InvisibleFence: Non-Lethal Edge-Optimized AI for Human Wildlife Coexistence

Overview

Human-wildlife conflicts in residential and agricultural settings often rely on ineffective or ecologically harmful deterrents (e.g., rodenticides, fences, lethal traps), forcing gardeners and small-scale farmers to abandon crops, causing ecosystem damage, and plagued by high false-positive rates and habituation. InvisibleFence addresses these challenges via an edge-optimized framework for accurate wildlife detection and targeted, non-lethal deterrence.

InvisibleFence modular hardware: Vision Pod (left) with 2K IR camera and 240° motion sensing, Sound Pod (right) with 6-speaker ultrasonic array

Edge Intelligence

2K video processed in <0.8s on embedded hardware

Species-Specific Response

Modular deterrents activate only for target wildlife, sparing pets

Power Efficiency

Continuous 24/7 operation under 5W, fully cloud-free

Motivation

Current Ineffective Solutions

Traditional fences and generic ultrasonic deterrents fail to prevent wildlife intrusion effectively

The Problem

Gardeners abandon crops due to wildlife damage
Traditional deterrents harm ecosystems
Motion sensors have high false-positive rates
Animals habituate to generic deterrents

Our Solution

AI-powered species identification
Targeted, non-lethal deterrence
Edge computing for privacy & efficiency
Adaptive response to prevent habituation

System Design

A modular framework combining edge AI detection with adaptive deterrence

Vision Pod

→ 3D-printed enclosure with weatherproof design
→ 2560×1440 IR camera with 174° fisheye lens
→ 240° tri-PIR array for motion detection
→ Raspberry Pi 5 running YOLOv8-Nano NCNN
→ Sub-5W power consumption

Deterrent Modules

→ Sound Pod: 6-speaker ultrasonic array
→ Light Module: Programmable LED strobes
→ Spray System: Water/scent deterrents
→ MQTT Control: Real-time activation
→ Human Safety: Auto-suppression when humans detected

Dataset

44,000 real-world images spanning 11 wildlife species classes

Our comprehensive dataset captures diverse wildlife in various conditions, collected from actual farms and backyards for real-world relevance.

44,000 Real-World Images

11 Wildlife Species

24/7 Day/Night Coverage

Species Coverage: Deer, fox, raccoon, rabbit, birds, squirrel, and other common garden intruders
Condition Variety: Day/night captures with IR and standard illumination across all seasons
Field-Collected Data: Images sourced from actual deployment sites

Request Dataset Access

Results

State-of-the-art performance with minimal computational footprint

Model	Precision	Recall	F1	mAP	Size (MB)	Time (s)
YOLOv7-416	0.855	0.802	0.82	0.84	12.3	692.38
YOLOv7-640	0.856	0.811	0.83	0.855	12.3	699.52
YOLOv8-N NCNN (Ours)	0.868	0.811	0.83	0.867	11.7	157.09
YOLOv11-S	0.889	0.804	0.87	0.893	19.2	1884.12

Species Classification Performance

Confusion Matrix showing classification accuracy

Confusion matrix demonstrating high accuracy across 11 wildlife species with minimal false positives

Key Performance Metrics:

→ 86.7% mAP with only 11.7 MB model size
→ 5x faster inference than YOLOv11
→ Sub-second processing on Raspberry Pi 5
→ Species-specific accuracy: Deer (0.87), Raccoon (0.95), Rabbit (0.90)

Live Demo

See InvisibleFence in action

The Vision Pod distinguishes animals from humans in its 240° view. When an animal appears, it triggers deterrent devices via MQTT. Detection stops if a human enters the frame, ensuring safety. A live dashboard displays detection confidence and distance to target.

Publications

Demo Paper

Snehalraj Chugh, Milind Rampure, Elijah Polyakov, Bipendra Basnyat, and Nirmalya Roy. Demo: InvisibleFence: Non-Lethal Edge-Optimized AI for Human Wildlife Coexistence and Crop Protection. In ACM MobiSys '25, June 23-27, 2025, Anaheim, CA, USA.

Demo Paper PDF

Poster PDF

Citation

@inproceedings{invisiblefence2025,
  author = {Snehalraj Chugh and Milind Rampure and Elijah Polyakov and Bipendra Basnyat and Nirmalya Roy},
  title = {Demo: InvisibleFence: Non-Lethal Edge-Optimized AI for Human Wildlife Coexistence and Crop Protection},
  booktitle = {The 23rd Annual International Conference on Mobile Systems, Applications and Services (MobiSys '25)},
  year = {2025},
  month = {June},
  days = {23-27},
  location = {Anaheim, CA, USA},
  publisher = {ACM},
  doi = {10.1145/3711875.3734379}
}

Team

Supported by NSF CAREER Award #1750936, NSF CNS EAGER Grant #2233879, NSF REU Site Grant #2050999, NSF I-Corps Grant #2502886, UMBC FEAT and TCF Grants.

InvisibleFence

Non-Lethal Edge-Optimized AI for Human Wildlife Coexistence and Crop Protection

ACM MobiSys 2025 Demo