app.py

import os
import sys
import subprocess
import logging
import warnings
import cv2
import gradio as gr
import torch
import numpy as np
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
import tempfile
import shutil
import tenacity

# ========================== # Configuration and Setup # ==========================
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)

def check_ffmpeg():
    try:
        subprocess.run(["ffmpeg", "-version"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
        logger.info("FFmpeg is available.")
        return True
    except (subprocess.CalledProcessError, FileNotFoundError):
        logger.error("FFmpeg is not installed or not found in PATH. Video processing may fail.")
        return False

FFMPEG_AVAILABLE = check_ffmpeg()

# ========================== # ByteTrack Implementation # ==========================
class BYTETracker:
    def __init__(self, track_thresh=0.3, track_buffer=90, match_thresh=0.3, frame_rate=30, max_distance=100):
        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 = {}
        self.worker_history = {}
        self.last_positions = {}
        self.recently_removed = {}  # Store recently removed tracks for re-identification
        self.track_attributes = {}  # Store additional attributes like appearance features
        self.active_workers = set()  # Track currently active workers
        self.worker_violation_history = {}  # Track violations per worker
        self.max_worker_distance = max_distance

    def update(self, dets, scores, cls):
        tracks = []
        current_time = time.time()
        
        # Prune stale tracks
        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:
            # Store recently removed tracks for re-identification (for 2 seconds)
            self.recently_removed[track_id] = {
                'bbox': self.tracks[track_id]['bbox'],
                'last_seen': current_time,
                'last_position': self.last_positions.get(track_id, [0, 0]),
                'appearance': self.track_attributes.get(track_id, {}).get('appearance', None)
            }
            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]
            if track_id in self.active_workers:
                self.active_workers.remove(track_id)

        # Clean up recently_removed tracks older than 2 seconds
        to_remove = []
        for track_id, info in self.recently_removed.items():
            if current_time - info['last_seen'] > 2.0:
                to_remove.append(track_id)
        for track_id in to_remove:
            del self.recently_removed[track_id]

        # Process new detections
        active_tracks = {}
        
        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 active tracks
            for track_id, track_info in self.tracks.items():
                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
                })
                
                if 'appearance' not in self.track_attributes.get(best_track_id, {}):
                    self.track_attributes[best_track_id] = {'appearance': self._extract_appearance_features([x, y, w, h])}
                
                if best_track_id not in self.worker_history:
                    self.worker_history[best_track_id] = []
                
                self.worker_history[best_track_id].append({'pos': [x, y], 'time': current_time})
                
                if len(self.worker_history[best_track_id]) > 30:
                    self.worker_history[best_track_id] = self.worker_history[best_track_id][-30:]
                
                self.last_positions[best_track_id] = [x, y]
                self.active_workers.add(best_track_id)
                
                if cl is not None:
                    if best_track_id not in self.worker_violation_history:
                        self.worker_violation_history[best_track_id] = set()
                    self.worker_violation_history[best_track_id].add(int(cl))
                
                active_tracks[best_track_id] = {
                    'id': best_track_id,
                    'bbox': [x, y, w, h],
                    'score': score,
                    'cls': cl
                }
            else:
                # Try to re-identify with recently removed tracks
                reidentified = False
                for track_id, info in self.recently_removed.items():
                    if self._is_same_worker([x, y], info['last_position']):
                        self.tracks[track_id] = {
                            'bbox': [x, y, w, h],
                            'score': score,
                            'cls': cl,
                            'last_seen': current_time
                        }
                        if track_id not in self.worker_history:
                            self.worker_history[track_id] = []
                        self.worker_history[track_id].append({'pos': [x, y], 'time': current_time})
                        self.last_positions[track_id] = [x, y]
                        self.active_workers.add(track_id)
                        
                        if cl is not None:
                            if track_id not in self.worker_violation_history:
                                self.worker_violation_history[track_id] = set()
                            self.worker_violation_history[track_id].add(int(cl))
                        
                        active_tracks[track_id] = {
                            'id': track_id,
                            'bbox': [x, y, w, h],
                            'score': score,
                            'cls': cl
                        }
                        reidentified = True
                        del self.recently_removed[track_id]
                        break
                
                if not reidentified:
                    # Try to match with last positions of existing tracks via distance
                    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
                            }
                            
                            if worker_id not in self.worker_history:
                                self.worker_history[worker_id] = []
                            self.worker_history[worker_id].append({'pos': [x, y], 'time': current_time})
                            self.last_positions[worker_id] = [x, y]
                            self.active_workers.add(worker_id)
                            
                            if cl is not None:
                                if worker_id not in self.worker_violation_history:
                                    self.worker_violation_history[worker_id] = set()
                                self.worker_violation_history[worker_id].add(int(cl))
                            
                            active_tracks[worker_id] = {
                                'id': worker_id,
                                'bbox': [x, y, w, h],
                                'score': score,
                                'cls': cl
                            }
                            same_worker = True
                            break
                    
                    if not same_worker:
                        # Register a new track
                        new_id = self.next_id
                        self.tracks[new_id] = {
                            'bbox': [x, y, w, h],
                            'score': score,
                            'cls': cl,
                            'last_seen': current_time
                        }
                        self.track_attributes[new_id] = {'appearance': self._extract_appearance_features([x, y, w, h])}
                        self.worker_history[new_id] = [{'pos': [x, y], 'time': current_time}]
                        self.last_positions[new_id] = [x, y]
                        self.active_workers.add(new_id)
                        
                        if cl is not None:
                            if new_id not in self.worker_violation_history:
                                self.worker_violation_history[new_id] = set()
                            self.worker_violation_history[new_id].add(int(cl))
                        
                        active_tracks[new_id] = {
                            'id': new_id,
                            'bbox': [x, y, w, h],
                            'score': score,
                            'cls': cl
                        }
                        self.next_id += 1
        
        return list(active_tracks.values())

    def _calculate_iou(self, box1, box2):
        x1, y1, w1, h1 = box1
        x2, y2, w2, h2 = box2
        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)
        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
    
    def _is_same_worker(self, pos1, pos2):
        x1, y1 = pos1
        x2, y2 = pos2
        distance = np.sqrt((x1 - x2)**2 + (y1 - y2)**2)
        return distance < self.max_worker_distance
    
    def _extract_appearance_features(self, bbox):
        """Simple appearance feature extraction (placeholder)"""
        _, _, w, h = bbox
        return [w, h, w/h]
    
    def get_active_worker_count(self):
        return len(self.active_workers)
    
    def get_worker_violation_types(self, worker_id):
        return self.worker_violation_history.get(worker_id, set())
    
    def get_all_workers(self):
        return set(list(self.tracks.keys()) + list(self.recently_removed.keys()))

# ========================== # Optimized Configuration # ==========================
CONFIG = {
    "MODEL_PATH": "yolov8_safety.pt",
    "FALLBACK_MODEL": "yolov8n.pt",
    "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),
        "no_harness": (0, 165, 255),
        "unsafe_posture": (0, 255, 0),
        "unsafe_zone": (255, 0, 0),
        "improper_tool_use": (255, 255, 0)
    },
    "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": os.getenv("SF_USERNAME", "prashanth1ai@safety.com"),
        "password": os.getenv("SF_PASSWORD", "SaiPrash461"),
        "security_token": os.getenv("SF_SECURITY_TOKEN", "AP4AQnPoidIKPvSvNEfAHyoK"),
        "domain": "login"
    },
    "PUBLIC_URL_BASE": "https://huggingface.co/spaces/PrashanthB461/AI_Safety_Demo2/resolve/main/static/output/",
    "CONFIDENCE_THRESHOLDS": {
        "no_helmet": 0.4,
        "no_harness": 0.25,
        "unsafe_posture": 0.25,
        "unsafe_zone": 0.25,
        "improper_tool_use": 0.25
    },
    "MIN_VIOLATION_FRAMES": 1,
    "VIOLATION_COOLDOWN": 30.0,
    "WORKER_TRACKING_DURATION": 10.0,
    "MAX_PROCESSING_TIME": 60,
    "FRAME_SKIP": 1,
    "BATCH_SIZE": 15,
    "PARALLEL_WORKERS": max(1, cpu_count() - 1),
    "TRACK_BUFFER": 150,  # 5.0 seconds at 30 fps
    "TRACK_THRESH": 0.3,
    "MATCH_THRESH": 0.3,
    "SNAPSHOT_QUALITY": 95,
    "MAX_WORKER_DISTANCE": 100,
    "TARGET_RESOLUTION": (384, 384)
}

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)
        if device.type == "cuda":
            model.model.half()
        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):
    target_res = CONFIG["TARGET_RESOLUTION"]
    frame = cv2.resize(frame, target_res, interpolation=cv2.INTER_LINEAR)
    frame = cv2.convertScaleAbs(frame, alpha=1.2, beta=20)
    return frame

def draw_detections(frame, detections):
    result_frame = frame.copy()
    
    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))
        
        cv2.rectangle(result_frame, (x1, y1), (x2, y2), color, 3)
        
        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)
        
        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)
        
    return result_frame

def calculate_safety_score(violations):
    penalties = {
        "no_helmet": 25,
        "no_harness": 30,
        "unsafe_posture": 20,
        "unsafe_zone": 35,
        "improper_tool_use": 25
    }
   
    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)
   
    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, output_dir):
    try:
        pdf_filename = f"violations_{int(time.time())}.pdf"
        pdf_path = os.path.join(output_dir, pdf_filename)
        pdf_file = BytesIO()
        c = canvas.Canvas(pdf_file, pagesize=letter)
       
        c.setFont("Helvetica-Bold", 16)
        c.drawString(1 * inch, 10 * inch, "Worksite Safety Violation Report")
       
        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')}")
       
        c.setFont("Helvetica-Bold", 14)
        c.drawString(1 * inch, 8.7 * inch, f"Safety Compliance Score: {score}%")

        y_position = 8.2 * inch
        c.setFont("Helvetica-Bold", 12)
        c.drawString(1 * inch, y_position, "Summary:")
        y_position -= 0.3 * inch
       
        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

        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
           
            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}"
               
                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
               
                if y_position < 1 * inch:
                    c.showPage()
                    c.setFont("Helvetica", 10)
                    y_position = 10 * inch

        c.save()
        pdf_file.seek(0)

        with open(pdf_path, "wb") as f:
            f.write(pdf_file.getvalue())
           
        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():
    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):
    try:
        if not pdf_file:
            logger.error("No PDF file provided for upload")
            return ""
           
        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 ""
           
        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):
    try:
        sf = connect_to_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)
           
            violations_text += f"Worker {worker_id}: {display_name} at {timestamp:.2f}s (Conf: {confidence:.2f})\n"
       
        if not violations_text:
            violations_text = "No violations detected."
           
        pdf_url = f"{CONFIG['PUBLIC_URL_BASE']}{os.path.basename(pdf_path)}" if pdf_path else ""

        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}")
       
        try:
            record = sf.Safety_Video_Report__c.create(record_data)
            logger.info(f"Created Safety_Video_Report__c record: {record['id']}")
        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']}")
           
        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}")
        return "N/A", "Salesforce integration failed."

@tenacity.retry(
    stop=tenacity.stop_after_attempt(3),
    wait=tenacity.wait_fixed(1),
    retry=tenacity.retry_if_exception_type((IOError, OSError)),
    before_sleep=lambda retry_state: logger.info(f"Retrying file access (attempt {retry_state.attempt_number}/3)...")
)
def verify_and_open_video(video_path):
    if not os.path.exists(video_path):
        raise FileNotFoundError(f"Temporary video file not found: {video_path}")
   
    file_size = os.path.getsize(video_path)
    if file_size == 0:
        raise ValueError(f"Temporary video file is empty: {video_path}")
   
    with open(video_path, "rb") as f:
        f.read(1)
   
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError("Could not open video file. Ensure the video format is supported (e.g., MP4) and FFmpeg is installed.")
   
    return cap

def process_video(video_data, temp_dir):
    video_path = None
    output_dir = os.path.join(temp_dir, "output")
    os.makedirs(output_dir, exist_ok=True)
    os.environ['YOLO_CONFIG_DIR'] = temp_dir
    
    try:
        if not video_data:
            raise ValueError("Empty video data provided.")
        
        logger.info(f"Received video data size: {len(video_data)} bytes")
        if len(video_data) == 0:
            raise ValueError("Video data is empty.")

        with tempfile.NamedTemporaryFile(suffix=".mp4", dir=temp_dir, delete=False) as temp_file:
            temp_file.write(video_data)
            temp_file.flush()
            video_path = temp_file.name
        logger.info(f"Video saved to temporary file: {video_path}")

        if not os.path.exists(video_path):
            raise FileNotFoundError(f"Temporary video file not found: {video_path}")
        file_size = os.path.getsize(video_path)
        if file_size == 0:
            raise ValueError(f"Temporary video file is empty: {video_path}")
        logger.info(f"Temporary video file size: {file_size} bytes")

        cap = verify_and_open_video(video_path)
        logger.info(f"Successfully opened video file: {video_path}")

        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}")

        if total_frames <= 0:
            raise ValueError("Video has no frames.")

        tracker = BYTETracker(
            track_thresh=CONFIG["TRACK_THRESH"],
            track_buffer=CONFIG["TRACK_BUFFER"],
            match_thresh=CONFIG["MATCH_THRESH"],
            frame_rate=fps,
            max_distance=CONFIG["MAX_WORKER_DISTANCE"]
        )

        unique_violations = {}
        violation_frames = {}
        violation_confidences = {}
        start_time = time.time()
        frame_skip = CONFIG["FRAME_SKIP"]
        processed_frames = 0
        last_yield_time = start_time
        
        logger.info("First pass: Worker detection and tracking")
        all_workers = set()
        worker_first_seen = {}
        worker_last_seen = {}
        
        while processed_frames < total_frames:
            batch_frames = []
            batch_indices = []
            batch_timestamps = []
            
            for _ in range(CONFIG["BATCH_SIZE"]):
                # Skip frames BEFORE reading to speed up
                for _ in range(frame_skip - 1):
                    if not cap.grab():
                        break
                
                frame_idx = int(cap.get(cv2.CAP_PROP_POS_FRAMES))
                if frame_idx >= total_frames:
                    break
                
                ret, frame = cap.read()
                if not ret:
                    logger.warning(f"Failed to read frame {frame_idx}. Skipping.")
                    break
                
                frame = preprocess_frame(frame)
                timestamp = frame_idx / fps
                
                batch_frames.append(frame)
                batch_indices.append(frame_idx)
                batch_timestamps.append(timestamp)
                processed_frames += 1

            if not batch_frames:
                logger.info("No more frames to process.")
                break

            try:
                batch_frames_np = np.array(batch_frames)
                batch_frames_tensor = torch.from_numpy(batch_frames_np).permute(0, 3, 1, 2).float() / 255.0
                batch_frames_tensor = batch_frames_tensor.to(device)
                if device.type == "cuda":
                    batch_frames_tensor = batch_frames_tensor.half()

                results = model(batch_frames_tensor, device=device, conf=0.1, verbose=False)
            except Exception as e:
                logger.error(f"Model inference failed: {e}")
                raise ValueError(f"Failed to process video frames with YOLO model: {str(e)}")
            finally:
                if device.type == "cuda":
                    torch.cuda.empty_cache()

            current_time = time.time()
            if current_time - last_yield_time > 0.1:
                progress = (processed_frames / total_frames) * 100
                elapsed_time = current_time - start_time
                fps_processed = processed_frames / elapsed_time if elapsed_time > 0 else 0
                yield f"Processing video... {progress:.1f}% complete (Frame {processed_frames}/{total_frames}, {fps_processed:.1f} FPS)", "", "", ""
                last_yield_time = current_time

            for i, (result, frame_idx, timestamp) in enumerate(zip(results, batch_indices, batch_timestamps)):
                boxes = result.boxes
                track_inputs = []
                
                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])
                )
                
                for obj in tracked_objects:
                    tracker_id = obj['id']
                    all_workers.add(tracker_id)
                    
                    if tracker_id not in worker_first_seen:
                        worker_first_seen[tracker_id] = timestamp
                    worker_last_seen[tracker_id] = timestamp
                    
                    label = CONFIG["VIOLATION_LABELS"].get(int(obj['cls']), None)
                    conf = obj['score']
                    
                    if label is None:
                        continue
                    
                    violation_key = (tracker_id, label)
                    
                    if violation_key not in unique_violations or conf > violation_confidences.get(violation_key, 0.0):
                        unique_violations[violation_key] = timestamp
                        violation_frames[violation_key] = frame_idx
                        violation_confidences[violation_key] = conf

        cap.release()
        processing_time = time.time() - start_time
        logger.info(f"Processing complete in {processing_time:.2f}s")
        
        total_workers = len(all_workers)
        logger.info(f"Total unique workers detected: {total_workers}")

        violations = []
        for (worker_id, label), detection_time in unique_violations.items():
            violations.append({
                "worker_id": worker_id,
                "violation": label,
                "timestamp": detection_time,
                "confidence": violation_confidences.get((worker_id, label), 0.0),
                "frame_idx": violation_frames[(worker_id, label)]
            })

        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"
            return

        snapshots = []
        cap = cv2.VideoCapture(video_path)
        for violation in violations:
            frame_idx = violation["frame_idx"]
            cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
            ret, frame = cap.read()
            if not ret:
                logger.warning(f"Failed to read frame {frame_idx} for snapshot.")
                continue

            frame = preprocess_frame(frame)
            frame_tensor = torch.from_numpy(frame).permute(2, 0, 1).float() / 255.0
            frame_tensor = frame_tensor.unsqueeze(0).to(device)
            if device.type == "cuda":
                frame_tensor = frame_tensor.half()

            result = model(frame_tensor, device=device, conf=0.1, verbose=False)[0]
            boxes = result.boxes

            for box in boxes:
                cls = int(box.cls)
                conf = float(box.conf)
                label = CONFIG["VIOLATION_LABELS"].get(cls, None)
                if label == violation["violation"]:
                    violation["confidence"] = round(conf, 2)
                    bbox = box.xywh.cpu().numpy()[0]
                    detection = {
                        "worker_id": violation["worker_id"],
                        "violation": label,
                        "confidence": violation["confidence"],
                        "bounding_box": bbox,
                        "timestamp": violation["timestamp"]
                    }
                    snapshot_frame = frame.copy()
                    snapshot_frame = draw_detections(snapshot_frame, [detection])
                    cv2.putText(
                        snapshot_frame,
                        f"Time: {violation['timestamp']:.2f}s",
                        (10, 30),
                        cv2.FONT_HERSHEY_SIMPLEX,
                        0.7,
                        (255, 255, 255),
                        2
                    )
                    snapshot_filename = f"violation_{label}worker{violation['worker_id']}{int(violation['timestamp']*100)}.jpg"
                    snapshot_path = os.path.join(output_dir, snapshot_filename)
                    cv2.imwrite(
                        snapshot_path,
                        snapshot_frame,
                        [cv2.IMWRITE_JPEG_QUALITY, CONFIG["SNAPSHOT_QUALITY"]]
                    )
                    snapshots.append({
                        "violation": label,
                        "worker_id": violation["worker_id"],
                        "timestamp": violation["timestamp"],
                        "snapshot_path": snapshot_path,
                        "snapshot_url": f"{CONFIG['PUBLIC_URL_BASE']}{snapshot_filename}",
                        "confidence": violation["confidence"]
                    })
                    logger.info(f"Captured snapshot for {label} violation by worker {violation['worker_id']} at {violation['timestamp']:.2f}s")
                    break

        cap.release()

        score = calculate_safety_score(violations)
        pdf_path, pdf_url, pdf_file = generate_violation_pdf(violations, score, output_dir)
        
        record_id, final_pdf_url = push_report_to_salesforce(violations, score, pdf_path, pdf_file)

        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)
        
        violation_table = f"## Total Workers Detected: {total_workers}\n\n"
        violation_table += "| Worker ID | Violation | Time (s) | Confidence |\n"
        violation_table += "|-----------|-----------|----------|------------|\n"
        
        for worker_id, vios in sorted(worker_violations.items()):
            vios.sort(key=lambda x: x.get("violation", ""))
            
            for v in vios:
                display_name = CONFIG["DISPLAY_NAMES"].get(v.get("violation", "Unknown"), "Unknown")
                timestamp = v.get("timestamp", 0.0)
                confidence = v.get("confidence", 0.0)
                violation_table += f"| {worker_id} | {display_name} | {timestamp:.2f} | {confidence:.2f} |\n"

        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"

        if not snapshots_text:
            snapshots_text = "No snapshots captured."

        yield (
            violation_table,
            f"Safety Score: {score}% (Based on {total_workers} workers)",
            snapshots_text,
            final_pdf_url
        )

    except Exception as e:
        logger.error(f"Error processing video: {str(e)}", exc_info=True)
        yield f"Error processing video: {str(e)}", "", "", ""
    finally:
        if video_path and os.path.exists(video_path):
            try:
                os.remove(video_path)
                logger.info(f"Cleaned up temporary video file: {video_path}")
            except Exception as e:
                logger.error(f"Failed to clean up temporary video file {video_path}: {e}")
        if device.type == "cuda":
            torch.cuda.empty_cache()

def gradio_interface(video_file):
    temp_dir = None
    local_video_path = None
    try:
        if not video_file:
            return "No file uploaded.", "", "No file uploaded.", ""
        
        temp_dir = tempfile.mkdtemp(prefix="Ultralytics_")
        logger.info(f"Created temporary directory for video processing: {temp_dir}")

        with open(video_file, "rb") as f:
            video_data = f.read()
        logger.info(f"Read Gradio video file: {video_file}, size: {len(video_data)} bytes")
        
        if len(video_data) == 0:
            return "Uploaded video file is empty.", "", "", ""

        with tempfile.NamedTemporaryFile(suffix=".mp4", dir=temp_dir, delete=False) as temp_file:
            temp_file.write(video_data)
            temp_file.flush()
            local_video_path = temp_file.name
        logger.info(f"Copied Gradio video to local temporary file: {local_video_path}")

        if not FFMPEG_AVAILABLE:
            return "FFmpeg is not available in the environment. Please install FFmpeg to process videos.", "", "", ""

        for status, score, snapshots_text, details_url in process_video(video_data, temp_dir):
            yield status, score, snapshots_text, details_url
            
    except Exception as e:
        logger.error(f"Error in Gradio interface: {e}", exc_info=True)
        yield f"Error: {str(e)}", "", "Error in processing.", ""
    finally:
        if local_video_path and os.path.exists(local_video_path):
            try:
                os.remove(local_video_path)
                logger.info(f"Cleaned up local temporary video file: {local_video_path}")
            except Exception as e:
                logger.error(f"Failed to clean up local temporary video file {local_video_path}: {e}")
        
        if temp_dir and os.path.exists(temp_dir):
            shutil.rmtree(temp_dir, ignore_errors=True)
            logger.info(f"Cleaned up temporary directory: {temp_dir}")
        if device.type == "cuda":
            torch.cuda.empty_cache()

# ========================== # 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="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). The system tracks individual workers and their specific violations.",
    allow_flagging="never"
)

if __name__ == "__main__":
    logger.info("Launching Enhanced Safety Analyzer App...")
    interface.launch()