matrice-analytics 0.1.97__py3-none-any.whl → 0.1.124__py3-none-any.whl

matrice_analytics/post_processing/usecases/crowdflow.py

@@ -0,0 +1,1088 @@
+from typing import Any, Dict, List, Optional
+from dataclasses import asdict
+import time
+import math
+import numpy as np
+from datetime import datetime, timezone
+from collections import defaultdict, deque
+
+from ..core.base import BaseProcessor, ProcessingContext, ProcessingResult, ConfigProtocol, ResultFormat
+from ..utils import (
+    filter_by_confidence,
+    filter_by_categories,
+    apply_category_mapping,
+    count_objects_by_category,
+    count_objects_in_zones,
+    calculate_counting_summary,
+    match_results_structure,
+    bbox_smoothing,
+    BBoxSmoothingConfig,
+    BBoxSmoothingTracker
+)
+from dataclasses import dataclass, field
+from ..core.config import BaseConfig, AlertConfig, ZoneConfig
+
+class TrajectoryCorrector:
+    """
+    Handles Velocity-Fusion logic to correct model orientation errors.
+    Stores history of track centers and applies EMA smoothing.
+    """
+    def __init__(self):
+        # track_id -> { "centers": deque, "smooth_angle": float }
+        self.history = defaultdict(lambda: {
+            "centers": deque(maxlen=10), 
+            "angles": deque(maxlen=5), 
+            "smooth_angle": None        
+        })
+
+    def get_direction_label(self, angle):
+        """
+        Your custom logic for Front/Back/Left/Right
+        """
+        if angle is None: return "unknown"
+        angle = angle % 360
+        if 45 <= angle < 135: return "back"
+        elif 135 <= angle < 225: return "left"
+        elif 225 <= angle < 315: return "front"
+        else: return "right"
+
+    def update_and_get_label(self, track_id, center, raw_angle_deg):
+        """
+        1. Fixes Angle (+90)
+        2. Calculates Velocity
+        3. Applies EMA Smoothing
+        4. Returns (Smooth_Angle, Label_String)
+        """
+        state = self.history[track_id]
+        state["centers"].append(center)
+
+        # --- FIX 1: ROTATE MODEL ANGLE ---
+        if raw_angle_deg is None: raw_angle_deg = 0.0
+        fixed_raw_angle = (raw_angle_deg + 90) % 360
+        state["angles"].append(fixed_raw_angle)
+
+        # --- FIX 2: CALCULATE VELOCITY ---
+        motion_angle = self._compute_motion_angle(state["centers"])
+        
+        # Decide Target Angle
+        if motion_angle is not None:
+            target_angle = motion_angle
+        elif fixed_raw_angle is not None:
+            target_angle = fixed_raw_angle
+        elif state["smooth_angle"] is not None:
+            target_angle = state["smooth_angle"]
+        else:
+            target_angle = 0.0
+
+        # --- FIX 3: EMA SMOOTHING ---
+        alpha = 0.2  
+        
+        if state["smooth_angle"] is None:
+            state["smooth_angle"] = target_angle
+        else:
+            prev_rad = math.radians(state["smooth_angle"])
+            curr_rad = math.radians(target_angle)
+            
+            new_sin = (1 - alpha) * math.sin(prev_rad) + alpha * math.sin(curr_rad)
+            new_cos = (1 - alpha) * math.cos(prev_rad) + alpha * math.cos(curr_rad)
+            
+            state["smooth_angle"] = math.degrees(math.atan2(new_sin, new_cos)) % 360
+
+        final_angle = state["smooth_angle"]
+        label = self.get_direction_label(final_angle)
+
+        return final_angle, label
+
+    def _compute_motion_angle(self, centers):
+        if len(centers) < 2: 
+            return None
+            
+        # Look back 5 frames for stability
+        lookback = min(len(centers), 5)
+        (x_past, y_past) = centers[-lookback]
+        (x_now, y_now) = centers[-1]
+        
+        dx = x_now - x_past
+        dy = y_now - y_past
+        
+        # THRESHOLD: 2.5 pixels
+        if math.hypot(dx, dy) < 0.5:
+            return None
+            
+        return math.degrees(math.atan2(-dy, dx)) % 360
+    
+
+
+@dataclass
+class CrowdflowConfig(BaseConfig):
+    """Configuration for footfall use case."""
+    
+    # Smoothing configuration
+    enable_smoothing: bool = False
+    smoothing_algorithm: str = "observability"  # "window" or "observability"
+    smoothing_window_size: int = 20
+    smoothing_cooldown_frames: int = 5
+    smoothing_confidence_range_factor: float = 0.5
+    
+    # Zone configuration
+    zone_config: Optional[ZoneConfig] = None
+    
+    # Counting parameters
+    enable_unique_counting: bool = True
+    time_window_minutes: int = 60
+    
+    # Category mapping
+    person_categories: List[str] = field(default_factory=lambda: ["person"])
+    index_to_category: Optional[Dict[int, str]] = None
+    
+    # Alert configuration
+    alert_config: Optional[AlertConfig] = None
+
+    target_categories: List[str] = field(
+        default_factory=lambda: ['person']
+
+    )
+    
+    def validate(self) -> List[str]:
+        """Validate people counting configuration."""
+        errors = super().validate()
+        
+        if self.time_window_minutes <= 0:
+            errors.append("time_window_minutes must be positive")
+        
+        if not self.person_categories:
+            errors.append("person_categories cannot be empty")
+        
+        # Validate nested configurations
+        if self.zone_config:
+            errors.extend(self.zone_config.validate())
+        
+        if self.alert_config:
+            errors.extend(self.alert_config.validate())
+        
+        return errors
+
+
+
+class CrowdflowUseCase(BaseProcessor):
+    # CATEGORY_DISPLAY = {
+    #     "person": "Person"
+    # }
+
+    def __init__(self):
+        super().__init__("crowdflow")
+        self.category = "retail"
+        self.CASE_TYPE: Optional[str] = 'crowdflow'
+        self.CASE_VERSION: Optional[str] = '1.1'
+        self.target_categories = ['person']
+        self.smoothing_tracker = None
+        self.tracker = None
+
+        self.GRID_SIZE = 180
+        self.MIN_TRACKS_PER_GRID = 6
+
+        self._track_trajectories = defaultdict(lambda: deque(maxlen=30))
+
+        self._grid_trajectories = defaultdict(
+            lambda: defaultdict(lambda: deque(maxlen=30))
+        )
+        # -------------------------------
+        self._flow_grid = defaultdict(lambda: {
+            "sum_sin": 0.0,
+            "sum_cos": 0.0,
+            "count": 0,
+
+            "frame_tracks": {
+                "front": set(),
+                "back": set(),
+                "left": set(),
+                "right": set()
+            },
+
+            "window_tracks": {
+                "front": set(),
+                "back": set(),
+                "left": set(),
+                "right": set()
+            }
+        })
+
+
+        # Initialize the Velocity Logic 
+        self.trajectory_corrector = TrajectoryCorrector()
+        self._total_frame_counter = 0
+        self._global_frame_offset = 0
+        self._tracking_start_time = None
+        self._track_aliases: Dict[Any, Any] = {}
+        self._canonical_tracks: Dict[Any, Dict[str, Any]] = {}
+        self._track_merge_iou_threshold: float = 0.05
+        self._track_merge_time_window: float = 7.0
+        self._ascending_alert_list: List[int] = []
+        self.current_incident_end_timestamp: str = "N/A"
+        self.start_timer = None
+
+    
+    def _compute_avg_displacement(self, trajectories, min_len=6):
+        starts, ends = [], []
+
+        for t in trajectories:
+            if len(t) >= min_len:
+                starts.append(t[0])
+                ends.append(t[-1])
+
+        if not starts:
+            return None
+
+        sx = int(sum(p[0] for p in starts) / len(starts))
+        sy = int(sum(p[1] for p in starts) / len(starts))
+        ex = int(sum(p[0] for p in ends) / len(ends))
+        ey = int(sum(p[1] for p in ends) / len(ends))
+
+        dx, dy = ex - sx, ey - sy
+        angle = math.degrees(math.atan2(-dy, dx)) % 360
+
+        return {
+            "start": (sx, sy),
+            "end": (ex, ey),
+            "angle": angle,
+            "track_count": len(starts)
+        }
+
+    def process(self, data: Any, config: ConfigProtocol, context: Optional[ProcessingContext] = None,
+                stream_info: Optional[Dict[str, Any]] = None) -> ProcessingResult:
+        processing_start = time.time()
+        if not isinstance(config, CrowdflowConfig):
+            return self.create_error_result("Invalid config type", usecase=self.name, category=self.category, context=context)
+        if context is None:
+            context = ProcessingContext()
+
+        fps = 25
+        if stream_info:
+            fps = stream_info.get("input_settings", {}).get("original_fps", 25)
+
+        decay_interval = int(2 * fps)
+
+        # Temporal decay of window tracks (~2 sec behavior)
+        if self._total_frame_counter % decay_interval == 0:
+            for cell in self._flow_grid.values():
+                for d in cell["window_tracks"]:
+                    cell["window_tracks"][d] = {
+                        tid for tid in cell["window_tracks"][d]
+                        if np.random.rand() < 0.90
+                    }
+
+
+
+        # Reset per-frame snapshot FIRST
+        for cell in self._flow_grid.values():
+            for d in cell["frame_tracks"]:
+                cell["frame_tracks"][d].clear()
+
+        # Then apply temporal decay
+        for cell in self._flow_grid.values():
+            cell["sum_sin"] *= 0.92
+            cell["sum_cos"] *= 0.92
+            cell["count"] *= 0.92
+
+
+
+        input_format = match_results_structure(data)
+        context.input_format = input_format
+        context.confidence_threshold = config.confidence_threshold
+
+        # ... [Keep your standard filtering logic here: confidence, mapping, categories] ...
+        if config.confidence_threshold is not None:
+            processed_data = filter_by_confidence(data, config.confidence_threshold)
+        else:
+            processed_data = data
+
+        if config.index_to_category:
+            processed_data = apply_category_mapping(processed_data, config.index_to_category)
+
+        if config.target_categories:
+            processed_data = [d for d in processed_data if d.get('category') in self.target_categories]
+
+        try:
+            from ..advanced_tracker import AdvancedTracker
+            from ..advanced_tracker.config import TrackerConfig
+            if self.tracker is None:
+                tracker_config = TrackerConfig(
+                    track_high_thresh=0.4,
+                    track_low_thresh=0.05,
+                    new_track_thresh=0.3,
+                    match_thresh=0.8)
+                self.tracker = AdvancedTracker(tracker_config)
+                self.logger.info("Initialized AdvancedTracker for People Counting")
+            
+            # 1. Run Standard Tracker (Assigns IDs)
+            processed_data = self.tracker.update(processed_data)
+            
+            # =========================================================
+            # NEW: INJECT VELOCITY FUSION LOGIC (CORRECTED)
+            # =========================================================
+            for det in processed_data:
+                track_id = det.get("track_id")
+
+                # STREAM-SAFE bbox normalization
+                bbox = det.get("bbox") or det.get("bounding_box")
+
+                if isinstance(bbox, dict):
+                    bbox = [
+                        bbox.get("xmin"),
+                        bbox.get("ymin"),
+                        bbox.get("xmax"),
+                        bbox.get("ymax"),
+                    ]
+
+                # Hard safety guard
+                if not bbox or len(bbox) < 4:
+                    continue
+
+                
+                # Check for 'raw_angle' (from predict.py) or 'orientation'
+                raw_angle = det.get("angle", det.get("raw_angle", det.get("orientation", 0.0)))
+                
+                if track_id is not None and bbox:
+                    # Calculate Center (cx, cy)
+                    cx = int((bbox[0] + bbox[2]) / 2)
+                    cy = int((bbox[1] + bbox[3]) / 2)
+                    
+                    # Run Correction (Velocity + EMA + 90 Fix)
+                    # FIX: Unpack both values (Angle AND Label)
+                    final_angle, direction_label = self.trajectory_corrector.update_and_get_label(
+                        track_id, 
+                        (cx, cy), 
+                        raw_angle
+                    )
+                    
+                    # OVERWRITE the detection angle
+                    det["orientation"] = final_angle  # For UI
+                    det["angle"] = final_angle        # For Analytics
+                    
+                    # FIX: SAVE THE DIRECTION LABEL
+                    det["direction"] = direction_label # "front", "back", etc.
+
+                    # -------------------------------
+                    # Crowd Flow Trajectory Capture
+                    # -------------------------------
+                    gx = int(cx // self.GRID_SIZE)
+                    gy = int(cy // self.GRID_SIZE)
+
+                    self._track_trajectories[track_id].append((cx, cy))
+                    self._grid_trajectories[(gx, gy)][track_id].append((cx, cy))
+
+
+                    cell = self._flow_grid[(gx, gy)]
+
+                    rad = math.radians(final_angle)
+                    cell["sum_cos"] += math.cos(rad)
+                    cell["sum_sin"] += math.sin(rad)
+                    cell["count"] += 1
+
+                    cell["frame_tracks"][direction_label].add(track_id)
+                    cell["window_tracks"][direction_label].add(track_id)
+
+
+            # =========================================================
+
+        except Exception as e:
+            self.logger.warning(f"AdvancedTracker/Velocity failed: {e}")
+
+        # ... [The rest of your process method remains exactly the same] ...
+        # CLEAN DEAD TRACKS (CRITICAL – prevents memory leak)
+        # ---------------------------------------------------------
+        alive_ids = {
+            d.get("track_id")
+            for d in processed_data
+            if d.get("track_id") is not None
+        }
+
+        # Clean grid-level trajectories
+        for grid_key in list(self._grid_trajectories.keys()):
+            for tid in list(self._grid_trajectories[grid_key].keys()):
+                if tid not in alive_ids:
+                    del self._grid_trajectories[grid_key][tid]
+
+        # Clean global trajectory memory
+        for tid in list(self._track_trajectories.keys()):
+            if tid not in alive_ids:
+                self._track_trajectories.pop(tid, None)
+    
+
+        self._update_tracking_state(processed_data)
+
+        frame_number = None
+        if stream_info:
+            input_settings = stream_info.get("input_settings", {})
+            start_frame = input_settings.get("start_frame")
+            end_frame = input_settings.get("end_frame")
+            if start_frame is not None and end_frame is not None and start_frame == end_frame:
+                frame_number = start_frame
+
+        general_counting_summary = calculate_counting_summary(data)
+        counting_summary = self._count_categories(processed_data, config)
+        total_counts = self.get_total_counts()
+        counting_summary['total_counts'] = total_counts
+
+        alerts = self._check_alerts(counting_summary, frame_number, config)
+        predictions = self._extract_predictions(processed_data)
+
+        incidents_list = self._generate_incidents(counting_summary, alerts, config, frame_number, stream_info)
+        tracking_stats_list = self._generate_tracking_stats(counting_summary, alerts, config, frame_number, stream_info)
+        business_analytics_list = self._generate_business_analytics(counting_summary, alerts, config, stream_info, is_empty=True)
+        summary_list = self._generate_summary(counting_summary, incidents_list, tracking_stats_list, business_analytics_list, alerts)
+
+        incidents = incidents_list[0] if incidents_list else {}
+        tracking_stats = tracking_stats_list[0] if tracking_stats_list else {}
+        business_analytics = business_analytics_list[0] if business_analytics_list else {}
+        summary = summary_list[0] if summary_list else {}
+        agg_summary = {str(frame_number): {
+            "incidents": incidents,
+            "tracking_stats": tracking_stats,
+            "business_analytics": business_analytics,
+            "alerts": alerts,
+            "human_text": summary}
+        }
+
+        context.mark_completed()
+        result = self.create_result(
+            data={"agg_summary": agg_summary},
+            usecase=self.name,
+            category=self.category,
+            context=context
+        )
+        proc_time = time.time() - processing_start
+        processing_latency_ms = proc_time * 1000.0
+        processing_fps = (1.0 / proc_time) if proc_time > 0 else None
+        # Log the performance metrics using the module-level logger
+        self._total_frame_counter += 1
+        print("latency in ms:",processing_latency_ms,"| Throughput fps:",processing_fps,"| Frame_Number:",self._total_frame_counter)
+        return result
+
+    def _check_alerts(self, summary: dict, frame_number: Any, config: CrowdflowConfig) -> List[Dict]:
+        def get_trend(data, lookback=900, threshold=0.6):
+            window = data[-lookback:] if len(data) >= lookback else data
+            if len(window) < 2:
+                return True
+            increasing = 0
+            total = 0
+            for i in range(1, len(window)):
+                if window[i] >= window[i - 1]:
+                    increasing += 1
+                total += 1
+            ratio = increasing / total
+            return ratio >= threshold
+
+        frame_key = str(frame_number) if frame_number is not None else "current_frame"
+        alerts = []
+        total_detections = summary.get("total_count", 0)
+        total_counts_dict = summary.get("total_counts", {})
+        per_category_count = summary.get("per_category_count", {})
+
+        if not config.alert_config:
+            return alerts
+
+        if hasattr(config.alert_config, 'count_thresholds') and config.alert_config.count_thresholds:
+            for category, threshold in config.alert_config.count_thresholds.items():
+                if category == "all" and total_detections > threshold:
+                    alerts.append({
+                        "alert_type": getattr(config.alert_config, 'alert_type', ['Default']),
+                        "alert_id": f"alert_{category}_{frame_key}",
+                        "incident_category": self.CASE_TYPE,
+                        "threshold_level": threshold,
+                        "ascending": get_trend(self._ascending_alert_list, lookback=900, threshold=0.8),
+                        "settings": {t: v for t, v in zip(getattr(config.alert_config, 'alert_type', ['Default']),
+                                                         getattr(config.alert_config, 'alert_value', ['JSON']))}
+                    })
+                elif category in per_category_count and per_category_count[category] > threshold:
+                    alerts.append({
+                        "alert_type": getattr(config.alert_config, 'alert_type', ['Default']),
+                        "alert_id": f"alert_{category}_{frame_key}",
+                        "incident_category": self.CASE_TYPE,
+                        "threshold_level": threshold,
+                        "ascending": get_trend(self._ascending_alert_list, lookback=900, threshold=0.8),
+                        "settings": {t: v for t, v in zip(getattr(config.alert_config, 'alert_type', ['Default']),
+                                                         getattr(config.alert_config, 'alert_value', ['JSON']))}
+                    })
+        return alerts
+
+    def _generate_incidents(self, counting_summary: Dict, alerts: List, config: CrowdflowConfig,
+                           frame_number: Optional[int] = None, stream_info: Optional[Dict[str, Any]] = None) -> List[Dict]:
+        incidents = []
+        total_detections = counting_summary.get("total_count", 0)
+        current_timestamp = self._get_current_timestamp_str(stream_info)
+        camera_info = self.get_camera_info_from_stream(stream_info)
+
+        self._ascending_alert_list = self._ascending_alert_list[-900:] if len(self._ascending_alert_list) > 900 else self._ascending_alert_list
+
+        if total_detections > 0:
+            level = "low"
+            intensity = 5.0
+            start_timestamp = self._get_start_timestamp_str(stream_info)
+            if start_timestamp and self.current_incident_end_timestamp == 'N/A':
+                self.current_incident_end_timestamp = 'Incident still active'
+            elif start_timestamp and self.current_incident_end_timestamp == 'Incident still active':
+                if len(self._ascending_alert_list) >= 15 and sum(self._ascending_alert_list[-15:]) / 15 < 1.5:
+                    self.current_incident_end_timestamp = current_timestamp
+            elif self.current_incident_end_timestamp != 'Incident still active' and self.current_incident_end_timestamp != 'N/A':
+                self.current_incident_end_timestamp = 'N/A'
+
+            if config.alert_config and config.alert_config.count_thresholds:
+                threshold = config.alert_config.count_thresholds.get("all", 15)
+                intensity = min(10.0, (total_detections / threshold) * 10)
+                if intensity >= 9:
+                    level = "critical"
+                    self._ascending_alert_list.append(3)
+                elif intensity >= 7:
+                    level = "significant"
+                    self._ascending_alert_list.append(2)
+                elif intensity >= 5:
+                    level = "medium"
+                    self._ascending_alert_list.append(1)
+                else:
+                    level = "low"
+                    self._ascending_alert_list.append(0)
+            else:
+                if total_detections > 30:
+                    level = "critical"
+                    intensity = 10.0
+                    self._ascending_alert_list.append(3)
+                elif total_detections > 25:
+                    level = "significant"
+                    intensity = 9.0
+                    self._ascending_alert_list.append(2)
+                elif total_detections > 15:
+                    level = "medium"
+                    intensity = 7.0
+                    self._ascending_alert_list.append(1)
+                else:
+                    level = "low"
+                    intensity = min(10.0, total_detections / 3.0)
+                    self._ascending_alert_list.append(0)
+
+            human_text_lines = [f"COUNTING INCIDENTS DETECTED @ {current_timestamp}:"]
+            human_text_lines.append(f"\tSeverity Level: {(self.CASE_TYPE, level)}")
+            human_text = "\n".join(human_text_lines)
+
+            alert_settings = []
+            if config.alert_config and hasattr(config.alert_config, 'alert_type'):
+                alert_settings.append({
+                    "alert_type": getattr(config.alert_config, 'alert_type', ['Default']),
+                    "incident_category": self.CASE_TYPE,
+                    "threshold_level": config.alert_config.count_thresholds if hasattr(config.alert_config, 'count_thresholds') else {},
+                    "ascending": True,
+                    "settings": {t: v for t, v in zip(getattr(config.alert_config, 'alert_type', ['Default']),
+                                                     getattr(config.alert_config, 'alert_value', ['JSON']))}
+                })
+
+            event = self.create_incident(
+                incident_id=f"{self.CASE_TYPE}_{frame_number}",
+                incident_type=self.CASE_TYPE,
+                severity_level=level,
+                human_text=human_text,
+                camera_info=camera_info,
+                alerts=alerts,
+                alert_settings=alert_settings,
+                start_time=start_timestamp,
+                end_time=self.current_incident_end_timestamp,
+                level_settings={"low": 1, "medium": 3, "significant": 4, "critical": 7}
+            )
+            incidents.append(event)
+        else:
+            self._ascending_alert_list.append(0)
+            incidents.append({})
+        return incidents
+
+    def _generate_tracking_stats(self, counting_summary: Dict, alerts: List, config: CrowdflowConfig,
+                                frame_number: Optional[int] = None, stream_info: Optional[Dict[str, Any]] = None) -> List[Dict]:
+        camera_info = self.get_camera_info_from_stream(stream_info)
+        tracking_stats = []
+        total_detections = counting_summary.get("total_count", 0)
+        total_counts_dict = counting_summary.get("total_counts", {})
+        per_category_count = counting_summary.get("per_category_count", {})
+        current_timestamp = self._get_current_timestamp_str(stream_info, precision=False)
+        start_timestamp = self._get_start_timestamp_str(stream_info, precision=False)
+        high_precision_start_timestamp = self._get_current_timestamp_str(stream_info, precision=True)
+        high_precision_reset_timestamp = self._get_start_timestamp_str(stream_info, precision=True)
+
+        total_counts = [{"category": cat, "count": count} for cat, count in total_counts_dict.items() if count > 0]
+        current_counts = [{"category": cat, "count": count} for cat, count in per_category_count.items() if count > 0 or total_detections > 0]
+
+        detections = []
+        for detection in counting_summary.get("detections", []):
+            bbox = detection.get("bounding_box", {})
+            category = detection.get("category", "person")
+            if detection.get("masks"):
+                segmentation = detection.get("masks", [])
+                detection_obj = self.create_detection_object(category, bbox, segmentation=segmentation)
+            elif detection.get("segmentation"):
+                segmentation = detection.get("segmentation")
+                detection_obj = self.create_detection_object(category, bbox, segmentation=segmentation)
+            elif detection.get("mask"):
+                segmentation = detection.get("mask")
+                detection_obj = self.create_detection_object(category, bbox, segmentation=segmentation)
+            else:
+                detection_obj = self.create_detection_object(category, bbox)
+            detections.append(detection_obj)
+
+        alert_settings = []
+        if config.alert_config and hasattr(config.alert_config, 'alert_type'):
+            alert_settings.append({
+                "alert_type": getattr(config.alert_config, 'alert_type', ['Default']),
+                "incident_category": self.CASE_TYPE,
+                "threshold_level": config.alert_config.count_thresholds if hasattr(config.alert_config, 'count_thresholds') else {},
+                "ascending": True,
+                "settings": {t: v for t, v in zip(getattr(config.alert_config, 'alert_type', ['Default']),
+                                                 getattr(config.alert_config, 'alert_value', ['JSON']))}
+            })
+
+        human_text_lines = []
+        human_text_lines.append(f"CURRENT FRAME @ {current_timestamp}:")
+        for cat, count in per_category_count.items():
+            human_text_lines.append(f"\t- People Detected: {count}")
+        human_text_lines.append("")
+        # human_text_lines.append(f"TOTAL SINCE {start_timestamp}")
+        # for cat, count in total_counts_dict.items():
+        #     if count > 0:
+        #         human_text_lines.append("")
+        #         human_text_lines.append(f"\t- Total unique people count: {count}")
+        # if alerts:
+        #     for alert in alerts:
+        #         human_text_lines.append(f"Alerts: {alert.get('settings', {})} sent @ {current_timestamp}")
+        # else:
+        #     human_text_lines.append("Alerts: None")
+        human_text = "\n".join(human_text_lines)
+
+        reset_settings = [{"interval_type": "daily", "reset_time": {"value": 9, "time_unit": "hour"}}]
+        tracking_stat = self.create_tracking_stats(
+            total_counts=total_counts,
+            current_counts=current_counts,
+            detections=detections,
+            human_text=human_text,
+            camera_info=camera_info,
+            alerts=alerts,
+            alert_settings=alert_settings,
+            reset_settings=reset_settings,
+            start_time=high_precision_start_timestamp,
+            reset_time=high_precision_reset_timestamp
+        )
+        # -------------------------------
+        # Crowd Flow Aggregation (GRID)
+        # -------------------------------
+        crowd_flow = []
+
+        for (gx, gy), tracks in self._grid_trajectories.items():
+            avg = self._compute_avg_displacement(tracks.values())
+            if not avg or avg["track_count"] < self.MIN_TRACKS_PER_GRID:
+                continue
+
+            cell = self._flow_grid[(gx, gy)]
+
+
+            front = len(cell["window_tracks"]["front"])
+            back  = len(cell["window_tracks"]["back"])
+            left  = len(cell["window_tracks"]["left"])
+            right = len(cell["window_tracks"]["right"])
+
+            conflict = (
+                (front >= self.MIN_TRACKS_PER_GRID and back >= self.MIN_TRACKS_PER_GRID) or
+                (left  >= self.MIN_TRACKS_PER_GRID and right >= self.MIN_TRACKS_PER_GRID)
+            )
+
+            
+            avg_direction_angle = None
+            if cell["count"] >= self.MIN_TRACKS_PER_GRID:
+                avg_direction_angle = (
+                    math.degrees(math.atan2(cell["sum_sin"], cell["sum_cos"])) % 360
+                )
+
+
+
+            crowd_flow.append({
+                "grid_x": gx,
+                "grid_y": gy,
+                "conflict": conflict,
+
+                # 🔵 BLUE: average displacement (trajectory-based)
+                "avg_displacement": {
+                    "start": avg["start"],
+                    "end": avg["end"],
+                    "angle": avg["angle"],
+                    "track_count": avg["track_count"]
+                },
+
+                # 🟢 GREEN: directional flow aggregation
+                "direction_flow": {
+                    "avg_angle": avg_direction_angle,
+                    "counts": {
+                        "front": len(cell["frame_tracks"]["front"]),
+                        "back": len(cell["frame_tracks"]["back"]),
+                        "left": len(cell["frame_tracks"]["left"]),
+                        "right": len(cell["frame_tracks"]["right"]),
+                    }
+                }
+            })
+
+
+
+        tracking_stat["crowd_flow"] = crowd_flow
+
+        tracking_stat['target_categories'] = self.target_categories
+        tracking_stats.append(tracking_stat)
+        return tracking_stats
+
+    def _generate_business_analytics(self, counting_summary: Dict, alerts: Any, config: CrowdflowConfig,
+                                    stream_info: Optional[Dict[str, Any]] = None, is_empty=False) -> List[Dict]:
+        if is_empty:
+            return []
+
+    def _generate_summary(self, summary: dict, incidents: List, tracking_stats: List, business_analytics: List, alerts: List) -> List[str]:
+        """
+        Generate a human_text string for the tracking_stat, incident, business analytics and alerts.
+        """
+        lines = []
+        lines.append("Application Name: "+self.CASE_TYPE)
+        lines.append("Application Version: "+self.CASE_VERSION)
+        # if len(incidents) > 0:
+        #     lines.append("Incidents: "+f"\n\t{incidents[0].get('human_text', 'No incidents detected')}")
+        if len(tracking_stats) > 0:
+            lines.append("Tracking Statistics: "+f"\t{tracking_stats[0].get('human_text', 'No tracking statistics detected')}")
+        if len(business_analytics) > 0:
+            lines.append("Business Analytics: "+f"\t{business_analytics[0].get('human_text', 'No business analytics detected')}")
+
+        if len(incidents) == 0 and len(tracking_stats) == 0 and len(business_analytics) == 0:
+            lines.append("Summary: "+"No Summary Data")
+
+        return ["\n".join(lines)]
+
+    def _get_track_ids_info(self, detections: list) -> Dict[str, Any]:
+        frame_track_ids = set()
+        for det in detections:
+            tid = det.get('track_id')
+            if tid is not None:
+                frame_track_ids.add(tid)
+        total_track_ids = set()
+        for s in getattr(self, '_per_category_total_track_ids', {}).values():
+            total_track_ids.update(s)
+        return {
+            "total_count": len(total_track_ids),
+            "current_frame_count": len(frame_track_ids),
+            "total_unique_track_ids": len(total_track_ids),
+            "current_frame_track_ids": list(frame_track_ids),
+            "last_update_time": time.time(),
+            "total_frames_processed": getattr(self, '_total_frame_counter', 0)
+        }
+
+    def _update_tracking_state(self, detections: list):
+        if not hasattr(self, "_per_category_total_track_ids"):
+            self._per_category_total_track_ids = {cat: set() for cat in self.target_categories}
+        self._current_frame_track_ids = {cat: set() for cat in self.target_categories}
+
+        for det in detections:
+            cat = det.get("category")
+            raw_track_id = det.get("track_id")
+            if cat not in self.target_categories or raw_track_id is None:
+                continue
+            bbox = det.get("bbox") or det.get("bounding_box")
+
+            if isinstance(bbox, dict):
+                bbox = [
+                    bbox.get("xmin"),
+                    bbox.get("ymin"),
+                    bbox.get("xmax"),
+                    bbox.get("ymax"),
+                ]
+
+            if not bbox or len(bbox) < 4:
+                continue
+
+            canonical_id = self._merge_or_register_track(raw_track_id, bbox)
+
+            det["track_id"] = canonical_id
+            self._per_category_total_track_ids.setdefault(cat, set()).add(canonical_id)
+            self._current_frame_track_ids[cat].add(canonical_id)
+
+    def get_total_counts(self):
+        return {cat: len(ids) for cat, ids in getattr(self, '_per_category_total_track_ids', {}).items()}
+
+    def _format_timestamp_for_stream(self, timestamp: float) -> str:
+        dt = datetime.fromtimestamp(timestamp, tz=timezone.utc)
+        return dt.strftime('%Y:%m:%d %H:%M:%S')
+
+    def _format_timestamp_for_video(self, timestamp: float) -> str:
+        hours = int(timestamp // 3600)
+        minutes = int((timestamp % 3600) // 60)
+        seconds = round(float(timestamp % 60), 2)
+        return f"{hours:02d}:{minutes:02d}:{seconds:.1f}"
+
+    def _format_timestamp(self, timestamp: Any) -> str:
+        """Format a timestamp to match the current timestamp format: YYYY:MM:DD HH:MM:SS.
+
+        The input can be either:
+        1. A numeric Unix timestamp (``float`` / ``int``) – it will be converted to datetime.
+        2. A string in the format ``YYYY-MM-DD-HH:MM:SS.ffffff UTC``.
+
+        The returned value will be in the format: YYYY:MM:DD HH:MM:SS (no milliseconds, no UTC suffix).
+
+        Example
+        -------
+        >>> self._format_timestamp("2025-10-27-19:31:20.187574 UTC")
+        '2025:10:27 19:31:20'
+        """
+
+        # Convert numeric timestamps to datetime first
+        if isinstance(timestamp, (int, float)):
+            dt = datetime.fromtimestamp(timestamp, timezone.utc)
+            return dt.strftime('%Y:%m:%d %H:%M:%S')
+
+        # Ensure we are working with a string from here on
+        if not isinstance(timestamp, str):
+            return str(timestamp)
+
+        # Remove ' UTC' suffix if present
+        timestamp_clean = timestamp.replace(' UTC', '').strip()
+
+        # Remove milliseconds if present (everything after the last dot)
+        if '.' in timestamp_clean:
+            timestamp_clean = timestamp_clean.split('.')[0]
+
+        # Parse the timestamp string and convert to desired format
+        try:
+            # Handle format: YYYY-MM-DD-HH:MM:SS
+            if timestamp_clean.count('-') >= 2:
+                # Replace first two dashes with colons for date part, third with space
+                parts = timestamp_clean.split('-')
+                if len(parts) >= 4:
+                    # parts = ['2025', '10', '27', '19:31:20']
+                    formatted = f"{parts[0]}:{parts[1]}:{parts[2]} {'-'.join(parts[3:])}"
+                    return formatted
+        except Exception:
+            pass
+
+        # If parsing fails, return the cleaned string as-is
+        return timestamp_clean
+
+    def _get_current_timestamp_str(self, stream_info: Optional[Dict[str, Any]], precision=False, frame_id: Optional[str]=None) -> str:
+        """Get formatted current timestamp based on stream type."""
+        
+        if not stream_info:
+            return "00:00:00.00"
+        if precision:
+            if stream_info.get("input_settings", {}).get("start_frame", "na") != "na":
+                if frame_id:
+                    start_time = int(frame_id)/stream_info.get("input_settings", {}).get("original_fps", 30)
+                else:
+                    start_time = stream_info.get("input_settings", {}).get("start_frame", 30)/stream_info.get("input_settings", {}).get("original_fps", 30)
+                stream_time_str = self._format_timestamp_for_video(start_time)
+                
+                return self._format_timestamp(stream_info.get("input_settings", {}).get("stream_time", "NA"))
+            else:
+                return datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+
+        if stream_info.get("input_settings", {}).get("start_frame", "na") != "na":
+            if frame_id:
+                start_time = int(frame_id)/stream_info.get("input_settings", {}).get("original_fps", 30)
+            else:
+                start_time = stream_info.get("input_settings", {}).get("start_frame", 30)/stream_info.get("input_settings", {}).get("original_fps", 30)
+
+            stream_time_str = self._format_timestamp_for_video(start_time)
+           
+
+            return self._format_timestamp(stream_info.get("input_settings", {}).get("stream_time", "NA"))
+        else:
+            stream_time_str = stream_info.get("input_settings", {}).get("stream_info", {}).get("stream_time", "")
+            if stream_time_str:
+                try:
+                    timestamp_str = stream_time_str.replace(" UTC", "")
+                    dt = datetime.strptime(timestamp_str, "%Y-%m-%d-%H:%M:%S.%f")
+                    timestamp = dt.replace(tzinfo=timezone.utc).timestamp()
+                    return self._format_timestamp_for_stream(timestamp)
+                except:
+                    return self._format_timestamp_for_stream(time.time())
+            else:
+                return self._format_timestamp_for_stream(time.time())
+
+    def _get_start_timestamp_str(self, stream_info: Optional[Dict[str, Any]], precision=False) -> str:
+        """Get formatted start timestamp for 'TOTAL SINCE' based on stream type."""
+        if not stream_info:
+            return "00:00:00"
+        
+        if precision:
+            if self.start_timer is None:
+                candidate = stream_info.get("input_settings", {}).get("stream_time")
+                if not candidate or candidate == "NA":
+                    candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                self.start_timer = candidate
+                return self._format_timestamp(self.start_timer)
+            elif stream_info.get("input_settings", {}).get("start_frame", "na") == 1:
+                candidate = stream_info.get("input_settings", {}).get("stream_time")
+                if not candidate or candidate == "NA":
+                    candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                self.start_timer = candidate
+                return self._format_timestamp(self.start_timer)
+            else:
+                return self._format_timestamp(self.start_timer)
+
+        if self.start_timer is None:
+            # Prefer direct input_settings.stream_time if available and not NA
+            candidate = stream_info.get("input_settings", {}).get("stream_time")
+            if not candidate or candidate == "NA":
+                # Fallback to nested stream_info.stream_time used by current timestamp path
+                stream_time_str = stream_info.get("input_settings", {}).get("stream_info", {}).get("stream_time", "")
+                if stream_time_str:
+                    try:
+                        timestamp_str = stream_time_str.replace(" UTC", "")
+                        dt = datetime.strptime(timestamp_str, "%Y-%m-%d-%H:%M:%S.%f")
+                        self._tracking_start_time = dt.replace(tzinfo=timezone.utc).timestamp()
+                        candidate = datetime.fromtimestamp(self._tracking_start_time, timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                    except:
+                        candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                else:
+                    candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+            self.start_timer = candidate
+            return self._format_timestamp(self.start_timer)
+        elif stream_info.get("input_settings", {}).get("start_frame", "na") == 1:
+            candidate = stream_info.get("input_settings", {}).get("stream_time")
+            if not candidate or candidate == "NA":
+                stream_time_str = stream_info.get("input_settings", {}).get("stream_info", {}).get("stream_time", "")
+                if stream_time_str:
+                    try:
+                        timestamp_str = stream_time_str.replace(" UTC", "")
+                        dt = datetime.strptime(timestamp_str, "%Y-%m-%d-%H:%M:%S.%f")
+                        ts = dt.replace(tzinfo=timezone.utc).timestamp()
+                        candidate = datetime.fromtimestamp(ts, timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                    except:
+                        candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+                else:
+                    candidate = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H:%M:%S.%f UTC")
+            self.start_timer = candidate
+            return self._format_timestamp(self.start_timer)
+        
+        else:
+            if self.start_timer is not None and self.start_timer != "NA":
+                return self._format_timestamp(self.start_timer)
+
+            if self._tracking_start_time is None:
+                stream_time_str = stream_info.get("input_settings", {}).get("stream_info", {}).get("stream_time", "")
+                if stream_time_str:
+                    try:
+                        timestamp_str = stream_time_str.replace(" UTC", "")
+                        dt = datetime.strptime(timestamp_str, "%Y-%m-%d-%H:%M:%S.%f")
+                        self._tracking_start_time = dt.replace(tzinfo=timezone.utc).timestamp()
+                    except:
+                        self._tracking_start_time = time.time()
+                else:
+                    self._tracking_start_time = time.time()
+
+            dt = datetime.fromtimestamp(self._tracking_start_time, tz=timezone.utc)
+            dt = dt.replace(minute=0, second=0, microsecond=0)
+            return dt.strftime('%Y:%m:%d %H:%M:%S')
+
+    def _count_categories(self, detections: list, config: CrowdflowConfig) -> dict:
+        counts = {}
+        for det in detections:
+            cat = det.get("direction") or "unknown"
+            counts[cat] = counts.get(cat, 0) + 1
+        return {
+            "total_count": sum(counts.values()),
+            "per_category_count": counts,
+            "detections": [
+                {
+                    "bounding_box": det.get("bounding_box"),
+                    "category": det.get("category"),
+                    "direction": det.get("direction"),
+                    "confidence": det.get("confidence"),
+                    "track_id": det.get("track_id"),
+                    "frame_id": det.get("frame_id"),
+                    "angle": det.get("angle"),          
+                    "orientation": det.get("orientation") #for UI arrows
+                }
+                for det in detections
+            ]
+        }
+
+    def _extract_predictions(self, detections: list) -> List[Dict[str, Any]]:
+        return [
+            {
+                "category": det.get("category", "unknown"),
+                "confidence": det.get("confidence", 0.0),
+                "bounding_box": det.get("bounding_box", {})
+            }
+            for det in detections
+        ]
+
+    def _compute_iou(self, box1: Any, box2: Any) -> float:
+        def _bbox_to_list(bbox):
+            if bbox is None:
+                return []
+            if isinstance(bbox, list):
+                return bbox[:4] if len(bbox) >= 4 else []
+            if isinstance(bbox, dict):
+                if "xmin" in bbox:
+                    return [bbox["xmin"], bbox["ymin"], bbox["xmax"], bbox["ymax"]]
+                if "x1" in bbox:
+                    return [bbox["x1"], bbox["y1"], bbox["x2"], bbox["y2"]]
+                values = [v for v in bbox.values() if isinstance(v, (int, float))]
+                return values[:4] if len(values) >= 4 else []
+            return []
+
+        l1 = _bbox_to_list(box1)
+        l2 = _bbox_to_list(box2)
+        if len(l1) < 4 or len(l2) < 4:
+            return 0.0
+        x1_min, y1_min, x1_max, y1_max = l1
+        x2_min, y2_min, x2_max, y2_max = l2
+        x1_min, x1_max = min(x1_min, x1_max), max(x1_min, x1_max)
+        y1_min, y1_max = min(y1_min, y1_max), max(y1_min, y1_max)
+        x2_min, x2_max = min(x2_min, x2_max), max(x2_min, x2_max)
+        y2_min, y2_max = min(y2_min, y2_max), max(y2_min, y2_max)
+        inter_x_min = max(x1_min, x2_min)
+        inter_y_min = max(y1_min, y2_min)
+        inter_x_max = min(x1_max, x2_max)
+        inter_y_max = min(y1_max, y2_max)
+        inter_w = max(0.0, inter_x_max - inter_x_min)
+        inter_h = max(0.0, inter_y_max - inter_y_min)
+        inter_area = inter_w * inter_h
+        area1 = (x1_max - x1_min) * (y1_max - y1_min)
+        area2 = (x2_max - x2_min) * (y2_max - y2_min)
+        union_area = area1 + area2 - inter_area
+        return (inter_area / union_area) if union_area > 0 else 0.0
+
+    def _merge_or_register_track(self, raw_id: Any, bbox: Any) -> Any:
+        if raw_id is None or bbox is None:
+            return raw_id
+        now = time.time()
+        if raw_id in self._track_aliases:
+            canonical_id = self._track_aliases[raw_id]
+            track_info = self._canonical_tracks.get(canonical_id)
+            if track_info is not None:
+                track_info["last_bbox"] = bbox
+                track_info["last_update"] = now
+                track_info["raw_ids"].add(raw_id)
+            return canonical_id
+        for canonical_id, info in self._canonical_tracks.items():
+            if now - info["last_update"] > self._track_merge_time_window:
+                continue
+            iou = self._compute_iou(bbox, info["last_bbox"])
+            if iou >= self._track_merge_iou_threshold:
+                self._track_aliases[raw_id] = canonical_id
+                info["last_bbox"] = bbox
+                info["last_update"] = now
+                info["raw_ids"].add(raw_id)
+                return canonical_id
+        canonical_id = raw_id
+        self._track_aliases[raw_id] = canonical_id
+        self._canonical_tracks[canonical_id] = {
+            "last_bbox": bbox,
+            "last_update": now,
+            "raw_ids": {raw_id},
+        }
+        return canonical_id
+
+    def _get_tracking_start_time(self) -> str:
+        if self._tracking_start_time is None:
+            return "N/A"
+        return self._format_timestamp(self._tracking_start_time)
+
+    def _set_tracking_start_time(self) -> None:
+        self._tracking_start_time = time.time()