💻 computer science

EMA-Det: Hardware-Aware Edge Verification for Two-Stage UAV Detection

This paper introduces EMA-Det, a hardware-aware, early-exit region-verification network for two-stage UAV detection that prioritizes recall over mAP to overcome system bottlenecks, achieving superior performance on edge devices and demonstrating the necessity of domain-specific training for thermal-IR applications.

Original authors: Enes Mert AYDIN, Osman TOKLUOGLU, Enver CAVUS

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Enes Mert AYDIN, Osman TOKLUOGLU, Enver CAVUS

Original paper licensed under CC BY 4.0 (https://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

In the skies above our cities and borders, a new kind of silent challenge has emerged: small, slow, and low-flying drones. These unmanned aircraft, often no larger than a bird, can slip past traditional radar and cameras, posing risks to airports, stadiums, and critical infrastructure. Defending against them requires a system that can see, decide, and act in a fraction of a second. However, the computers that would need to run these defenses are often small, battery-powered devices with limited memory and processing power. They cannot simply scan a massive, high-definition video feed with the same heavy software used in a supercomputer; the data stream is too wide, and the memory too narrow. To solve this, engineers often use a two-step approach. First, a quick, simple sensor scan flags a few small areas of the sky that might contain a drone. Then, a second, more careful system zooms in on just those tiny patches to confirm whether a real threat is there. The success of the entire defense depends on this second step: if it misses a drone, the system fails, regardless of how well the first step worked.

Researchers at Ankara Yıldırım Beyazıt University have developed a new method called EMA-Det to make this second step faster and more reliable on these small, edge devices. Instead of trying to build a single, massive network that does everything, they designed a specialized verifier that understands the unique nature of the task. They realized that most of the small image patches sent for verification are actually empty sky. A standard computer vision system would waste time analyzing every single patch with the same heavy effort, whether it contains a drone or just blue sky. The new system changes this by adding "early exits." As the image data flows through the network, simple checks happen at the very beginning. If the system is highly confident that a patch is just empty sky, it stops processing immediately and discards it, saving valuable time and energy. Only the patches that look suspicious are allowed to travel through the deeper, more complex layers of the network. This design turns the fact that most candidates are harmless into a speed advantage, allowing the device to handle more potential threats without slowing down.

To teach this system what to look for, the researchers had to be extremely careful about the data they used. They gathered thousands of photos of drones and empty skies from public sources, but they noticed that many images were nearly identical copies taken from the same video frames. To ensure the system was truly learning and not just memorizing duplicates, they created a strict cleaning process. They removed the duplicates and split the remaining unique images into training and testing groups, making sure no single photo appeared in both. They also started the system's learning process with a clever trick: instead of letting the computer guess what edges and shapes look like from scratch, they gave it a head start by programming the first layer to recognize basic lines and edges, much like a human sketching the outline of an object before filling it in. This physical head start helped the system learn faster and more stably.

The results of this approach were striking when tested against standard methods. While other popular detection systems, which are designed to scan entire wide-angle views, managed to find about 74 percent of the drones in these small patches, the new system found nearly 96 percent. The researchers found that the standard way of measuring success, which looks at a complex average score, was actually misleading for this specific job. A system could look good on that average score but still miss too many drones, which is a fatal flaw for a defense system. By focusing on the ability to catch every single target, the new design proved far superior. The team also tested how well the system would work if trained on visible light cameras but then asked to look at thermal infrared images, which show heat rather than color. Without any retraining, the system still recognized drones with high precision, though it missed more of them than it did with visible light. However, when they trained the system from scratch using thermal images, it performed exceptionally well, proving that the gap between the two types of vision was due to the difference in the images themselves, not a limit of the system's design.

To prove this system could actually run on real hardware, the team took the first part of their network and built a digital blueprint for a specific type of chip used in field devices. They verified that this blueprint would fit within the strict memory and power limits of the chip, confirming that the design is not just a theory but a practical solution for real-world deployment. The work demonstrates that for protecting against small, slow drones, the key is not just building a smarter brain, but building a brain that knows when to stop thinking. By letting the system quickly dismiss the empty sky, it saves its energy for the moments that matter, ensuring that when a drone is present, it is seen.

Technical Summary: EMA-Det

Problem Statement
The detection of Low-Slow-Small (LSS) Unmanned Aerial Vehicles (UAVs) on edge devices faces a critical bottleneck in the "sense-decide-act" loop. While autonomous defense systems typically employ a two-stage pipeline—where a first stage (e.g., radar, motion detection) proposes Regions of Interest (RoIs) and a second stage verifies them—current approaches often adapt general-purpose full-frame detectors (like YOLO) to the verification task. The authors argue this is suboptimal for two reasons:

Resource Constraints: Edge devices, particularly FPGA-based systems, are often bottlenecked by memory bandwidth rather than raw compute (FLOPs). Standard lightweight backbones (e.g., depthwise-separable convolutions) can exacerbate memory pressure on BRAM-constrained fabrics.
Metric Misalignment: In a two-stage cascade, the total system recall ( $R_{total}$ ) is the product of the proposal recall and the verifier recall ( $R_{proposal} \times R_{verifier}$ ). A verifier optimized for Mean Average Precision (mAP) may fail to minimize false negatives, creating a system-level bottleneck. The verification stage specifically requires high recall to prevent system-level misses, even if it means sacrificing some mAP.

Methodology: EMA-Det
The paper proposes EMA-Det, a hardware-aware, two-stage region verification network designed specifically for the second stage of UAV detection. Key architectural and methodological components include:

Physically Initialized First Layer: To address the limited spatial information in small targets (128×128 resolution), the first convolutional layer is initialized with a deterministic bank of Laplacian and Sobel operators. This provides an inductive bias for edge detection, reducing training variance compared to random initialization.
Dynamic Early-Exit Gates: Leveraging the asymmetric distribution of candidate regions (where most are background), the network incorporates two confidence-gated early-exit points (after Block 2 and Block 4). If a candidate is confidently classified as background, the inference halts, skipping subsequent layers. This converts linear latency scaling into an asymptotic one, significantly reducing average inference time.
Decoupled Spatial Bounding-Box Head: Unlike standard architectures that pool features before regression, EMA-Det decouples the classification head (global average pooled) from the bounding-box refinement head. The box head reads the pre-pooling feature map directly to preserve spatial layout information, improving localization without significant parameter overhead.
Loss Function: The network utilizes a combined loss of Binary Cross-Entropy for classification, Complete-IoU (CIoU) for box regression, and additional Binary Cross-Entropy terms for the early-exit gates.
Hardware-Aware Design: The architecture avoids depthwise-separable convolutions in favor of standard convolutions to mitigate memory bandwidth pressure on FPGAs.

Experimental Setup
The study employs a rigorous, controlled evaluation protocol:

Dataset Construction: A source-level-split, deduplicated dataset was created from public UAV and sky/bird imagery. An 8×8 average-hash comparison removed near-duplicates, ensuring a strict train/test boundary with no source-level overlap. The final corpus contains 4,136 unique patches.
Training Protocol: All models were trained for 150 epochs using AdamW, cosine annealing, and N=5 random seeds to ensure statistical robustness. Data augmentation excluded Mosaic/MixUp to preserve small-target shapes.
Baselines: Comparisons were made against YOLOv8n, MobileNetV3-Small, and ShuffleNetV2, all trained from scratch with identical heads and protocols.
Hardware Validation: A static resource analysis and synthesized implementation of the network's first block were performed on a Zynq-7020 FPGA using Vitis HLS 2025.2.

Key Results

Recall vs. mAP: EMA-Det achieved a recall of 96.08 ± 0.47%, significantly outperforming the full-detection baseline YOLOv8n (73.75 ± 1.86%) under identical conditions. While YOLOv8n achieved higher mAP (49.59% vs. 28.04%), the authors demonstrate that the lower recall of YOLOv8n would bottleneck the overall system performance.
Ablation Studies:
- CBAM: The inclusion of CBAM attention modules did not improve mAP but reduced recall stability and average recall. The authors conclude that CBAM's benefits in other contexts may substitute for spatial information lost by pooling, which is mitigated by EMA-Det's decoupled box head.
- Spatial Box Head: Decoupling the box head improved mAP@0.75 by 5.7 points and recall stability (reducing standard deviation from 2.54 to 0.47) with negligible FLOP cost.
Early-Exit Efficiency: By tuning the exit threshold ( $\tau$ ), the system can save up to 21.2% of FLOPs while maintaining mAP, with a significant portion of background candidates exiting early.
Cross-Sensor Generalization: When an EO-trained EMA-Det was applied zero-shot to thermal-IR data (Anti-UAV410 benchmark), precision remained high (~95%), but recall dropped to 80–86% and mAP to ~39–52%. However, training EMA-Det from scratch on thermal data achieved 89.52% mAP@0.5 and 91.64% recall, confirming the performance gap is due to domain shift rather than architectural limitations.
Hardware Feasibility: The synthesized Block 1 on Zynq-7020 (using 16-bit fixed-point quantization) utilized 65% of the DSP budget and 26% of LUTs, achieving a latency of ~365 µs/image. Full unrolling of deeper blocks was deemed infeasible due to DSP constraints, necessitating time-multiplexed MAC arrays for the full network.

Significance and Claims
The paper claims that EMA-Det represents a shift in design philosophy for edge-based UAV detection:

Metric Reorientation: It argues that for the verification stage of a two-stage system, recall is the critical optimization target, not mAP. Optimizing for mAP can lead to system-level failures.
Hardware-Aware Architecture: It demonstrates that shallow, streaming-friendly architectures with standard convolutions are more suitable for FPGA edge deployment than deeper, FLOPs-optimized backbones that suffer from memory bandwidth bottlenecks.
Methodological Rigor: The authors emphasize the importance of source-level deduplication and strict train/test boundaries in UAV detection research to prevent data leakage and overestimation of performance.
Domain Shift Quantification: The study provides a quantitative baseline for the EO-to-thermal domain shift, showing that while zero-shot transfer is imperfect, the architecture is capable of high performance when trained on the target modality.

The authors conclude that while the full FPGA mapping of Blocks 2–6 remains future work, the verified implementation of Block 1 and the algorithmic design substantiate the viability of EMA-Det for resource-constrained edge verification.

Technical Summary: EMA-Det

More like this