Scholarlink

<p>Unmanned aerial vehicles (UAVs) are increasingly utilized in urban management, environmental monitoring, and emergency response, yet accurate detection of small objects remains highly challenging due to scale variations, complex backgrounds, and dense target distributions. To address these issues, we propose Aero-LiteNet, an efficient detection framework designed for UAV-based small object detection. The network incorporates four key innovations. First, a multi-scale spatial pyramid fusion (MSPF) module extends feature representations through parallel multi-resolution branches and coordinate attention, effectively retaining fine-grained details of small targets. Second, a cross-dimensional aerial small-object attention module (CASAM) establishes collaborative interactions across channel and spatial dimensions, enhancing the discriminability of weak features under occlusion and background clutter. Third, an adaptive spatial feature fusion (ASFF) strategy aligns multi-level features via scale normalization and learnable weight allocation, thereby mitigating inconsistencies and redundancy across feature layers. Finally, a neighborhood-aware IoU (NAIoU) loss function introduces local constraints to penalize excessive overlap in dense scenes, significantly improving bounding box regression accuracy. Comprehensive experiments on the VisDrone2019 benchmark demonstrate that Aero-LiteNet achieves an mAP@50 of 48.1%, surpassing YOLOv8s and several state-of-the-art efficient detectors while maintaining low complexity. Additional validation on RSOD and TT100K datasets further confirms the generalization capability of the proposed model. To validate practical deployment feasibility, we conducted extensive embedded platform experiments on a Rockchip RK3588 system-on-chip (SoC), demonstrating that the INT8-quantized Aero-LiteNet achieves 53 FPS inference throughput while maintaining a competitive mAP@50 of 46.8%, substantially exceeding the 25 FPS real-time requirement for UAV missions. These results indicate that Aero-LiteNet provides an effective and efficient solution for UAV-based small object detection, balancing accuracy and real-time performance for deployment on resource-constrained platforms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Aero-LiteNet: robust aerial small object detection via multi-scale fusion and neighborhood-aware optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月13日

查看全文
  • Hui Liu,
  • Tong Su

摘要

Unmanned aerial vehicles (UAVs) are increasingly utilized in urban management, environmental monitoring, and emergency response, yet accurate detection of small objects remains highly challenging due to scale variations, complex backgrounds, and dense target distributions. To address these issues, we propose Aero-LiteNet, an efficient detection framework designed for UAV-based small object detection. The network incorporates four key innovations. First, a multi-scale spatial pyramid fusion (MSPF) module extends feature representations through parallel multi-resolution branches and coordinate attention, effectively retaining fine-grained details of small targets. Second, a cross-dimensional aerial small-object attention module (CASAM) establishes collaborative interactions across channel and spatial dimensions, enhancing the discriminability of weak features under occlusion and background clutter. Third, an adaptive spatial feature fusion (ASFF) strategy aligns multi-level features via scale normalization and learnable weight allocation, thereby mitigating inconsistencies and redundancy across feature layers. Finally, a neighborhood-aware IoU (NAIoU) loss function introduces local constraints to penalize excessive overlap in dense scenes, significantly improving bounding box regression accuracy. Comprehensive experiments on the VisDrone2019 benchmark demonstrate that Aero-LiteNet achieves an mAP@50 of 48.1%, surpassing YOLOv8s and several state-of-the-art efficient detectors while maintaining low complexity. Additional validation on RSOD and TT100K datasets further confirms the generalization capability of the proposed model. To validate practical deployment feasibility, we conducted extensive embedded platform experiments on a Rockchip RK3588 system-on-chip (SoC), demonstrating that the INT8-quantized Aero-LiteNet achieves 53 FPS inference throughput while maintaining a competitive mAP@50 of 46.8%, substantially exceeding the 25 FPS real-time requirement for UAV missions. These results indicate that Aero-LiteNet provides an effective and efficient solution for UAV-based small object detection, balancing accuracy and real-time performance for deployment on resource-constrained platforms.