Scholarlink

Nighttime UAV tracking poses significant challenges due to low-light conditions, low resolution, and poor contrast, making it inherently more difficult than generic object tracking and daytime UAV tracking. To address these challenges, we propose SPAR-T, a novel Self-Prompting AutoRegressive Tracker that combines self-prompting vision-language learning with autoregressive trajectory prediction. Specifically, SPAR-T employs a self-prompting multimodal encoder to dynamically integrate visual inputs and language descriptions, along with an autoregressive decoder to iteratively generate accurate target coordinates by conditioning on the historical object trajectory. Autoregressive temporal modeling coupled with integrated vision-language representations enables robust performance in nighttime environments. To support research in this domain, we construct NUT318, a large-scale dataset comprising 318 video sequences with 39 object categories, offering 207K frames annotated with high-quality bounding boxes, language descriptions, and 14 attributes—filling a critical gap in nighttime UAV tracking. Compared to existing nighttime UAV tracking datasets, the proposed NUT318 offers greater diversity (e.g., scenes, categories) and richer annotations, enabling a more thorough evaluation of tracking algorithms. Extensive experiments demonstrate that SPAR-T consistently outperforms state-of-the-art methods on both NUT318 and existing nighttime UAV tracking benchmarks. Our dataset and code will be released at here .

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

Boosting Nighttime UAV Tracking via Self-prompting Autoregressive Learning and a New Benchmark

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chunhui Zhang,
  • Li Liu,
  • Hao Wen,
  • Xi Zhou,
  • Yanfeng Wang

摘要

Nighttime UAV tracking poses significant challenges due to low-light conditions, low resolution, and poor contrast, making it inherently more difficult than generic object tracking and daytime UAV tracking. To address these challenges, we propose SPAR-T, a novel Self-Prompting AutoRegressive Tracker that combines self-prompting vision-language learning with autoregressive trajectory prediction. Specifically, SPAR-T employs a self-prompting multimodal encoder to dynamically integrate visual inputs and language descriptions, along with an autoregressive decoder to iteratively generate accurate target coordinates by conditioning on the historical object trajectory. Autoregressive temporal modeling coupled with integrated vision-language representations enables robust performance in nighttime environments. To support research in this domain, we construct NUT318, a large-scale dataset comprising 318 video sequences with 39 object categories, offering 207K frames annotated with high-quality bounding boxes, language descriptions, and 14 attributes—filling a critical gap in nighttime UAV tracking. Compared to existing nighttime UAV tracking datasets, the proposed NUT318 offers greater diversity (e.g., scenes, categories) and richer annotations, enabling a more thorough evaluation of tracking algorithms. Extensive experiments demonstrate that SPAR-T consistently outperforms state-of-the-art methods on both NUT318 and existing nighttime UAV tracking benchmarks. Our dataset and code will be released at here .