, w, h],
-                        'score': score,
-                        'cls': cl
-                    })
-                    self.next_id += 1
-        
-        # Clean up old tracks
-        current_time = time.time()
-        stale_ids = []
-        for track_id, track_info in self.tracks.items():
-            if current_time - track_info['last_seen'] > self.track_buffer / self.frame_rate:
-                stale_ids.append(track_id)
-        
-        for track_id in stale_ids:
-            del self.tracks[track_id]
-            if track_id in self.worker_history:
-                del self.worker_history[track_id]
-            if track_id in self.last_positions:
-                del self.last_positions[track_id]
+# ========================== # Face Recognition # ==========================
+class FaceTracker:
+    def __init__(self):
+        try:
+            # Try to load face detection model
+            self.face_cascade = cv2.CascadeClassifier(
+                cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
+            self.enabled = True
+        except:
+            logger.warning("Face detection model not available - falling back to position tracking")
+            self.enabled = False
+    
+    def detect_faces(self, frame: np.ndarray) -> List[Tuple[int, int, int, int]]:
+        """Detect faces in a frame and return bounding boxes"""
+        if not self.enabled:
+            return []
             
-        return tracks
+        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+        faces = self.face_cascade.detectMultiScale(
+            gray,
+            scaleFactor=1.1,
+            minNeighbors=5,
+            minSize=(30, 30)
+        return faces
+
+# ========================== # Violation Tracker # ==========================
+class ViolationTracker:
+    def __init__(self):
+        self.tracked_workers = {}  # {track_id: {last_seen, violations, face_bbox, position}}
+        self.violation_history = defaultdict(list)  # {violation_type: [detection_times]}
+        self.next_id = 1
+        self.face_tracker = FaceTracker() if Config.FACE_TRACKING else None
     
-    def _calculate_iou(self, box1, box2):
-        """Calculate IOU between two boxes"""
+    def _calculate_iou(self, box1: Tuple, box2: Tuple) -> float:
+        """Calculate Intersection over Union for two bounding boxes"""
         x1, y1, w1, h1 = box1
         x2, y2, w2, h2 = box2
         
-        # Calculate intersection coordinates
-        x_left = max(x1 - w1/2, x2 - w2/2)
-        y_top = max(y1 - h1/2, y2 - h2/2)
-        x_right = min(x1 + w1/2, x2 + w2/2)
-        y_bottom = min(y1 + h1/2, y2 + h2/2)
+        # Calculate coordinates
+        x_left = max(x1, x2)
+        y_top = max(y1, y2)
+        x_right = min(x1 + w1, x2 + w2)
+        y_bottom = min(y1 + h1, y2 + h2)
         
         if x_right < x_left or y_bottom < y_top:
             return 0.0
             
         intersection_area = (x_right - x_left) * (y_bottom - y_top)
-        
         box1_area = w1 * h1
         box2_area = w2 * h2
         
-        iou = intersection_area / (box1_area + box2_area - intersection_area)
-        return iou
+        return intersection_area / (box1_area + box2_area - intersection_area)
     
-    def _is_same_worker(self, pos1, pos2, threshold=100):
-        """Check if two positions likely belong to the same worker"""
+    def _is_same_position(self, pos1: Tuple, pos2: Tuple) -> bool:
+        """Check if two positions are close enough to be considered the same"""
         x1, y1 = pos1
         x2, y2 = pos2
         distance = np.sqrt((x1 - x2)**2 + (y1 - y2)**2)
-        return distance < threshold
-
-# ========================== # Optimized Configuration # ==========================
-CONFIG = {
-    "MODEL_PATH": "yolov8_safety.pt",
-    "FALLBACK_MODEL": "yolov8n.pt",
-    "OUTPUT_DIR": "static/output",
-    "VIOLATION_LABELS": {
-        0: "no_helmet",
-        1: "no_harness",
-        2: "unsafe_posture",
-        3: "unsafe_zone",
-        4: "improper_tool_use"
-    },
-    "CLASS_COLORS": {
-        "no_helmet": (0, 0, 255),          # Red
-        "no_harness": (0, 165, 255),       # Orange
-        "unsafe_posture": (0, 255, 0),     # Green
-        "unsafe_zone": (255, 0, 0),        # Blue
-        "improper_tool_use": (255, 255, 0) # Cyan
-    },
-    "DISPLAY_NAMES": {
-        "no_helmet": "No Helmet Violation",
-        "no_harness": "No Harness Violation",
-        "unsafe_posture": "Unsafe Posture",
-        "unsafe_zone": "Unsafe Zone Entry",
-        "improper_tool_use": "Improper Tool Use"
-    },
-    "SF_CREDENTIALS": {
-        "username": "prashanth1ai@safety.com",
-        "password": "SaiPrash461",
-        "security_token": "AP4AQnPoidIKPvSvNEfAHyoK",
-        "domain": "login"
-    },
-    "PUBLIC_URL_BASE": "https://huggingface.co/spaces/PrashanthB461/AI_Safety_Demo2/resolve/main/static/output/",
-    "CONFIDENCE_THRESHOLDS": {
-        "no_helmet": 0.5,
-        "no_harness": 0.3,
-        "unsafe_posture": 0.3,
-        "unsafe_zone": 0.3,
-        "improper_tool_use": 0.3
-    },
-    "MIN_VIOLATION_FRAMES": 1,
-    "VIOLATION_COOLDOWN": 30.0,  # Increased cooldown period
-    "WORKER_TRACKING_DURATION": 5.0,
-    "MAX_PROCESSING_TIME": 60,
-    "FRAME_SKIP": 2,  # Skip more frames for faster processing
-    "BATCH_SIZE": 16,
-    "PARALLEL_WORKERS": max(1, cpu_count() - 1),
-    "TRACK_BUFFER": 30,
-    "TRACK_THRESH": 0.3,
-    "MATCH_THRESH": 0.7,
-    "SNAPSHOT_QUALITY": 95,  # Higher quality for better visibility
-    "MAX_WORKER_DISTANCE": 100  # Maximum pixel distance to consider same worker
-}
-
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-logger.info(f"Using device: {device}")
-
-def load_model():
-    try:
-        if os.path.isfile(CONFIG["MODEL_PATH"]):
-            model_path = CONFIG["MODEL_PATH"]
-            logger.info(f"Model loaded: {model_path}")
-        else:
-            model_path = CONFIG["FALLBACK_MODEL"]
-            logger.warning("Using fallback model. Train yolov8_safety.pt for best results.")
-            if not os.path.isfile(model_path):
-                logger.info(f"Downloading fallback model: {model_path}")
-                torch.hub.download_url_to_file('https://github.com/ultralytics/assets/releases/download/v8.3.0/yolov8n.pt', model_path)
-                
-        model = YOLO(model_path).to(device)
-        logger.info(f"Model classes: {model.names}")
-        return model
-    except Exception as e:
-        logger.error(f"Failed to load model: {e}")
-        raise
-
-model = load_model()
-
-# ========================== # Helper Functions # ==========================
-def preprocess_frame(frame):
-    """Apply basic preprocessing to enhance detection"""
-    frame = cv2.convertScaleAbs(frame, alpha=1.2, beta=20)
-    return frame
-
-def draw_detections(frame, detections):
-    """Draw bounding boxes and labels on detection frame with improved visibility"""
-    result_frame = frame.copy()
+        return distance < Config.POSITION_THRESHOLD
     
-    for det in detections:
-        label = det.get("violation", "Unknown")
-        confidence = det.get("confidence", 0.0)
-        x, y, w, h = det.get("bounding_box", [0, 0, 0, 0])
-        worker_id = det.get("worker_id", "Unknown")
-
-        x1 = int(x - w/2)
-        y1 = int(y - h/2)
-        x2 = int(x + w/2)
-        y2 = int(y + h/2)
-        
-        color = CONFIG["CLASS_COLORS"].get(label, (0, 0, 255))
-        
-        # Draw thicker rectangle with border
-        cv2.rectangle(result_frame, (x1, y1), (x2, y2), color, 3)
-        
-        # Add black background behind text
-        display_text = f"{CONFIG['DISPLAY_NAMES'].get(label, label)} (Worker {worker_id})"
-        text_size = cv2.getTextSize(display_text, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)[0]
-        cv2.rectangle(result_frame, (x1, y1-text_size[1]-10), (x1+text_size[0]+10, y1), (0, 0, 0), -1)
-        cv2.putText(result_frame, display_text, (x1+5, y1-5), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 255, 255), 2)
-        
-        # Add confidence score
-        conf_text = f"Conf: {confidence:.2f}"
-        cv2.putText(result_frame, conf_text, (x1+5, y2+20), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 2)
+    def _get_face_id(self, frame: np.ndarray, bbox: Tuple) -> Optional[int]:
+        """Try to identify a person by their face"""
+        if not self.face_tracker or not self.face_tracker.enabled:
+            return None
+            
+        x, y, w, h = bbox
+        face_region = frame[max(0, y):y+h, max(0, x):x+w]
         
-    return result_frame
-
-def calculate_safety_score(violations):
-    """Calculate safety score based on detected violations"""
-    penalties = {
-        "no_helmet": 25,
-        "no_harness": 30,
-        "unsafe_posture": 20,
-        "unsafe_zone": 35,
-        "improper_tool_use": 25
-    }
-    
-    # Count unique violation types per worker
-    worker_violations = {}
-    for v in violations:
-        worker_id = v.get("worker_id", "Unknown")
-        violation_type = v.get("violation", "Unknown")
-        
-        if worker_id not in worker_violations:
-            worker_violations[worker_id] = set()
-        worker_violations[worker_id].add(violation_type)
-    
-    # Calculate total penalty
-    total_penalty = 0
-    for worker_violations_set in worker_violations.values():
-        worker_penalty = sum(penalties.get(v, 0) for v in worker_violations_set)
-        total_penalty += worker_penalty
-    
-    score = max(0, 100 - total_penalty)
-    return score
-
-def generate_violation_pdf(violations, score):
-    """Generate a PDF report for the detected violations"""
-    try:
-        pdf_filename = f"violations_{int(time.time())}.pdf"
-        pdf_path = os.path.join(CONFIG["OUTPUT_DIR"], pdf_filename)
-        pdf_file = BytesIO()
-        c = canvas.Canvas(pdf_file, pagesize=letter)
-        
-        # Title
-        c.setFont("Helvetica-Bold", 16)
-        c.drawString(1 * inch, 10 * inch, "Worksite Safety Violation Report")
-        
-        # Basic Information
-        c.setFont("Helvetica", 12)
-        c.drawString(1 * inch, 9.5 * inch, f"Date: {time.strftime('%Y-%m-%d')}")
-        c.drawString(1 * inch, 9.2 * inch, f"Time: {time.strftime('%H:%M:%S')}")
-        
-        # Safety Score
-        c.setFont("Helvetica-Bold", 14)
-        c.drawString(1 * inch, 8.7 * inch, f"Safety Compliance Score: {score}%")
-
-        # Violation Summary
-        y_position = 8.2 * inch
-        c.setFont("Helvetica-Bold", 12)
-        c.drawString(1 * inch, y_position, "Summary:")
-        y_position -= 0.3 * inch
-        
-        # Group violations by worker
-        worker_violations = {}
-        for v in violations:
-            worker_id = v.get("worker_id", "Unknown")
-            if worker_id not in worker_violations:
-                worker_violations[worker_id] = []
-            worker_violations[worker_id].append(v)
-        
-        c.setFont("Helvetica", 10)
-        summary_data = {
-            "Total Workers with Violations": len(worker_violations),
-            "Total Violations Found": len(violations),
-            "Analysis Timestamp": time.strftime("%Y-%m-%d %H:%M:%S")
-        }
-        
-        for key, value in summary_data.items():
-            c.drawString(1 * inch, y_position, f"{key}: {value}")
-            y_position -= 0.25 * inch
-
-        # Detailed Violations by Worker
-        y_position -= 0.5 * inch
-        c.setFont("Helvetica-Bold", 12)
-        c.drawString(1 * inch, y_position, "Violations by Worker:")
-        y_position -= 0.3 * inch
-        
-        c.setFont("Helvetica", 10)
-        for worker_id, worker_vios in worker_violations.items():
-            c.drawString(1 * inch, y_position, f"Worker {worker_id}:")
-            y_position -= 0.2 * inch
+        faces = self.face_tracker.detect_faces(face_region)
+        if len(faces) == 0:
+            return None
             
-            for v in worker_vios:
-                display_name = CONFIG["DISPLAY_NAMES"].get(v.get("violation", "Unknown"), "Unknown")
-                time_str = f"{v.get('timestamp', 0.0):.2f}s"
-                conf_str = f"{v.get('confidence', 0.0):.2f}"
+        # Get the largest face in the region
+        largest_face = max(faces, key=lambda f: f[2]*f[3])
+        fx, fy, fw, fh = largest_face
+        
+        # Check if we've seen this face before
+        for track_id, info in self.tracked_workers.items():
+            if 'face_bbox' not in info:
+                continue
                 
-                violation_text = f"  - {display_name} at {time_str} (Confidence: {conf_str})"
-                c.drawString(1.2 * inch, y_position, violation_text)
-                y_position -= 0.2 * inch
+            # Compare with stored face
+            iou = self._calculate_iou((fx, fy, fw, fh), info['face_bbox'])
+            if iou > 0.5:  # Threshold for face matching
+                return track_id
                 
-                if y_position < 1 * inch:
-                    c.showPage()
-                    c.setFont("Helvetica", 10)
-                    y_position = 10 * inch
-
-        c.save()
-        pdf_file.seek(0)
-
-        # Save PDF file
-        with open(pdf_path, "wb") as f:
-            f.write(pdf_file.getvalue())
+        return None
+    
+    def update(self, frame: np.ndarray, detections: List[Dict], timestamp: float) -> List[Dict]:
+        """Update tracker with new detections and return unique violations"""
+        current_time = time.time()
+        unique_violations = []
+        
+        # First pass - try to match with existing tracks
+        for det in detections:
+            bbox = det['bbox']
+            violation_type = det['violation']
+            confidence = det['confidence']
+            x, y, w, h = bbox
+            center = (x + w/2, y + h/2)
             
-        public_url = f"{CONFIG['PUBLIC_URL_BASE']}{pdf_filename}"
-        logger.info(f"PDF generated: {public_url}")
-        return pdf_path, public_url, pdf_file
-    except Exception as e:
-        logger.error(f"Error generating PDF: {e}")
-        return "", "", None
-
-@retry(stop_max_attempt_number=3, wait_fixed=2000)
-def connect_to_salesforce():
-    """Connect to Salesforce with retry logic"""
-    try:
-        sf = Salesforce(**CONFIG["SF_CREDENTIALS"])
-        logger.info("Connected to Salesforce")
-        sf.describe()
-        return sf
-    except Exception as e:
-        logger.error(f"Salesforce connection failed: {e}")
-        raise
-
-def upload_pdf_to_salesforce(sf, pdf_file, report_id):
-    """Upload PDF report to Salesforce"""
-    try:
-        if not pdf_file:
-            logger.error("No PDF file provided for upload")
-            return ""
+            # Try face recognition for helmet violations
+            track_id = None
+            if violation_type == "no_helmet" and self.face_tracker:
+                track_id = self._get_face_id(frame, bbox)
             
-        encoded_pdf = base64.b64encode(pdf_file.getvalue()).decode('utf-8')
-        content_version_data = {
-            "Title": f"Safety_Violation_Report_{int(time.time())}",
-            "PathOnClient": f"safety_violation_{int(time.time())}.pdf",
-            "VersionData": encoded_pdf,
-            "FirstPublishLocationId": report_id
-        }
-        content_version = sf.ContentVersion.create(content_version_data)
-        result = sf.query(f"SELECT Id, ContentDocumentId FROM ContentVersion WHERE Id = '{content_version['id']}'")
-        
-        if not result['records']:
-            logger.error("Failed to retrieve ContentVersion")
-            return ""
+            # If no face match, try position matching
+            if track_id is None:
+                for tid, info in self.tracked_workers.items():
+                    if current_time - info['last_seen'] > Config.TRACK_BUFFER / 30.0:
+                        continue
+                        
+                    if 'position' in info and self._is_same_position(center, info['position']):
+                        track_id = tid
+                        break
             
-        file_url = f"https://{sf.sf_instance}/sfc/servlet.shepherd/version/download/{content_version['id']}"
-        logger.info(f"PDF uploaded to Salesforce: {file_url}")
-        return file_url
-    except Exception as e:
-        logger.error(f"Error uploading PDF to Salesforce: {e}")
-        return ""
-
-def push_report_to_salesforce(violations, score, pdf_path, pdf_file):
-    """Push violation report to Salesforce"""
-    try:
-        sf = connect_to_salesforce()
-        
-        # Format violations for Salesforce
-        violations_text = ""
-        for v in violations:
-            display_name = CONFIG['DISPLAY_NAMES'].get(v.get('violation', 'Unknown'), 'Unknown')
-            worker_id = v.get('worker_id', 'Unknown')
-            timestamp = v.get('timestamp', 0.0)
-            confidence = v.get('confidence', 0.0)
+            # If still no match, create new track
+            if track_id is None:
+                track_id = self.next_id
+                self.next_id += 1
+                self.tracked_workers[track_id] = {
+                    'last_seen': current_time,
+                    'violations': set(),
+                    'position': center
+                }
+                
+                # Store face if available
+                if violation_type == "no_helmet" and self.face_tracker:
+                    faces = self.face_tracker.detect_faces(frame[y:y+h, x:x+w])
+                    if len(faces) > 0:
+                        self.tracked_workers[track_id]['face_bbox'] = max(
+                            faces, key=lambda f: f[2]*f[3])
+            
+            # Update track info
+            self.tracked_workers[track_id].update({
+                'last_seen': current_time,
+                'position': center,
+                'bbox': bbox
+            })
             
-            violations_text += f"Worker {worker_id}: {display_name} at {timestamp:.2f}s (Conf: {confidence:.2f})\n"
+            # Check if this is a new violation for this track
+            if violation_type not in self.tracked_workers[track_id]['violations']:
+                # Check if this violation type has been seen recently
+                recent_violations = [t for t in self.violation_history[violation_type] 
+                                    if current_time - t < Config.VIOLATION_COOLDOWN]
+                
+                if not recent_violations:
+                    self.tracked_workers[track_id]['violations'].add(violation_type)
+                    self.violation_history[violation_type].append(current_time)
+                    
+                    unique_violations.append({
+                        'track_id': track_id,
+                        'violation': violation_type,
+                        'bbox': bbox,
+                        'confidence': confidence,
+                        'timestamp': timestamp
+                    })
         
-        if not violations_text:
-            violations_text = "No violations detected."
+        # Clean up old tracks
+        stale_ids = [tid for tid, info in self.tracked_workers.items() 
+                     if current_time - info['last_seen'] > Config.TRACK_BUFFER / 30.0]
+        for tid in stale_ids:
+            del self.tracked_workers[tid]
             
-        pdf_url = f"{CONFIG['PUBLIC_URL_BASE']}{os.path.basename(pdf_path)}" if pdf_path else ""
+        return unique_violations
 
-        record_data = {
-            "Compliance_Score__c": score,
-            "Violations_Found__c": len(violations),
-            "Violations_Details__c": violations_text,
-            "Status__c": "Pending",
-            "PDF_Report_URL__c": pdf_url
-        }
-        
-        logger.info(f"Creating Salesforce record with data: {record_data}")
+# ========================== # Video Processor # ==========================
+class VideoProcessor:
+    def __init__(self):
+        self.model = self._load_model()
+        self.tracker = ViolationTracker()
         
+    def _load_model(self):
+        """Load YOLOv8 model with fallback"""
         try:
-            record = sf.Safety_Video_Report__c.create(record_data)
-            logger.info(f"Created Safety_Video_Report__c record: {record['id']}")
+            if os.path.exists(Config.MODEL_PATH):
+                model = YOLO(Config.MODEL_PATH).to(Config.DEVICE)
+                logger.info(f"Loaded custom model from {Config.MODEL_PATH}")
+            else:
+                model = YOLO(Config.FALLBACK_MODEL).to(Config.DEVICE)
+                logger.warning("Using fallback YOLOv8n model")
+            return model
         except Exception as e:
-            logger.error(f"Failed to create Safety_Video_Report__c: {e}")
-            record = sf.Account.create({"Name": f"Safety_Report_{int(time.time())}"})
-            logger.warning(f"Fell back to Account record: {record['id']}")
+            logger.error(f"Failed to load model: {e}")
+            raise
             
-        record_id = record["id"]
-
-        if pdf_file:
-            uploaded_url = upload_pdf_to_salesforce(sf, pdf_file, record_id)
-            if uploaded_url:
-                try:
-                    sf.Safety_Video_Report__c.update(record_id, {"PDF_Report_URL__c": uploaded_url})
-                    logger.info(f"Updated record {record_id} with PDF URL: {uploaded_url}")
-                except Exception as e:
-                    logger.error(f"Failed to update Safety_Video_Report__c: {e}")
-                    sf.Account.update(record_id, {"Description": uploaded_url})
-                    logger.info(f"Updated Account record {record_id} with PDF URL")
-                pdf_url = uploaded_url
-
-        return record_id, pdf_url
-    except Exception as e:
-        logger.error(f"Salesforce record creation failed: {e}", exc_info=True)
-        return None, ""
-
-def process_video(video_data):
-    """Process video to detect safety violations"""
-    try:
-        os.makedirs(CONFIG["OUTPUT_DIR"], exist_ok=True)
-        logger.info(f"Output directory ensured: {CONFIG['OUTPUT_DIR']}")
-
-        video_path = os.path.join(CONFIG["OUTPUT_DIR"], f"temp_{int(time.time())}.mp4")
-        with open(video_path, "wb") as f:
-            f.write(video_data)
-        logger.info(f"Video saved: {video_path}")
-
+    def _draw_detection(self, frame: np.ndarray, detection: Dict) -> np.ndarray:
+        """Draw a single detection on the frame"""
+        x, y, w, h = detection['bbox']
+        label = detection['violation']
+        confidence = detection['confidence']
+        track_id = detection.get('track_id', 0)
+        
+        color = Config.COLOR_MAP.get(label, (0, 0, 255))
+        
+        # Draw bounding box
+        cv2.rectangle(frame, (x, y), (x+w, y+h), color, 2)
+        
+        # Draw label background
+        label_text = f"{label} (ID: {track_id})"
+        (text_width, text_height), _ = cv2.getTextSize(
+            label_text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
+        
+        cv2.rectangle(frame, (x, y - text_height - 10), 
+                     (x + text_width + 10, y), color, -1)
+        
+        # Draw label text
+        cv2.putText(frame, label_text, (x + 5, y - 5), 
+                   cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 255, 255), 1)
+        
+        # Draw confidence
+        conf_text = f"{confidence:.2f}"
+        cv2.putText(frame, conf_text, (x + 5, y + h + 15), 
+                   cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 1)
+        
+        return frame
+    
+    def _process_frame(self, frame: np.ndarray, frame_idx: int, fps: float) -> Tuple[List[Dict], np.ndarray]:
+        """Process a single video frame"""
+        timestamp = frame_idx / fps
+        
+        # Run YOLO detection
+        results = self.model(frame, verbose=False)
+        detections = []
+        
+        for result in results:
+            for box in result.boxes:
+                cls_id = int(box.cls)
+                conf = float(box.conf)
+                label = Config.VIOLATION_LABELS.get(cls_id)
+                
+                if label is None or conf < Config.MIN_VIOLATION_CONFIDENCE:
+                    continue
+                
+                # Convert bounding box to (x, y, w, h) format
+                x1, y1, x2, y2 = map(int, box.xyxy[0])
+                detections.append({
+                    'bbox': (x1, y1, x2 - x1, y2 - y1),
+                    'violation': label,
+                    'confidence': conf
+                })
+        
+        # Update tracker and get unique violations
+        unique_violations = self.tracker.update(frame, detections, timestamp)
+        
+        # Draw detections on frame
+        output_frame = frame.copy()
+        for violation in unique_violations:
+            output_frame = self._draw_detection(output_frame, violation)
+        
+        return unique_violations, output_frame
+    
+    def process_video(self, video_path: str) -> Tuple[List[Dict], List[Tuple[float, np.ndarray]]]:
+        """Process entire video file"""
         cap = cv2.VideoCapture(video_path)
         if not cap.isOpened():
-            os.remove(video_path)
-            raise ValueError("Could not open video file")
-
+            raise ValueError(f"Could not open video: {video_path}")
+        
+        fps = cap.get(cv2.CAP_PROP_FPS)
         total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
-        fps = cap.get(cv2.CAP_PROP_FPS) or 30
-        duration = total_frames / fps
-        width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-        height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
-        logger.info(f"Video properties: {duration:.2f}s, {total_frames} frames, {fps:.1f} FPS, {width}x{height}")
-
-        tracker = BYTETracker(
-            track_thresh=CONFIG["TRACK_THRESH"],
-            track_buffer=CONFIG["TRACK_BUFFER"],
-            match_thresh=CONFIG["MATCH_THRESH"],
-            frame_rate=fps
-        )
-
-        # Track unique violations by worker ID
-        unique_violations = {}  # {worker_id: {violation_type: first_detection_time}}
+        violations = []
         snapshots = []
-        start_time = time.time()
-        frame_skip = CONFIG["FRAME_SKIP"]
-        processed_frames = 0
-
-        while processed_frames < total_frames:
-            batch_frames = []
-            batch_indices = []
-            
-            for _ in range(CONFIG["BATCH_SIZE"]):
-                frame_idx = int(cap.get(cv2.CAP_PROP_POS_FRAMES))
-                if frame_idx >= total_frames:
-                    break
-                
-                ret, frame = cap.read()
-                if not ret:
-                    break
-                
-                frame = preprocess_frame(frame)
+        
+        logger.info(f"Processing video: {video_path} ({total_frames} frames, {fps:.1f} FPS)")
+        
+        frame_idx = 0
+        while cap.isOpened():
+            ret, frame = cap.read()
+            if not ret:
+                break
                 
-                # Skip frames if needed
-                for _ in range(frame_skip - 1):
-                    if not cap.grab():
-                        break
+            # Skip frames for efficiency
+            if frame_idx % Config.FRAME_SKIP != 0:
+                frame_idx += 1
+                continue
                 
-                batch_frames.append(frame)
-                batch_indices.append(frame_idx)
-                processed_frames += 1
-
-            if not batch_frames:
-                break
-
-            # Process batch with YOLO model
-            results = model(batch_frames, device=device, conf=0.1, verbose=False)
+            # Process frame
+            frame_violations, output_frame = self._process_frame(frame, frame_idx, fps)
             
-            for i, (result, frame_idx) in enumerate(zip(results, batch_indices)):
-                current_time = frame_idx / fps
+            # Save snapshots for new violations
+            for violation in frame_violations:
+                violations.append(violation)
                 
-                # Update progress every second
-                if time.time() - start_time > 1.0:
-                    progress = (processed_frames / total_frames) * 100
-                    yield f"Processing video... {progress:.1f}% complete (Frame {processed_frames}/{total_frames})", "", "", "", ""
-                    start_time = time.time()
-
-                boxes = result.boxes
-                track_inputs = []
+                # Save snapshot
+                timestamp = frame_idx / fps
+                snapshot_path = os.path.join(
+                    Config.OUTPUT_DIR,
+                    f"violation_{violation['violation']}_id{violation['track_id']}_{int(timestamp*100)}.jpg")
                 
-                for box in boxes:
-                    cls = int(box.cls)
-                    conf = float(box.conf)
-                    label = CONFIG["VIOLATION_LABELS"].get(cls, None)
-                    
-                    if label is None:
-                        continue
-                        
-                    if conf < CONFIG["CONFIDENCE_THRESHOLDS"].get(label, 0.25):
-                        continue
-
-                    bbox = box.xywh.cpu().numpy()[0]
-                    track_inputs.append({
-                        "bbox": bbox,
-                        "conf": conf,
-                        "cls": cls
-                    })
-
-                if not track_inputs:
-                    continue
-                    
-                tracked_objects = tracker.update(
-                    np.array([t["bbox"] for t in track_inputs]),
-                    np.array([t["conf"] for t in track_inputs]),
-                    np.array([t["cls"] for t in track_inputs])
-                )
-
-                # Process tracked objects for violations
-                for obj in tracked_objects:
-                    worker_id = obj['id']
-                    label = CONFIG["VIOLATION_LABELS"].get(int(obj['cls']), None)
-                    conf = obj['score']
-                    bbox = obj['bbox']
-                    
-                    if label is None:
-                        continue
-                        
-                    # Initialize worker if not seen before
-                    if worker_id not in unique_violations:
-                        unique_violations[worker_id] = {}
-                    
-                    # Check if this violation type has been recorded for this worker
-                    if label not in unique_violations[worker_id]:
-                        # This is a new violation type for this worker
-                        unique_violations[worker_id][label] = current_time
-                        
-                        # Create detection object
-                        detection = {
-                            "worker_id": worker_id,
-                            "violation": label,
-                            "confidence": round(conf, 2),
-                            "bounding_box": bbox,
-                            "timestamp": current_time
-                        }
-                        
-                        # Take snapshot for the new violation
-                        snapshot_frame = batch_frames[i].copy()
-                        snapshot_frame = draw_detections(snapshot_frame, [detection])
-                        
-                        # Add timestamp to snapshot
-                        cv2.putText(
-                            snapshot_frame,
-                            f"Time: {current_time:.2f}s",
-                            (10, 30),
-                            cv2.FONT_HERSHEY_SIMPLEX,
-                            0.7,
-                            (255, 255, 255),
-                            2
-                        )
-                        
-                        # Save snapshot with high quality
-                        snapshot_filename = f"violation_{label}_worker{worker_id}_{int(current_time*100)}.jpg"
-                        snapshot_path = os.path.join(CONFIG["OUTPUT_DIR"], snapshot_filename)
-                        
-                        cv2.imwrite(
-                            snapshot_path,
-                            snapshot_frame,
-                            [cv2.IMWRITE_JPEG_QUALITY, CONFIG["SNAPSHOT_QUALITY"]]
-                        )
-                        
-                        snapshots.append({
-                            "violation": label,
-                            "worker_id": worker_id,
-                            "timestamp": current_time,
-                            "snapshot_path": snapshot_path,
-                            "snapshot_url": f"{CONFIG['PUBLIC_URL_BASE']}{snapshot_filename}"
-                        })
-                        
-                        logger.info(f"Captured snapshot for {label} violation by worker {worker_id} at {current_time:.2f}s")
-
-        cap.release()
-        if os.path.exists(video_path):
-            os.remove(video_path)
+                cv2.imwrite(snapshot_path, output_frame, 
+                           [cv2.IMWRITE_JPEG_QUALITY, Config.SNAPSHOT_QUALITY])
+                
+                snapshots.append((timestamp, snapshot_path))
+                logger.info(f"Detected {violation['violation']} at {timestamp:.2f}s (ID: {violation['track_id']})")
             
-        processing_time = time.time() - start_time
-        logger.info(f"Processing complete in {processing_time:.2f}s")
-
-        # Convert tracked violations to final violation list
-        violations = []
-        for worker_id, worker_violations in unique_violations.items():
-            for label, detection_time in worker_violations.items():
-                violation = {
-                    "worker_id": worker_id,
-                    "violation": label,
-                    "timestamp": detection_time
-                }
-                violations.append(violation)
-
-        if not violations:
-            logger.info("No violations detected after processing")
-            yield "No violations detected in the video.", "Safety Score: 100%", "No snapshots captured.", "N/A", "N/A"
-            return
-
-        # Calculate safety score
-        score = calculate_safety_score(violations)
+            frame_idx += 1
+            
+            # Yield progress
+            if frame_idx % 10 == 0:
+                progress = frame_idx / total_frames
+                yield progress, violations, snapshots
         
-        # Generate PDF report
-        pdf_path, pdf_url, pdf_file = generate_violation_pdf(violations, score)
+        cap.release()
+        yield 1.0, violations, snapshots  # Final yield
+    
+    def calculate_safety_score(self, violations: List[Dict]) -> int:
+        """Calculate safety compliance score (0-100)"""
+        penalty = 0
+        violation_types = set()
         
-        # Push report to Salesforce
-        report_id, final_pdf_url = push_report_to_salesforce(violations, score, pdf_path, pdf_file)
-
-        # Format violations table for display
-        violation_table = "| Violation | Worker ID | Time (s) | Confidence |\n"
-        violation_table += "|-----------|-----------|----------|------------|\n"
-        
-        for v in sorted(violations, key=lambda x: (x.get("worker_id", "Unknown"), x.get("timestamp", 0.0))):
-            display_name = CONFIG["DISPLAY_NAMES"].get(v.get("violation", "Unknown"), "Unknown")
-            worker_id = v.get("worker_id", "Unknown")
-            timestamp = v.get("timestamp", 0.0)
-            confidence = v.get("confidence", 0.0)
-            
-            violation_table += f"| {display_name} | {worker_id} | {timestamp:.2f} | {confidence:.2f} |\n"
+        for violation in violations:
+            violation_types.add(violation['violation'])
+        
+        for v_type in violation_types:
+            penalty += Config.PENALTIES.get(v_type, 0)
+        
+        return max(0, 100 - penalty)
 
-        # Format snapshots for display
-        snapshots_text = ""
-        for s in snapshots:
-            display_name = CONFIG["DISPLAY_NAMES"].get(s["violation"], "Unknown")
-            worker_id = s.get("worker_id", "Unknown")
-            timestamp = s.get("timestamp", 0.0)
-            
-            snapshots_text += f"### {display_name} - Worker {worker_id} at {timestamp:.2f}s\n\n"
-            snapshots_text += f"![Violation]({s['snapshot_url']})\n\n"
+# ========================== # Gradio Interface # ==========================
+def setup_output_dir():
+    """Ensure output directory exists"""
+    os.makedirs(Config.OUTPUT_DIR, exist_ok=True)
+    logger.info(f"Output directory: {Config.OUTPUT_DIR}")
+
+def format_violation_table(violations: List[Dict]) -> str:
+    """Format violations as markdown table"""
+    if not violations:
+        return "No violations detected"
+    
+    table = "| Violation | Worker ID | Time (s) | Confidence |\n"
+    table += "|-----------|-----------|----------|------------|\n"
+    
+    for v in sorted(violations, key=lambda x: x['timestamp']):
+        table += f"| {v['violation']} | {v['track_id']} | {v['timestamp']:.2f} | {v['confidence']:.2f} |\n"
+    
+    return table
 
-        if not snapshots_text:
-            snapshots_text = "No snapshots captured."
+def format_snapshots(snapshots: List[Tuple[float, str]]) -> str:
+    """Format snapshots as markdown"""
+    if not snapshots:
+        return "No violations captured"
+    
+    markdown = ""
+    for timestamp, path in snapshots:
+        filename = os.path.basename(path)
+        markdown += f"### Violation at {timestamp:.2f}s\n\n"
+        markdown += f"![Violation](file/{path})\n\n"
+    
+    return markdown
 
+def process_video_wrapper(video_file: str):
+    """Wrapper for Gradio interface"""
+    setup_output_dir()
+    
+    # Create temporary video path
+    temp_video_path = os.path.join(Config.OUTPUT_DIR, f"temp_{int(time.time())}.mp4")
+    with open(temp_video_path, "wb") as f:
+        f.write(open(video_file, "rb").read())
+    
+    processor = VideoProcessor()
+    
+    try:
+        for progress, violations, snapshots in processor.process_video(temp_video_path):
+            yield (
+                f"Processing... {progress*100:.1f}% complete",
+                "",
+                "",
+                ""
+            )
+        
+        # Calculate final results
+        score = processor.calculate_safety_score(violations)
+        violation_table = format_violation_table(violations)
+        snapshots_md = format_snapshots(snapshots)
+        
         yield (
-            violation_table,
+            "Processing complete",
             f"Safety Score: {score}%",
-            snapshots_text,
-            f"Salesforce Record ID: {report_id or 'N/A'}",
-            final_pdf_url or "N/A"
+            violation_table,
+            snapshots_md
         )
-
-    except Exception as e:
-        logger.error(f"Error processing video: {e}", exc_info=True)
-        if 'video_path' in locals() and os.path.exists(video_path):
-            os.remove(video_path)
-        yield f"Error processing video: {e}", "", "", "", ""
-
-def gradio_interface(video_file):
-    """Gradio interface for the video processing"""
-    if not video_file:
-        return "No file uploaded.", "", "No file uploaded.", "", ""
         
-    try:
-        with open(video_file, "rb") as f:
-            video_data = f.read()
-
-        for status, score, snapshots_text, record_id, details_url in process_video(video_data):
-            yield status, score, snapshots_text, record_id, details_url
+    finally:
+        if os.path.exists(temp_video_path):
+            os.remove(temp_video_path)
+
+# ========================== # Main Application # ==========================
+def create_interface():
+    """Create Gradio interface"""
+    with gr.Blocks(title="Safety Compliance Analyzer") as interface:
+        gr.Markdown("# 🚧 Safety Compliance Video Analyzer")
+        gr.Markdown("Upload site videos to detect safety violations (No Helmet, No Harness, etc.)")
+        
+        with gr.Row():
+            with gr.Column():
+                video_input = gr.Video(label="Upload Site Video")
+                submit_btn = gr.Button("Analyze Video", variant="primary")
             
-    except Exception as e:
-        logger.error(f"Error in Gradio interface: {e}", exc_info=True)
-        yield f"Error: {str(e)}", "", "Error in processing.", "", ""
-
-# ========================== # Gradio Interface # ==========================
-interface = gr.Interface(
-    fn=gradio_interface,
-    inputs=gr.Video(label="Upload Site Video"),
-    outputs=[
-        gr.Markdown(label="Detected Safety Violations"),
-        gr.Textbox(label="Compliance Score"),
-        gr.Markdown(label="Snapshots"),
-        gr.Textbox(label="Salesforce Record ID"),
-        gr.Textbox(label="Violation Details URL")
-    ],
-    title="Worksite Safety Violation Analyzer",
-    description="Upload site videos to detect safety violations (No Helmet, No Harness, Unsafe Posture, Unsafe Zone, Improper Tool Use). Each unique violation is detected only once per worker.",
-    allow_flagging="never"
-)
+            with gr.Column():
+                progress_out = gr.Textbox(label="Status")
+                score_out = gr.Textbox(label="Safety Score")
+                violations_out = gr.Markdown(label="Detected Violations")
+                snapshots_out = gr.Markdown(label="Violation Snapshots")
+        
+        submit_btn.click(
+            fn=process_video_wrapper,
+            inputs=video_input,
+            outputs=[progress_out, score_out, violations_out, snapshots_out]
+        )
+    
+    return interface
 
 if __name__ == "__main__":
-    logger.info("Launching Enhanced Safety Analyzer App...")
+    setup_output_dir()
+    interface = create_interface()
     interface.launch()