Scholarlink

<p>With the rapid advancements in unmanned aerial vehicle (UAV) technologies, UAV-based visual object tracking has increasingly garnered attention. However, nighttime tracking remains challenging due to poor lighting conditions, which often impede the effectiveness of UAV trackers. To address this challenge, we introduce a novel nighttime UAV tracking method through joint low-light image enhancement using a hybrid attention transformer and denoising (named JLEDNet). In the first stage of JLEDNet (JLEDNet-stageI), we employ a U-shaped architecture for low-light image enhancement, proposing the spatial-channel hybrid attention transformer-based block (SCHATB) to enhance the network’s modeling capability, and introducing the spatial-frequency convolutional block (SFCB) to improve global feature learning. In JLEDNet-stageII, we introduce the corresponding denoising scheme to remove the latent noise in the darkness amplified after the first stage. Extensive evaluation on multiple nighttime public UAV benchmarks, including UAVDark135, DarkTrack2021, UAVDark70, and NAT2021-L, demonstrate that JLEDNet significantly enhances nighttime tracking performance.</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

JLEDNet: a nighttime UAV tracking method through joint low-light image enhancement using hybrid attention transformer and denoising

  • 期刊论文

  • 发表日期：

    2025年1月22日

查看全文
  • Yanmei Li,
  • Tao Yu,
  • Jian Luo,
  • Xiaoshuang Li,
  • Jingshi Deng,
  • Qibin Yang

摘要

With the rapid advancements in unmanned aerial vehicle (UAV) technologies, UAV-based visual object tracking has increasingly garnered attention. However, nighttime tracking remains challenging due to poor lighting conditions, which often impede the effectiveness of UAV trackers. To address this challenge, we introduce a novel nighttime UAV tracking method through joint low-light image enhancement using a hybrid attention transformer and denoising (named JLEDNet). In the first stage of JLEDNet (JLEDNet-stageI), we employ a U-shaped architecture for low-light image enhancement, proposing the spatial-channel hybrid attention transformer-based block (SCHATB) to enhance the network’s modeling capability, and introducing the spatial-frequency convolutional block (SFCB) to improve global feature learning. In JLEDNet-stageII, we introduce the corresponding denoising scheme to remove the latent noise in the darkness amplified after the first stage. Extensive evaluation on multiple nighttime public UAV benchmarks, including UAVDark135, DarkTrack2021, UAVDark70, and NAT2021-L, demonstrate that JLEDNet significantly enhances nighttime tracking performance.