• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Lightweight AI Network Enables Real-Time Aerial Surveillance

by
November 9, 2026
in Technology
Reading Time: 6 mins read
0
Lightweight AI Network Enables Real-Time Aerial Surveillance

Lightweight AI Network Enables Real-Time Aerial Surveillance

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

The study, published in Neural Computing and Applications, addresses the significant performance degradation that standard visible-light cameras experience in low-illumination environments. By fusing visible light imagery with infrared thermal data, the proposed model maintains high detection accuracy during nighttime operations while significantly reducing the computational overhead typically associated with multimodal fusion architectures.

The core innovation of the framework is a guided cross-modal fusion module that adaptively balances the contributions of visible and infrared features. Traditional fusion methods often introduce feature redundancy and high computational costs, making them unsuitable for the resource-constrained hardware found on drones. This new approach utilizes a difference-guided channel enhancement mechanism to isolate unique information from each modality, effectively suppressing redundant background noise and highlighting discriminative object features that remain visible even when lighting conditions are poor.

To further optimize efficiency, the researchers integrated a custom PCShuffle module into the YOLOv8 nano backbone. This module combines partial convolution, channel shuffling, and pointwise convolution to reduce the number of parameters and floating-point operations required for feature extraction.

The model was rigorously evaluated on four distinct datasets: DroneVehicle, VEDAI, RGBT-Tiny, and LLVIP. These datasets cover a wide range of environmental conditions, including daytime, night, and dark-night scenarios, as well as various flight altitudes and viewing angles. The DroneVehicle dataset, for instance, contains over 28,000 paired images with nearly one million annotated vehicle instances, providing a robust benchmark for testing the model’s ability to detect objects under challenging illumination.

On the DroneVehicle test set, the proposed PCShuffle+GCMF variant achieved a mean average precision of 78.0% at an Intersection over Union threshold of 0.5. This performance was attained with only 2.74 million parameters, representing a substantial reduction in model size compared to existing state-of-the-art methods. The model also demonstrated strong capability in detecting small targets, outperforming several lightweight baselines in the small-object scale category while maintaining a compact architecture.

Performance on the VEDAI dataset, which focuses on small object detection in remote sensing imagery, was particularly notable. The model achieved a mean average precision of 93.0%, significantly outperforming compared methods. It obtained the highest accuracy for specific vehicle categories, including trucks, tractors, and campers. The high performance on this dataset confirms the framework’s effectiveness in identifying small, distant objects from aerial viewpoints, a critical requirement for surveillance and search-and-rescue missions.

Real-time performance was a primary design goal, and the model met this objective on both desktop and embedded hardware. On an NVIDIA RTX 4080 laptop GPU, the model processed images in 2.9 milliseconds. On the Jetson Orin AGX, it achieved a throughput of 35.71 frames per second, while on the Jetson Orin Nano, it reached 26.3 frames per second.

Energy efficiency was also evaluated to ensure suitability for UAV platforms with limited power resources. The model demonstrated low energy consumption per frame across different power modes. On the Orin Nano in a 15-watt mode, the energy cost was calculated to be 0.053 Joules per frame. On the Orin AGX, the efficiency remained consistent at approximately 0.067 Joules per frame in the maximum performance configuration. This balance between inference speed and power consumption is crucial for extending the operational time of autonomous drones.

Ablation studies confirmed the individual contributions of the proposed modules. The guided cross-modal fusion module improved detection accuracy by effectively modeling complementary cross-modal cues, particularly in low-illumination conditions where simple fusion strategies like concatenation or addition failed. The PCShuffle backbone reduced the parameter count by 37% compared to the standard YOLOv8 nano variant while maintaining competitive detection performance. The inclusion of learnable parameters in the fusion process further enhanced accuracy without adding computational overhead, allowing the network to adaptively adjust modality contributions based on the input scene.

Qualitative analysis using Grad-CAM visualizations revealed that the proposed model produces more focused and localized activations compared to baseline models. While the baseline model exhibited diffuse activations across irrelevant background regions, the proposed framework concentrated its feature responses on the target objects. This improved localization is particularly evident in low-light scenarios, where the model successfully suppresses redundant background responses and enhances the representation of object-relevant regions.

The researchers also conducted a cross-dataset evaluation by training the model on the DroneVehicle dataset and testing it on the VEDAI dataset without fine-tuning. Despite the differences in imaging conditions and object distributions, the model maintained competitive detection performance. This robustness suggests that the framework can generalize well to unseen environments, a valuable trait for real-world surveillance applications where conditions are unpredictable.

Limitations of the model were identified in extreme scenarios. Under complete nighttime darkness, where the visible light modality provides no discriminative information, the model relies entirely on the infrared modality. In cases of thermal crossover, where different object types have similar thermal signatures, the model occasionally misclassifies trucks as freight cars. Additionally, sensor misalignment between the visible and infrared cameras can lead to localization errors, as the model may detect objects based on clearer edges in the infrared image rather than the actual positions in the visible image.

The study concludes that the proposed lightweight framework offers a favorable balance between detection accuracy and computational efficiency for aerial surveillance. By integrating efficient backbone design with adaptive multimodal fusion, the model achieves real-time performance on edge devices while maintaining high accuracy in challenging lighting conditions. Future work will focus on further deployment optimization through network pruning and quantization to reduce memory usage and power consumption on resource-constrained platforms.

The proposed framework is built upon the YOLOv8 nano architecture, a baseline chosen for its suitability in real-time detection tasks. The backbone network typically accounts for more than 50% of the total inference cost in general object detectors, making its optimization critical for edge deployment. To address this, the researchers developed a PCShuffle module that replaces the standard C2f modules in the YOLOv8 backbone. This module utilizes partial convolution, which processes only a fraction of the input channels, specifically one-quarter, to reduce computational complexity. By allowing the remaining channels to bypass the convolution, the design ensures regular memory access patterns and reduces redundancy. The partial convolution is followed by a single 1×1 pointwise convolution to integrate information across channels, replacing the two-layer expansion-projection structure found in the original FasterNet module. A skip connection is included to preserve feature details essential for small-object detection, while channel shuffling is applied before the partial convolution to enhance feature diversity and information exchange.

The fusion mechanism distinguishes itself from traditional attention-based methods by explicitly modeling and enhancing cross-modal difference features. In conventional approaches, difference information is often used merely to generate attention weights that reweight the original features. In contrast, the Guided Cross-Modal Fusion module applies weighting functions directly to the difference representations between the RGB and infrared feature maps. This process involves computing global channel statistics via global average pooling and applying a tanh activation to generate bidirectional modulation weights. The tanh function maps values into the range of negative one to positive one, allowing the network to assign negative weights to less informative or corrupted channels and positive weights to useful ones. This bidirectional modulation is particularly advantageous when one modality is degraded, as it enables the suppression of noisy channels and the highlighting of meaningful cues, resulting in a cleaner and more reliable fused representation.

Learnable parameters are integrated into the fusion process to dynamically adjust the relative importance of each modality during training. These scalar weights, initialized to one, are optimized via backpropagation to balance the contributions of the original features and their cross-modal differences. The ablation studies indicate that the inclusion of these learnable parameters improves detection accuracy without introducing additional computational overhead. On the DroneVehicle dataset, this adjustment resulted in a 0.9% increase in mean average precision, while on the VEDAI dataset, the improvement was 5.3%. This adaptive mechanism allows the network to respond to varying scene characteristics, ensuring that complementary information is preserved while preventing feature degradation during the fusion process.

The spatial-channel decoupled downsampling module further reduces computational costs by separating the spatial and channel transformation stages. Traditional downsampling using standard convolutions with a stride of two simultaneously reduces spatial resolution and doubles the number of channels, incurring significant floating-point operations. The proposed module first applies pointwise convolution to refine channel dimensions and capture cross-channel interactions, followed by a 3×3 depthwise convolution with a stride of two to reduce feature map dimensions. This decoupling strategy lowers the parameter count and computational complexity while preserving performance.

The evaluation protocol for the VEDAI dataset involved a 10-fold cross-validation approach to ensure robust performance assessment. The dataset was split into 10 folders, each containing 1,089 images for training and 121 for testing. The reported results for the proposed model correspond to the average of the 10-fold testing results, providing a comprehensive measure of its generalization capability. For ablation studies, a single fold was used to isolate the effects of specific modules. The input resolution for all experiments was standardized to 640×640 pixels, following the default configuration of the YOLOv8 framework. Images were resized using a letterbox strategy to preserve the original aspect ratio, which reduces computational load while maintaining sufficient spatial detail for accurate detection.

Subject of Research: Lightweight cross-modal RGB-IR fusion for real-time aerial object detection

Article Title: Lightweight cross-modal RGB-IR fusion network for real-time aerial surveillance

Article References: Habash, N., Abu Alqumsan, A., & Zhou, T. (2026). Lightweight cross-modal RGB-IR fusion network for real-time aerial surveillance. Neural Computing and Applications, 38(21), Article 878. https://doi.org/10.1007/s00521-026-12529-7

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12529-7

Keywords: object detection, RGB-IR fusion, unmanned aerial vehicles, real-time inference, edge computing, aerial surveillance, deep learning, low-light vision, YOLOv8, multimodal learning, Lightweight, cross-modal

News Source: Denise Maddox. (November 9, 2026). Lightweight AI Network Enables Real-Time Aerial Surveillance. Scienmag.

Tags: aerial surveillancecross-modaldeep learningEdge Computinglightweightlow-light visionmultimodal learningobject detectionreal-time inferenceRGB-IR fusionunmanned aerial vehiclesYOLOv8
Share12Tweet7Share2ShareShareShare1

Related Posts

Fungal Mycelium Optimization Algorithm Solves Complex Engineering and Benchmark Problems

Fungal Mycelium Optimization Algorithm Solves Complex Engineering and Benchmark Problems

November 17, 2026
Tuned XGBoost Predicts Abrasive Water Jet Machining Responses for Titanium Alloy

Tuned XGBoost Predicts Abrasive Water Jet Machining Responses for Titanium Alloy

November 17, 2026

Biopolymer Hydroxyapatite Nanocomposite Strengthens Portland Cement and Accelerates Early Hydration

November 17, 2026

Oil Palm Carbon Modified With Geothermal Silica Shows Coin Cell Anode Potential

November 17, 2026

POPULAR NEWS

  • Four Key Clinical Factors Associated With Heart Injury in Gastrointestinal Bleeding

    Four Key Clinical Factors Associated With Heart Injury in Gastrointestinal Bleeding

    29 shares
    Share 12 Tweet 7
  • Nanjing Hospital Study Finds 19.2 Percent HPV Positivity Among 39,335 Women

    29 shares
    Share 12 Tweet 7
  • Specialized EBI3-High Dendritic Cell Subset Identified in Rheumatoid Arthritis Synovial Fluid

    29 shares
    Share 12 Tweet 7
  • Early Tildrakizumab Treatment Predicts Complete Psoriasis Clearance in Five-Year Italian Study

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Review Outlines Deep Phenotyping Frameworks to Advance Precision Medical Genetics

Aspergillus tubingensis Yields Bioactive Antioxidant Pyrones Amid Melanin Extraction Challenges

Four Key Clinical Factors Associated With Heart Injury in Gastrointestinal Bleeding

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.