DeerAI Tracking Response

A wildlife management platform that transforms property protection through spatial intelligence. Built to detect, track, and deter deer and other animals using computer vision, Kalman filtering, and autonomous ground vehicles.

The Challenge

Property owners face constant wildlife intrusion damaging gardens, crops, and landscaping. Traditional deterrents are static and ineffective. Animals quickly learn to ignore them.

The solution? Dynamic, intelligent response.

System Architecture

Spatial Intelligence Pipeline

1. Detection → MegaDetector v5 & custom YOLOv8n models identify wildlife in each frame
2. Tracking → Kalman filter + appearance embeddings maintain identity across occlusions
3. Localization → AprilTag homography converts pixel coordinates to real-world positions (feet)
4. Behavior Analysis → Classify posture, speed, direction, and intent
5. Trajectory Forecasting → Predict paths and route deterrent systems

Dual-Environment Development

Indoor Tabletop (Rapid Iteration)

• 24"×30" test environment with Wyze v2 + Reolink E1 cameras
• CuteBot Pro unmanned ground vehicle (UGV)
• AprilTag calibration for precision coordinate mapping
• Iterate detection, tracking, and routing algorithms safely

Outdoor Production

• 4-5 Reolink cameras covering property perimeter
• Real-time RTSP stream processing
• GPU-accelerated inference on Ubuntu server (GTX 3080)
• Coordinate mapping across entire property

Technical Stack

• Computer Vision: OpenCV, PyTorch, Ultralytics YOLO, MegaDetector v5
• Tracking: Kalman filtering, Hungarian algorithm, MegaDescriptor embeddings
• Cameras: Wyze v2, Reolink E1 Zoom, Reolink Duo 3 PoE
• Robotics: CuteBot Pro (Micro:bit v2), custom BLE control
• Infrastructure: AWS S3, Samba shared storage, tmux monitoring dashboards
• Calibration: AprilTags (tag36h11 family)

Key Features

Real-Time Detection & Tracking

MegaDetector provides robust animal detection. Custom YOLOv8n models fine-tuned for local species and lighting conditions. Appearance embeddings ensure consistent ID across frames even with occlusions.

Precision Localization

AprilTag-based homography transforms camera pixels into yard coordinates (x_ft, y_ft). Know exactly where animals are, not just what they look like.

Autonomous Response

Route UGVs to intercept predicted trajectories. Activate targeted deterrents (lights, sound, motion) exactly when and where needed.

Continuous Retraining Loop

Drop new footage into the pipeline. Auto-generate draft labels. Correct and retrain. Quality gates enforce AP50, recall, and latency thresholds before deployment.

Workflow

# Data ingestion
make frames       # Extract frames from raw clips
make autolabel    # Generate YOLO draft labels
make qc           # Validate dataset integrity
make split        # Train/val/test splits

# Training & evaluation
make train        # YOLOv8n with locked hyperparameters
make eval         # Metrics + pytest quality gates

# Deployment
make export       # ONNX/TorchScript for edge inference

Inspiration

Based on the University of Minnesota's "Geofenced Unmanned Aerial Robotic Defender for Deer Detection and Deterrence (GUARD)" research. Adapted for ground-based systems with enhanced tracking and behavior analysis.

Current Status

Active development. Indoor tabletop system operational with functional detection, tracking, and UGV routing. Outdoor deployment in progress with multi-camera calibration and deterrent integration.

What's Next

• Multi-target trajectory prediction
• Behavior state classification (grazing, alert, fleeing)
• Coordinated multi-UGV swarm control
• Event highlight reels with automated video generation
• Integration with smart home systems (HomeKit, Home Assistant)

This is not just object detection. This is spatial intelligence.