⚡ electrical engineering

Frugal Geofencing via Energy-aware Sensing and Reporting

This paper proposes an energy-aware geofencing framework for battery-free, camera-equipped IoT devices that utilizes reinforcement learning to optimize device placement and coordinated sensing, thereby achieving timely and reliable intruder detection with fewer devices and lower energy consumption compared to conventional grid-based deployments.

Original authors: David E. Ruiz-Guirola, Miltiadis Filippou, Onel A. Lopez

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: David E. Ruiz-Guirola, Miltiadis Filippou, Onel A. Lopez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guard a very valuable treasure chest (the Protected Area) in the middle of a large, dark warehouse. You have a team of security guards (the IoT Devices), but there's a catch: these guards don't have batteries. Instead, they are like solar-powered robots that only have enough energy to work for a few minutes at a time before they need to "recharge" by soaking up light (or radio waves) from the environment.

If you wake them up too often to check for thieves, they run out of energy and die. If you leave them asleep too long, a thief might sneak right past them before they wake up.

This paper presents a smart new way to manage these energy-starved guards so they can protect the treasure effectively without burning out. Here is the breakdown using simple analogies:

1. The Problem: The "Sleepy Guard" Dilemma

Traditional security systems assume guards have infinite battery life. They stand there 24/7, watching everything. But our new guards are "frugal." They have tiny energy buffers.

The Old Way: You line them up in a perfect grid (like a checkerboard). They all wake up at the same time, look around, and go back to sleep.
- The Flaw: If a thief walks between two guards while they are both asleep, you miss them. If you wake them up too often to check, they run out of energy by noon.
The New Way: We need a system that knows when to wake up specific guards and where to point their cameras, saving energy for the moments that actually matter.

2. The Solution: The "Smart Warden" (Reinforcement Learning)

The authors propose a "Smart Warden" (an AI controller) that sits in a central tower (the Access Point). This Warden doesn't just tell everyone to wake up; it acts like a conductor in an orchestra.

The Outer Ring (The Early Warning Zone): Imagine a fence around the treasure, but with a wide "moat" outside it. The goal isn't just to catch the thief inside the fence; it's to spot them in the moat before they even touch the fence. This gives you time to react.
Directional Vision: These guards have cameras, not 360-degree eyes. They can only look in one direction at a time (like a flashlight beam). Turning the camera costs a tiny bit of energy.
The AI Strategy: The Warden uses Reinforcement Learning (a type of AI that learns by trial and error). It watches the energy levels of every guard and predicts where a thief is likely to go.
- Scenario: If the AI sees a thief approaching from the North, it tells the North guards: "Wake up! Point your cameras North!" It tells the South guards: "Stay asleep, you're not needed right now."

3. How It Saves Energy (The "Frugal" Part)

Think of the energy buffer as a gas tank.

Grid Method (Old): Every car in the parking lot wakes up every 5 minutes to check if someone is there, regardless of where they are. They run out of gas quickly.
RL Method (New): The Warden only starts the engines of the cars that are facing the direction of the intruder. The others stay parked and conserve gas.
- Result: You need fewer guards to do the same job because the ones you do have are used much more efficiently.

4. The Results: Smarter, Faster, Cheaper

The paper tested this against the old "checkerboard" method and found:

Fewer Devices Needed: Because the AI is so good at picking the right guards, you can get away with fewer cameras to achieve the same level of security. This saves money on hardware.
Earlier Detection: Because the system focuses on the "moat" (the outer zone) and wakes up the right people at the right time, it spots intruders sooner. It's like spotting a wolf in the woods before it reaches the cabin door, rather than waiting for it to break the window.
Longer Life: The guards don't burn out. The AI balances the workload so that no single guard is overworked and runs out of energy while others sit idle.

Summary Analogy

Imagine you are trying to catch a fly in a room with a net.

The Old Way: You swing the net randomly every few seconds, hoping to hit the fly. You get tired quickly, and you often miss.
The New Way: You have a smart assistant who watches the fly. The assistant only tells you to swing the net when the fly is about to fly right into it. You swing less often, you get tired less, and you catch the fly much more reliably.

In a nutshell: This paper teaches us how to use a smart AI to manage a team of low-energy, battery-free cameras so they can act like a super-efficient security team, spotting intruders early without running out of juice.

1. Problem Statement

The paper addresses the challenge of deploying Ultra-Low Power Internet of Things (IoTDs) powered by Energy Harvesting (EH) for geofencing and intrusion detection.

The Core Conflict: Traditional grid-based IoT deployments assume continuous power or high-duty cycles. However, EH devices have limited energy buffers and intermittent power availability. Frequent wake-ups for sensing and reporting can deplete energy, leading to device unavailability.
The Specific Challenge: Ensuring timely and accurate detection of intruders crossing a protected boundary (geofence) using directional sensors (cameras) with limited Field-of-View (FoV), while strictly adhering to energy constraints.
Gap in Existing Literature: Most current approaches do not jointly optimize sensing schedules, reporting, and device orientation under the constraints of energy intermittency and directional sensing limitations.

2. System Model & Architecture

The proposed system consists of a Protected Area surrounded by an Outer Perimeter Zone (early detection buffer).

Hardware:
- IoTDs: Equipped with optical cameras (directional FoV), Energy Harvesting (EH) modules, rechargeable buffers, Wake-up Radios (WuR), and lightweight Computer Vision (CV) models for local inference.
- Access Point (AP) & Mobile Edge Host (MEH): The AP provides connectivity. The MEH hosts a Facility Geofencing System (FGS) containing a Controller and a Fusion Center (FC).
Operational Flow:
1. IoTDs operate in sleep/sense/report cycles.
2. Local CV models detect objects; valid detections are reported to the AP.
3. The FC fuses data, predicts intruder trajectories, and determines if additional sensing is needed.
4. The Controller dynamically instructs the AP to send event-driven wake-up requests to specific sleeping IoTDs via WuR, adjusting their activation status and camera orientation to cover predicted trajectories.

3. Methodology

A. Directional Sensing Power Model

The authors introduce a probabilistic model for detection confidence, $p(r_t, \Delta\theta_t)$ , which combines:

Angular Masking: A binary function $I(\Delta\theta_t)$ that is 1 only if the object is within the camera's FoV ( $[\theta_{min}, \theta_{max}]$ ).
Distance Decay: An exponential decay function $p(r_t) = \exp(-\eta r_t)$ representing the drop in detection confidence as distance increases, nullifying at a maximum range $r_{max}$ .
Detection Probability: Defined as the maximum confidence over the time an object is within the detectable area.

B. Energy Model

Harvesting: Modeled as a discrete Bernoulli process where devices harvest energy units with probability $\lambda$ .
Consumption: Energy is consumed for transmission ( $P_{Tx}$ ), wake-up ( $P_{WuR}$ ), and camera reorientation ( $P_{rot,bin}$ ).
Constraint: If the energy buffer is insufficient, the action fails, and the device remains unavailable until recharged.

C. Optimization Problem

The goal is to minimize the Object-Level Detection Error ( $E_{obj}$ ), which penalizes:

Missed Detections (Error = 1).
Late Detections (Error = $\mu_1$ , where $0 < \mu_1 < 1$ ).
Early Detections (Error = 0).

The objective is to jointly optimize:

Activation Policy ( $\delta(t)$ ): Which devices to wake up.
Orientation Policy ( $\Delta\theta(t)$ ): How to rotate cameras.

D. Reinforcement Learning (RL) Solution

Since the environment is dynamic (moving intruders, stochastic energy), the authors propose a Centralized RL approach:

Agent: The FGS Controller.
State: Includes device positions, energy buffer levels, current orientation, and sensing confidence.
Action: Selecting the activation vector and updating camera orientations.
Reward Function: Designed to balance three factors:
1. Energy Efficiency: Penalizing unnecessary activations.
2. Spatial Coverage: Rewarding the fraction of the perimeter covered.
3. Detection Performance: Penalizing late or missed detections.

4. Key Contributions

Frugal Geofencing Framework: A novel system for camera-equipped EH IoTDs that utilizes an outer perimeter zone for early detection before intruders breach the protected area.
Directional Sensing-Power Function: A mathematical model fusing distance decay and FoV masking to quantify time-dependent detection confidence.
Joint Optimization: Formulation of a problem that dynamically adjusts device activity, reporting, and sensing orientation to balance reliability, timeliness, and energy sustainability.
RL-Based Deployment: Demonstration that RL-based, mobility-aware control significantly reduces the required device density compared to static grid deployments while maintaining high reliability.

5. Results & Performance Analysis

The system was evaluated against a Grid Baseline (static placement, periodic sleep/wake) using Key Performance Indicators (KPIs): Reliability ( $P_{det} \geq 99.9\%$ ), Timeliness (Early Detection), and Energy Availability.

Reliability vs. Device Density:
- The RL-based approach achieves the 99.9% reliability target with significantly fewer devices ( $N_{min}$ ) than the grid baseline, especially under tight energy constraints (long duty cycles).
- The grid method requires a rapidly increasing number of devices as the duty cycle lengthens, whereas the RL method adapts efficiently.
Timeliness (Early Detection):
- While the grid method performs slightly better with very short duty cycles (frequent sensing), the RL method maintains superior timeliness as duty cycles increase.
- By prioritizing devices near the predicted intruder path, the RL method reduces "blind spots" and detection delays even with longer sleep intervals.
Energy Availability:
- Crucial Finding: The RL method maintains significantly higher average energy levels across the network.
- In the grid scenario, many devices deplete their energy (becoming unavailable) during peak activity hours. The RL method distributes the load, ensuring more devices remain operational, which directly correlates to sustained barrier coverage.

6. Significance

This paper demonstrates that intelligent, coordinated control is essential for sustainable, battery-free IoT security.

Cost Reduction: By reducing the number of required devices to achieve high reliability, the solution lowers deployment and maintenance costs.
Sustainability: It enables "frugal" operation where devices can run for decades on harvested energy without battery replacement.
Security Efficacy: It shifts the paradigm from "always-on" monitoring to "smart-on-demand" monitoring, ensuring that detection happens early (in the outer zone) rather than just after a breach, providing time for countermeasures.

In conclusion, the proposed framework successfully bridges the gap between the strict energy limitations of Ambient IoT and the high reliability requirements of critical geofencing applications.