🎯 AI-Powered Video Analytics System

A production-grade, real-time object detection, multi-object tracking, and zone-based analytics system built with YOLO11 and BotSORT — served through an interactive multi-page Streamlit dashboard.

---

Features

• Real-time object detection — YOLO11 (detection + instance segmentation modes)
• Multi-object tracking — BotSORT with persistent track IDs across frames
• Zone-based analytics — configurable rectangular tracking zones with entry counting
• Three-rule alert system — zone intrusion, loitering detection, crowd alerts with per-track cooldowns
• Professional dashboard — multi-page Streamlit UI with real-time Plotly charts, session state, start/stop controls
• Typed configuration — config.yaml with dataclass validation and env-var overrides (APP_ prefix)
• Tested — pytest suite covering config, frame processor geometry, and alert service logic

---

Architecture

main.py                         ← Streamlit entrypoint (landing page)
pages/
  1_Live_Analytics.py           ← Video upload / camera → real-time processing
  2_Configuration.py            ← Model params, alert rules, zone defaults
  3_Alert_Log.py                ← Searchable alert history

app/
  config.py                     ← Typed Settings (dataclass) + YAML loader
  analytics/
    frame_processor.py          ← Central orchestrator (detection → tracking → zones → alerts)
    utils.py                    ← Zone geometry helpers
  alerts/
    alert_service.py            ← Unified alert evaluator (intrusion / loitering / crowd)
    models.py                   ← Alert, AlertType, Severity data models

tests/                          ← pytest suite
  conftest.py                   ← Shared fixtures
  test_config.py                ← Config loading, defaults, env overrides
  test_frame_processor.py       ← Geometry helpers, track map, stats
  test_alert_service.py         ← Intrusion, cooldown, crowd, loitering, reset

Data Flow

Video Source → YOLO11 Detection → BotSORT Tracking → Zone Analytics → AlertService
                                                            ↓
                                                    Streamlit Dashboard
                                                  (metrics, charts, alerts)

---

Tech Stack

Component	Technology	Version
Detection	Ultralytics YOLO11	≥ 8.3
Tracking	BotSORT (built-in)	—
Video I/O	OpenCV (headless)	≥ 4.11
Dashboard	Streamlit (multi-page)	≥ 1.44
Charts	Plotly	≥ 5.24
Config	PyYAML + dataclasses	—
Testing	pytest	≥ 8.0
Linting	Ruff	≥ 0.8

---

Quick Start

1. Clone & Install

git clone https://github.com/DIPESHGOEL27/video-analytics.git
cd video-analytics
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt

For GPU support, install PyTorch with CUDA first:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

2. Download Models

Place at project root:

• yolo11n.pt — detection model
• yolo11n-seg.pt — segmentation model

3. Run

streamlit run main.py

Or use the Makefile:

make run      # Launch dashboard
make test     # Run test suite
make lint     # Run ruff linter

---

Configuration

All settings live in config.yaml. Example:

yolo:
  model_path: "yolo11n.pt"
  confidence: 0.5

alert:
  enabled: true
  cooldown_seconds: 10
  loiter_threshold_seconds: 5.0
  crowd_threshold: 5

Environment variable overrides — prefix with APP_, use __ for nesting:

APP_YOLO__CONFIDENCE=0.7 streamlit run main.py

---

Alert Rules

Rule	Trigger	Severity
Zone Intrusion	Object centroid enters the tracking zone	Medium
Loitering	Object remains in zone > threshold seconds	High
Crowd	Simultaneous objects in zone ≥ threshold	Critical

All rules respect a per-track cooldown to prevent alert spam.

---

Testing

pip install -r requirements-dev.txt
python -m pytest tests/ -v

Tests cover:

• Config system — defaults, YAML overrides, env-var overrides, unknown key handling
• Frame processor — _rect_to_polygon, _is_point_in_zone boundary cases, _build_track_map, stats serialisation
• Alert service — intrusion detection, cooldown enforcement, crowd thresholds, disable/reset

---

License

MIT