christiankhoury05/Chickadee28M

CHICKADEE 28M - Multi-Drone Detection & Tracking System

🌐 Developed by Deep Autonomy - Advanced AI solutions for autonomous systems

Latest Update: Enhanced with 200K iteration training achieving 81.30% mAP performance

Overview

CHICKADEE 28M is a production-ready computer vision system for real-time multi-drone detection and tracking. Built on Detectron2 with a ResNet-18 backbone, the system provides robust object detection with advanced ByteTrack multi-object tracking capabilities and enterprise-grade reliability.

Model Specifications

Architecture

• Framework: Detectron2 (PyTorch)
• Base Model: Faster R-CNN
• Backbone: ResNet-18 with Feature Pyramid Network (FPN)
• Input Resolution: 1920x1080 (Full HD)
• Model Size: 216 MB (.pth format)
• Parameters: 28,269,781 total (28.1M trainable)

Performance Metrics (COCO Evaluation)

• Average Precision (AP): 81.30%
• AP50: 97.00%
• AP75: 92.80%
• AP Small: 77.93%
• AP Medium: 90.29%
• AP Large: 40.88%

Training Configuration

• Training Iterations: 200,000 (Extended training)
• Dataset: Synthetic drone dataset (143K images)
• Batch Size: 8
• Learning Rate: 0.0004 with decay steps at 140K, 180K
• Optimizer: SGD with momentum
• Data Augmentation: Random flip, resize shortest edge
• Warmup: 8,000 iterations

Detection Performance

• Confidence Threshold: 0.02 (configurable)
• NMS Threshold: 0.05 (aggressive suppression)
• RPN NMS Threshold: 0.1
• Maximum Detections: 1000 per image

Tracking Performance

• Tracking Algorithm: ByteTrack (simplified implementation)
• Track Threshold: 0.3
• Track Buffer: 30 frames
• Match Threshold: 0.7 (IoU-based association)
• Minimum Box Area: 10 pixels

Production Features

• Real Confidence Scores: Native PyTorch precision
• Enterprise Reliability: Battle-tested Detectron2 backend
• Professional Visualization: Detectron2 + custom track overlays
• Robust Tracking: Advanced ByteTrack implementation
• CPU/GPU Support: Automatic device detection
• Configurable Parameters: Flexible deployment options

System Requirements

Hardware

• CPU: Multi-core processor (Intel i5 or equivalent)
• RAM: 8GB minimum, 16GB recommended
• GPU: Optional (CUDA-compatible for acceleration)
• Storage: 500MB free space

Software Dependencies

• Python 3.8+
• PyTorch 1.9+
• Detectron2
• OpenCV 4.5+
• NumPy
• SciPy

File Structure

chickadee-28m/
├── README.md                           # This documentation
├── model_final.pth                     # Trained ResNet-18 model (216 MB)
├── config.json                         # Model configuration (enables download tracking)
├── model_index.json                    # Model metadata and file index
└── requirements.txt                    # Python dependencies

Usage

Model Loading

import torch
from detectron2.config import get_cfg
from detectron2.modeling import build_model
from detectron2.checkpoint import DetectionCheckpointer
from detectron2 import model_zoo

# Load the model
cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.BACKBONE.NAME = "build_resnet_fpn_backbone"
cfg.MODEL.RESNETS.DEPTH = 18
cfg.MODEL.RESNETS.RES2_OUT_CHANNELS = 64
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1
cfg.MODEL.WEIGHTS = "./model_final.pth"

model = build_model(cfg)
checkpointer = DetectionCheckpointer(model)
checkpointer.load(cfg.MODEL.WEIGHTS)
model.eval()

Basic Inference

import cv2
from detectron2.utils.visualizer import Visualizer
from detectron2.data import MetadataCatalog

# Load and process image/video frame
image = cv2.imread("input_image.jpg")
with torch.no_grad():
    predictions = model([{"image": torch.as_tensor(image.astype("float32").transpose(2, 0, 1))}])

# Visualize results
v = Visualizer(image[:, :, ::-1], MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale=1.0)
out = v.draw_instance_predictions(predictions[0]["instances"].to("cpu"))
result = out.get_image()[:, :, ::-1]

Configuration Options

Detection Parameters

• confidence_threshold: Minimum detection confidence (default: 0.02)
• nms_threshold: Non-maximum suppression threshold (default: 0.05)
• max_detections: Maximum detections per frame (default: 1000)

Tracking Parameters

• track_thresh: High-confidence track threshold (default: 0.3)
• track_buffer: Frame buffer for lost tracks (default: 30)
• match_thresh: IoU threshold for track association (default: 0.7)
• min_box_area: Minimum bounding box area (default: 10)

Output Format

Detection Results

Each detection includes:

• Bounding box coordinates (x1, y1, x2, y2)
• Confidence score (0.0 - 1.0)
• Class prediction (drone)

Tracking Results

Each track provides:

• Unique track ID
• Bounding box coordinates
• Confidence score
• Track state (active/lost/removed)
• Velocity estimation

Performance Benchmarks

Processing Statistics (Example Video)

• Input: 1920x1080, 417 frames, 23 FPS
• Total Detections: 11,188 (0.01 confidence)
• Average Detections/Frame: 28.25
• Detection Rate: 100% (detections in every frame)
• Average Active Tracks/Frame: 39.12
• Maximum Simultaneous Tracks: 99
• Processing Speed: 0.4 FPS (CPU) / 12.3 FPS (GPU RTX 3080)

Accuracy Metrics

• Classification Accuracy: 97.56%
• False Negative Rate: 6.30%
• Foreground Classification Accuracy: 93.70%

Technical Features

Advanced Detection

• Multi-scale feature extraction via FPN
• Aggressive non-maximum suppression for dense scenes
• Configurable confidence thresholds
• Robust bounding box filtering (size, aspect ratio)

Multi-Object Tracking

• Hungarian algorithm for optimal track-detection association
• Kalman filter-based motion prediction
• Track state management (new, tracked, lost, removed)
• Re-identification of temporarily lost tracks
• Velocity-based motion compensation

Visualization

• Professional bounding box rendering
• Consistent track ID coloring
• Real-time confidence score display
• Detectron2-compatible visualization pipeline

Model Validation

Training Convergence (200K Iterations)

• Improved Performance: +5.51% AP improvement over 50K training
• Extended Training: 4x longer training for better convergence
• Learning Rate Schedule: Decay at 140K and 180K iterations
• Stable Convergence: Excellent performance on validation set

Evaluation Protocol

• COCO-style evaluation metrics
• IoU thresholds: 0.5, 0.75, 0.5:0.95
• Size categories: Small (<32²), Medium (32²-96²), Large (>96²)

Deployment Considerations

Production Deployment

• PyTorch backend ensures consistent confidence scores
• CPU-compatible for broad hardware support
• Configurable parameters for different scenarios
• Model can be integrated into custom applications

Integration Notes

• Model expects RGB images in Detectron2 format
• Outputs standard COCO-format predictions
• Compatible with Detectron2 visualization tools

Performance Optimization

• GPU acceleration available with CUDA
• Batch processing for multiple images/videos
• Configurable detection thresholds for speed/accuracy trade-offs
• Memory-efficient inference

Limitations

Current Constraints

• Optimized for drone-like objects (small, aerial vehicles)
• Performance degrades with very large objects (AP Large: 23.47%)
• CPU-only processing limits real-time performance
• Requires Detectron2 framework for inference

Recommended Use Cases

• Surveillance and monitoring applications
• Multi-drone analysis
• Aerial vehicle counting and tracking
• Security and defense applications

Support and Maintenance

Model Updates

• Retrain with domain-specific data for improved accuracy
• Adjust NMS thresholds for different object densities
• Fine-tune tracking parameters for specific scenarios

Troubleshooting

• Ensure all dependencies are correctly installed
• Verify model file integrity (215.4 MB expected size)
• Check input video format compatibility
• Monitor memory usage for large videos

Version History

Version 1.0 (Updated)

• ResNet-18 backbone for improved efficiency
• 200K iteration training (4x extended)
• 81.30% mAP performance (+5.51% improvement)
• Enhanced ByteTrack implementation with sticky tracking
• Real confidence score integration
• GPU acceleration support (12.3 FPS on RTX 3080)
• Production-ready visualization with metrics display
• 28M parameter architecture
• Comprehensive documentation

---

About Deep Autonomy

CHICKADEE 28M is developed by Deep Autonomy, a leading provider of advanced AI solutions for autonomous systems. We specialize in computer vision, multi-object tracking, and real-time detection systems for defense, surveillance, and autonomous vehicle applications.

🌐 Visit us: deepautonomy.ai
📧 Contact: For enterprise solutions and custom AI development
🚀 Expertise: Computer Vision • Object Detection • Multi-Object Tracking • Real-time AI Systems

Download Metrics

This model includes download tracking via Hugging Face's metrics system:

• Primary Query File: config.json (enables download counting)
• Model Query File: model_final.pth (tracks model downloads)
• Tracking Method: Server-side HTTP request counting (no user data collected)

---

CHICKADEE 28M - Professional Multi-Drone Detection & Tracking System
Developed for production deployment with enterprise-grade reliability and performance.