🛡️ DRISHTI

Defense-Grade Early Warning Intelligence System

"Turning background noise into preventive defense insight."

Designed for border surveillance and high-security perimeters where threats emerge gradually.

---

🎯 The Problem

Traditional surveillance systems are REACTIVE:

• They alert you AFTER something bad happens
• Binary alerts: "NORMAL" or "ALARM"
• High false positive rate causes alert fatigue
• Misses gradual threats that evolve slowly

Real-world threats often emerge GRADUALLY:

• Border infiltrations happen in stages
• Insider threats develop over weeks
• Equipment fails progressively
• Crowd tension builds before incidents

---

💡 Our Solution

DRISHTI detects threats BEFORE they become obvious.

Instead of asking "Is this instant abnormal?" (reactive), we ask:

"Is behavior gradually changing from what's normal?" (proactive)

Key Innovation: Drift Detection, Not Anomaly Detection

---

🏗️ System Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                         DRISHTI PIPELINE                        │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────────────┐  │
│  │ 📹 VIDEO     │───→│ 🔬 FEATURE   │───→│ 🧠 LSTM-VAE          │  │
│  │  INGESTION   │    │  EXTRACTION  │    │  TEMPORAL LEARNING   │  │
│  └──────────────┘    └──────────────┘    └──────────────────────┘  │
│         │                    │                       │              │
│         │           24 Behavioral Features           │              │
│         │           • Motion Energy                  │              │
│         │           • Optical Flow                   │              │
│         │           • Scene Entropy                  │              │
│         │           • Direction Patterns             │              │
│         │                                            │              │
│         │                    ↓                       ↓              │
│         │         ┌────────────────────────────────────┐            │
│         │         │     📊 DRIFT INTELLIGENCE          │            │
│         │         │  • Threat Deviation Index (TDI)    │            │
│         │         │  • KL Divergence Analysis          │            │
│         │         │  • EWMA Smoothed Scoring           │            │
│         │         │  • Trend Detection (↑ → ↓)         │            │
│         │         └────────────────────────────────────┘            │
│         │                         │                                 │
│         │                         ↓                                 │
│         │         ┌────────────────────────────────────┐            │
│         │         │     🎯 RISK ZONE CLASSIFIER        │            │
│         │         │  🟢 NORMAL  → Standard monitoring  │            │
│         │         │  🟡 WATCH   → Increase attention   │            │
│         │         │  🟠 WARNING → Alert response team  │            │
│         │         │  🔴 CRITICAL→ Immediate action     │            │
│         │         └────────────────────────────────────┘            │
│         │                         │                                 │
│         │                         ↓                                 │
│         │         ┌────────────────────────────────────┐            │
│         │         │     🖥️ INTELLIGENCE DASHBOARD      │            │
│         │         │  • Real-time TDI visualization     │            │
│         │         │  • Feature attribution (XAI)       │            │
│         │         │  • Human-in-the-loop feedback      │            │
│         │         └────────────────────────────────────┘            │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

---

✨ Key Features

Feature	Description
🎓 Unsupervised Learning	Train only on normal behavior - no labeled anomalies needed
🧠 LSTM-VAE Architecture	Pure NumPy implementation - temporal normality modeling
� Ensemble Detection	LSTM-VAE + Isolation Forest + One-Class SVM + LOF
📊 Threat Deviation Index	0-100 scale for intuitive operator understanding
🎯 4-Tier Risk Zones	Graduated alerts reduce fatigue (Normal → Watch → Warning → Critical)
🏷️ Anomaly Classification	Categorize threats (Loitering, Intrusion, Crowd, etc.)
🌌 3D Latent Visualization	Visualize behavioral trajectories in latent space
🔍 Explainable AI (XAI)	Shows WHICH features are causing drift
🚨 Incident Logging	Track, export, and analyze all alerts
🔮 TDI Forecasting	Predict future threat levels
👤 Human-in-the-Loop	Operators provide feedback for baseline adaptation
📹 Real Video Support	Works with UCSD dataset and custom surveillance footage

---

🚀 Quick Start

Prerequisites

• Python 3.10 or higher
• Git

Installation

# 1. Clone the repository
git clone https://github.com/YourUsername/Noise_floor_project.git
cd Noise_floor_project

# 2. Create virtual environment
python -m venv .venv

# 3. Activate virtual environment
# Windows PowerShell:
.\.venv\Scripts\Activate.ps1
# Windows CMD:
.\.venv\Scripts\activate.bat
# Linux/Mac:
source .venv/bin/activate

# 4. Install dependencies
pip install -r requirements.txt

Running the Application

# Launch the main dashboard (RECOMMENDED)
streamlit run dashboard/app_main.py

The dashboard opens at http://localhost:8501

---

🖥️ Dashboard Guide

Dashboard Tabs

Tab	Description
📊 Intelligence Dashboard	Main TDI display, risk zones, feature attribution, AI explanations
🧠 AI Ensemble	Multi-model detection votes, 3D latent space, anomaly classification
📹 Camera Grid	6-camera surveillance view with live UCSD frames
🚨 Incident Log	Full history of all alerts with export capability
📈 Analytics	Session statistics, zone distribution, data export

Operating Modes

Mode	Description	Use Case
📹 UCSD Real Video	Processes actual surveillance footage from UCSD dataset	Production demo, validation
🔬 Synthetic Demo	Uses generated data with controlled drift	Quick testing, concept demo

Understanding the Display

Threat Deviation Index (TDI)

• 0-25: 🟢 NORMAL - All good
• 25-50: 🟡 WATCH - Something's slightly off
• 50-75: 🟠 WARNING - Confirmed drift, pay attention
• 75-100: 🔴 CRITICAL - Take action immediately

Drift Trend

• ↑ RISING - Threat is increasing
• → STABLE - No significant change
• ↓ FALLING - Returning to normal

Anomaly Categories

• 🧍 Loitering - Prolonged stationary activity
• ⚠️ Intrusion - Boundary crossing detected
• 👥 Crowd Formation - Unusual gathering
• 🌀 Erratic Movement - Abnormal motion patterns
• 🎯 Coordinated Activity - Synchronized movement
• ⚡ Speed Anomaly - Unusual velocity
• ↩️ Direction Anomaly - Unusual direction pattern

---

📁 Project Structure

Noise_floor_project/
├── 📂 dashboard/
│   └── app_main.py            # Main dashboard (USE THIS)
│
├── 📂 src/
│   ├── lstm_vae.py            # LSTM-VAE temporal model
│   ├── drift_intelligence.py  # TDI computation engine
│   ├── ensemble_detector.py   # Multi-model ensemble
│   ├── advanced_ai.py         # Anomaly classification
│   ├── risk_zones.py          # 4-tier zone classifier
│   ├── behavioral_features.py # 24 feature definitions
│   ├── video_features.py      # Real video processing
│   ├── explainability.py      # XAI attribution
│   ├── incident_logger.py     # Incident tracking
│   ├── feedback_system.py     # Human-in-the-loop
│   └── utils.py               # Utility functions
│
├── 📂 data/
│   └── UCSD_Anomaly_Dataset.v1p2/
│       ├── UCSDped1/
│       │   ├── Train/         # Normal pedestrian videos
│       │   └── Test/          # Contains anomalies
│       └── UCSDped2/
│
├── 📂 incident_logs/          # Logged incidents
├── 📂 feedback_data/          # Operator feedback logs
├── 📂 baseline_data/          # Frozen baseline snapshots
│
├── config.py                  # Configuration settings
├── main.py                    # CLI entry point
├── requirements.txt           # Python dependencies
└── README.md                  # This file

---

🔬 How It Works

Phase 1: Learning Normal Behavior

UCSD Train Data (Normal pedestrians)
    → Optical Flow Extraction
    → 24 Behavioral Features
    → LSTM-VAE Training
    → Ensemble Detector Fitting
    → Baseline Established ✓

Phase 2: Monitoring & Detection

UCSD Test Data (Contains bikes, carts, etc.)
    → Feature Extraction
    → LSTM-VAE Inference
    → Ensemble Voting (IF, SVM, LOF)
    → Compute TDI
    → Classify Risk Zone
    → Classify Anomaly Type
    → Log Incident
    → Generate Explanation

Phase 3: Operator Response

Dashboard displays:
    → Current TDI (e.g., 67)
    → Risk Zone (🟠 WARNING)
    → Trend (↑ RISING)
    → Top Features causing drift
    → AI explanation
    
Operator can:
    → Acknowledge alert
    → Mark as false positive
    → Request investigation
    → Update baseline (human-gated)

---

🎓 System Philosophy

"Defense systems manage CONFIDENCE, not panic."

"AI assists operators, it does NOT replace them."

"Baseline adaptation is human-gated."

These three principles guide every design decision:

1. Graduated Risk Zones - Reduce alert fatigue with progressive warnings
2. Explainable AI - Operators understand WHY alerts occur
3. Human-in-the-Loop - Critical decisions remain with humans
4. Baseline Protection - Prevents adversarial manipulation

---

📊 Dataset Information

UCSD Anomaly Detection Dataset

Used as proxy for border surveillance footage.

Subset	Train	Test	Anomalies
Ped1	34 clips	36 clips	Bikes, skateboards, carts
Ped2	16 clips	12 clips	Bikes, skateboards

How we use it:

• Train folder → Learn NORMAL pedestrian behavior
• Test folder → Detect DRIFT when anomalies appear (labels ignored - unsupervised)

---

🛠️ Technical Specifications

Component	Specification
ML Model	LSTM-VAE (Pure NumPy, no TensorFlow/PyTorch)
Features	24 behavioral metrics from optical flow
Latent Dim	8-dimensional latent space
Sequence Length	10 frames temporal window
Smoothing	EWMA with α=0.15
Technology Readiness	TRL-4 (Lab Validated)

---

🔧 Configuration

Edit config.py to customize:

# Data source
DATA_MODE = "real_video"  # "synthetic" | "real_video"

# UCSD Dataset
UCSD_SUBSET = "ped1"      # "ped1" or "ped2"

# Baseline protection
BASELINE_FREEZE_CONFIG = {
    'learning_window': 200,
    'freeze_after_learning': True,
    'adaptation_learning_rate': 0.01,
}

---

🤝 Use Cases

Domain	Application
Border Security	Detect infiltration patterns at perimeter fences
Airport Security	Monitor crowd behavior at checkpoints
Critical Infrastructure	Surveillance of power plants, data centers
Military Installations	Base perimeter monitoring
Corporate Security	Campus and facility protection

---

📈 Performance Metrics

Metric	Description
Detection Delay	Frames between actual drift start and system detection
False Positive Rate	Alerts during confirmed normal periods
Peak TDI	Maximum threat deviation observed
Zone Transitions	History of risk zone changes

---

🐛 Troubleshooting

Common Issues

1. "Dataset not found"

# Ensure UCSD dataset is in correct location:
data/UCSD_Anomaly_Dataset.v1p2/UCSDped1/Train/

2. "Module not found"

# Make sure virtual environment is activated
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

3. "Streamlit not starting"

pip install streamlit --upgrade
streamlit run dashboard/app_pro_v2.py

---

👥 Team

• Project: DRISHTI - Defense Intelligence System

---

🛡️ DRISHTI - Because early warning saves lives. 🛡️