Military Object Detection & Tracking (MODT)

Production-ready end-to-end object detection and tracking system for military applications.

Features

• Real-time detection using YOLOv8 (50-60 FPS on GPU)
• Multi-tracker support: Kalman, ByteTrack, DeepSORT, IoU
• Edge optimization: TensorRT, ONNX, INT8/FP16 quantization
• Trained model: 12 military object classes, 26K+ images

System Requirements

• Python: 3.10+
• GPU: NVIDIA GPU with CUDA support (RTX 3060+ recommended)
• VRAM: 6GB+ for training, 4GB+ for inference
• Disk: 50GB+ for datasets

Supported Objects (12 Classes)

ID	Class Name	ID	Class Name
0	Camouflage Soldier	6	Soldier
1	Weapon	7	Civilian Vehicle
2	Military Tank	8	Military Artillery
3	Military Truck	9	Trench
4	Military Vehicle	10	Military Aircraft
5	Civilian	11	Military Warship

Project Structure

MODT/
├── config/                        # Configuration files
│   ├── default.yaml               # Default settings
│   ├── training.yaml              # Training parameters
│   └── edge_jetson.yaml           # Jetson deployment config
│
├── data/                          # Datasets (Ignored in Git)
│   └── military/assets/           # Military dataset (26K images)
│
├── docs/                          # Documentation
│   ├── ARCHITECTURE.md            # System architecture
│   ├── TRAINING.md                # Training guide
│   └── DEPLOYMENT.md              # Deployment guide
│
├── models/                        # Model implementations
│   ├── __init__.py                # Package init
│   ├── dtea.py                    # Dynamic Temporal Enhanced Attention
│   ├── backbones/                 # Backbone networks
│   ├── detectors/                 # Detection models (YOLODetector)
│   └── trackers/                  # Tracking algorithms (5 trackers)
│
├── outputs/                       # Outputs & Demos
│   └── trained_model_demo/        # Local Demo results
│
├── scripts/                       # Executable scripts (14 total)
│   ├── train_detector.py          # Model training
│   ├── demo.py                    # Synthetic demo
│   ├── demo_trained_model.py      # Run trained model on images
│   ├── demo_video.py              # Video demo
│   ├── deploy_edge.py             # Edge deployment
│   ├── evaluate_model.py          # Model evaluation
│   ├── export_onnx.py             # ONNX export
│   ├── optimize_tensorrt.py       # TensorRT optimization
│   ├── compare_trackers.py        # Tracker comparison
│   ├── download_datasets.py       # Dataset management
│   ├── run_military_inference.py  # Military inference
│   ├── convert_airbus_to_yolo.py  # Airbus ship converter
│   ├── convert_flir_to_yolo.py    # FLIR thermal converter
│   └── merge_datasets.py          # Multi-dataset merger
│
├── src/                           # Core source code
│   ├── __init__.py                # Package init
│   ├── analytics/                 # Analytics (zones, trajectories)
│   ├── data/                      # Data loading, augmentation
│   ├── deployment/                # Deployment pipeline
│   ├── detection/                 # Detection (trainer, evaluator)
│   ├── optimization/              # Model optimization
│   ├── tracking/                  # Tracking utilities
│   └── utils/                     # Config, logger utilities
│
├── tests/                         # Unit tests (3 files)
│   ├── test_detection.py          # Detection tests
│   ├── test_tracking.py           # Tracking tests
│   └── test_integration.py        # Integration tests
│
├── weights/                       # Pre-trained weights
│   └── best.pt                    # Best trained YOLOv8s model
│
├── .env                           # Environment variables
├── .gitignore                     # Git ignore rules
├── requirements.txt               # Python dependencies
├── setup.py                       # Package setup
└── TECHNICAL_BRIEF.md             # Technical summary

Installation

# Create virtual environment
python -m venv .venv
.venv\Scripts\activate  # Windows

# Install dependencies
pip install -r requirements.txt

Quick Start

# Run demo with synthetic data
python scripts/demo.py

# Run with trained model on images
python scripts/demo_trained_model.py

# Real-time webcam detection
python -m src.deployment.pipeline --source 0

# Headless mode (server)
python -m src.deployment.pipeline --source 0 --no-display

Training

# Train with YOLOv8s (recommended for best results)
.venv\Scripts\yolo.exe detect train \
    data=data/military/assets/military_object_dataset/military_dataset.yaml \
    model=yolov8s.pt \
    epochs=150 \
    batch=16 \
    imgsz=640 \
    device=0

# Or use training script
python scripts/train_detector.py --epochs 100

Model Performance

Model	Epochs	mAP@0.5	mAP@0.5:0.95	Precision	Recall
YOLOv8n	Baseline	0.445	0.280	-	-
YOLOv8s	75	0.613	0.418	0.616	0.556

Training Growth:

• mAP@0.5 improved by +38% over baseline
• mAP@0.5:0.95 improved by +49% over baseline

Best performing classes: Military Aircraft, Military Tank, Soldier.

Trained Model

The best trained model is located at:

outputs/yolov8s_train/military_75ep/weights/best.pt

Use it for inference:

.venv\Scripts\yolo.exe detect predict model=outputs/yolov8s_train/military_75ep/weights/best.pt source=your_image.jpg

Dataset

• Main Dataset: Military Object Dataset (26,315 images, 12 classes)
• Format: YOLOv8 compatible
• Split: Train (21,978) / Val (2,941) / Test (1,396)

Documentation

• Architecture Guide
• Training Guide
• Deployment Guide

License

Internal Project