💻 computer science

Dot-Flik: A Scalable Edge AI Architecture for Distributed Insect Monitoring

This paper presents Dot-Flik, a scalable edge AI architecture for distributed insect monitoring that utilizes motion-informed frame filtering to reduce data transmission and energy consumption, enabling cost-effective, real-time biodiversity tracking on low-cost hardware without relying on centralized processing.

Original authors: Mattia Consani, Denisa-Andreea Constantinescu, Åse Håtveit, Titus Venverloo, Fabio Duarte, Carlo Ratti, David Atienza

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mattia Consani, Denisa-Andreea Constantinescu, Åse Håtveit, Titus Venverloo, Fabio Duarte, Carlo Ratti, David Atienza

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 count every single bee, butterfly, and beetle in a city park to understand how healthy the local ecosystem is. The problem? There are too many insects, and they move too fast for humans to count manually. If you try to use cameras, you run into a new problem: cameras generate a massive amount of video data. Most of that video is just empty sky or leaves blowing in the wind. Sending all that useless footage to a central computer is expensive, drains batteries quickly, and clogs up the network.

This paper introduces Dot-Flik, a smart, two-part system designed to solve this "data flood" problem so we can monitor insects on a large scale without breaking the bank.

Here is how it works, using simple analogies:

The Problem: The "Always-On" Camera Trap

Think of traditional insect monitoring like having a security guard at every single door of a massive hotel. Each guard (a camera) is constantly shouting, "I see something! I see something!" even if it's just a leaf falling or a shadow moving.

The Cost: You need a huge team of guards (expensive computers) to listen to all these shouts.
The Waste: 90% of the shouting is about nothing important. The guards get tired (batteries die), and the phone lines get jammed (network overload).

The Solution: The Dot-Flik Team

The authors propose splitting the job into two specialized roles, creating a "hierarchical" team.

1. The "Dot" Nodes (The Smart Bouncers)

These are the low-cost cameras placed all over the park. Instead of being "dumb" cameras that record everything, they act like smart bouncers at a club entrance.

How they work: They watch the video feed but only look for movement.
The Trick: They use a clever math trick (called "motion-informed filtering") to ignore the wind blowing through the trees. They know that wind moves everything slowly and evenly, while an insect moves quickly in a specific spot.
The Result: If the bouncer sees a leaf blowing, they say, "Ignore that." If they see a bee buzzing by, they say, "Stop! Send this to the boss!"
Efficiency: In calm or light-wind conditions, these bouncers throw away 60% to 80% of the video frames before they even leave the camera. They only send the "interesting" moments.

2. The "Flik" Node (The Expert Analyst)

This is the central computer (the "boss") that receives the filtered messages from the many "Dot" bouncers.

How it works: Because the "Dots" have already done the hard work of filtering out the noise, the "Flik" node doesn't have to listen to 100 cameras shouting at once. It only listens to the 20% of cameras that actually saw something interesting.
The Benefit: One powerful computer can now handle data from 5 to 6 different camera locations simultaneously, whereas before, it could barely handle one. It's like one expert being able to interview 5 people at once because the other 95 people stayed silent.

Why This Matters (The Real-World Results)

The researchers tested this system in a real outdoor garden with real wind and real insects. Here is what they found:

Massive Data Reduction: Under light wind, the system cut the data sent to the central computer by 60–80%. It's like sending a summary of a book instead of the whole book.
Battery Life: Because the cameras aren't constantly sending useless video, they save energy. The system showed up to 22.6% energy savings, which could extend the life of a battery-powered camera by about a full day.
Speed: The system runs fast enough to catch insects in real-time (30 frames per second) without getting stuck or lagging.
Cost: By using cheap, simple cameras for the "bouncer" job and only a few powerful computers for the "expert" job, you can cover a much larger area for much less money.

The Bottom Line

The paper argues that to monitor insects across a whole city, we can't just build more expensive, powerful cameras. Instead, we need to build smarter, simpler cameras that know how to ignore the wind and only report what matters.

By moving the "thinking" about what is interesting to the edge (the camera itself) and only sending the good stuff to the central brain, the Dot-Flik system makes it possible to create a dense, city-wide network of insect monitors that is affordable, energy-efficient, and scalable. It turns a chaotic flood of data into a manageable stream of useful information.

Technical Summary: Dot-Flik: A Scalable Edge AI Architecture for Distributed Insect Monitoring

Problem Statement
Global insect populations are facing precipitous declines, necessitating scalable, continuous monitoring systems to track biodiversity. While automated vision-based solutions exist, they face fundamental barriers to large-scale deployment. Current state-of-the-art approaches typically rely on expensive, power-hungry hardware or continuous cloud connectivity. These dependencies create a "monitoring gap": high hardware costs limit node density, high power consumption restricts placement to grid-connected areas, and cloud-centric architectures introduce bandwidth and privacy constraints. Furthermore, existing edge AI systems (such as the Flik platform) often operate near computational saturation, where each additional monitoring area requires a proportionally dense deployment of full-processing nodes, leading to unfavorable scaling of cost, energy, and compute resources.

Methodology
The paper proposes Dot-Flik, a distributed, hierarchical IoT architecture that decouples data acquisition from AI classification through edge-side data reduction. The system consists of two distinct layers:

Dot Nodes (Edge Sensing): Low-cost sensing units (implemented on Raspberry Pi Zero 2 W) equipped with standard camera modules. These nodes perform lightweight, motion-informed frame filtering at the source. Instead of transmitting continuous video, they discard background-dominated content locally.
Flik 2.0 Nodes (Edge AI Classification): Centralized processing units (implemented on Raspberry Pi 5 with Hailo-8 NPU) that receive only the filtered candidate frames from multiple Dot nodes. These nodes perform the computationally intensive deep learning inference for insect species classification.

Key Contributions
The authors present three primary contributions:

Motion-Informed Frame Filtering Algorithm: A lightweight algorithm running on the Dot nodes that discards irrelevant frames without requiring deep learning inference on the sensing device. The algorithm utilizes a three-stage workflow:
- Preprocessing: Grayscale conversion, downsampling to 320x320, and 5x5 Gaussian blurring to suppress high-frequency noise.
- Temporal Differencing & Amplification: Calculation of motion likelihood using a rolling buffer of three frames. The motion signal is normalized and gamma-corrected ( $\gamma=3.0$ ) to emphasize concentrated motion (insects) while suppressing dispersed environmental noise (e.g., wind-blown vegetation).
- Block-Based Density Analysis: The frame is partitioned into non-overlapping 20x20 pixel blocks. Using integral images for $O(1)$ complexity, the system calculates motion intensity per block. A frame is retained only if at least one block exceeds a motion threshold ( $\tau=2000$ ), distinguishing localized insect activity from uniform background movement.
Distributed Hierarchical Architecture: The Dot-Flik topology separates sensing from classification, enabling a "many-to-few" scaling model ( $N$ Dot nodes to $M$ Flik nodes, where $M \ll N$ ). This architecture reduces communication volume and allows backend processing demand to scale fractionally with the monitored area rather than linearly with the number of nodes. It supports asynchronous operation, where Dot nodes stream in real-time (30 FPS) while Flik nodes process data in batch mode to optimize computational utilization.
Real-World Validation: The system was deployed outdoors in an urban garden environment under varying wind conditions (no wind to extreme wind). The evaluation covered four axes: real-time performance, network scalability, hardware cost, and energy efficiency.

Results

Data Reduction: The motion-informed filtering achieved frame reduction rates of 60–80% under light-wind conditions (5–15 km/h) and 20–40% under strong winds. Manual inspection confirmed that insect activity visible in unfiltered streams was consistently preserved.
Real-Time Performance: The Dot nodes sustained 30 FPS operation with a total processing time of 33.32 ms per frame, leaving 12.8 ms of computational headroom. Camera capture accounted for 76.8% of the processing time.
Scalability: Based on the light-wind filtering rate ( $r \approx 0.72$ ) and the batch-processing capabilities of the Flik 2.0 node, the authors project a sustainable concurrency of 5–6 concurrent Dot streams per Flik 2.0 node. Network bandwidth was identified as non-binding; the compute capacity of the central node is the primary scalability constraint.
Energy Efficiency: Power consumption correlated with frame transmission rates. The system achieved energy savings of up to 22.6% at 100% frame drop rates (idle). Under moderate wind conditions (70% drop rate), the system demonstrated a 15.6% energy reduction, translating to approximately 9.2 additional hours of operation on a 20,000 mAh power bank.

Significance and Claims
The paper claims that Dot-Flik establishes a practical foundation for dense, low-cost biodiversity monitoring networks in urban environments. By relocating motion-informed filtering to a low-cost edge stage upstream of inference, the architecture addresses the scalability limitations of centralized models. The authors emphasize that the "edge-filter-then-infer" pattern is a necessary design principle for network-scale deployment on commodity hardware, particularly for vision-based IoT workloads where events of interest occur infrequently relative to total data capture.

The authors maintain a modest scope regarding their claims:

They do not claim to have solved the detection rate for all insect species; the evaluation of detection accuracy relies on the upstream Flik pipeline (a separate project) and was qualitatively validated in this study.
The scalability projections (5–6 streams) are derived from single-node measurements; validation with multiple concurrent physical nodes is noted as future work.
The system's effectiveness is currently bounded by wind sensitivity; high wind conditions reduce filtering efficacy due to vegetation motion, though this coincides with naturally suppressed insect activity.
The study focuses on wind variations; broader environmental factors like precipitation, fog, and seasonal illumination changes were not systematically assessed.

The work positions itself as a step toward solving the "monitoring gap" by enabling economically and technically feasible city-scale insect monitoring through architectural optimization rather than solely through algorithmic improvements in classification.