Scholarlink

<p>Unmanned Aerial Vehicles (UAVs) or drones are pivotal in urban remote sensing, yet single-drone object tracking often falters in complex environments characterised by occlusions, scale variations, and illumination changes. Multi-drone systems offer a solution but face persistent challenges in fusing heterogeneous visual features and facilitating effective cross-platform target recovery. Here, we propose TSMMR, a novel template sharing-enhanced mixed mamba network with a redetection strategy for multi-drone single object tracking. To address scale variance, we design a Group-adaptive Multi-scale Feature Aggregation (GAMFA) module. For effective inter-drone collaboration, we introduce a Dynamic Template Interaction Adapter (DTIA) that enables complementary template fusion. The core of TSMMR is a mixed Vision Mamba backbone, which combines the long-sequence modelling efficiency of state-space models with self-attention blocks to capture global context. Furthermore, we devise a text-guided cross-drone redetection strategy that leverages semantic similarity to re-localise lost objects by referencing drones with higher tracking confidence. Experimental results on the MDOT dataset demonstrate that TSMMR achieves a success rate of 58.4% and a precision of 75.1%, outperforming existing state-of-the-art multi-drone trackers. This work underscores the potential of integrating state-space models with semantic cues to advance robust collaborative perception in aerial systems. Our source code is available at: <a href="https://github.com/JerryBryant24/TSMMR">https://github.com/JerryBryant24/TSMMR</a>.</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

TSMMR: a mixed Mamba network with cross-drone template fusion and text-guided redetection for robust aerial tracking

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Xiangqian Liu,
  • Guangwei Zhang,
  • Yifan Zhu,
  • Bing Zhou

摘要

Unmanned Aerial Vehicles (UAVs) or drones are pivotal in urban remote sensing, yet single-drone object tracking often falters in complex environments characterised by occlusions, scale variations, and illumination changes. Multi-drone systems offer a solution but face persistent challenges in fusing heterogeneous visual features and facilitating effective cross-platform target recovery. Here, we propose TSMMR, a novel template sharing-enhanced mixed mamba network with a redetection strategy for multi-drone single object tracking. To address scale variance, we design a Group-adaptive Multi-scale Feature Aggregation (GAMFA) module. For effective inter-drone collaboration, we introduce a Dynamic Template Interaction Adapter (DTIA) that enables complementary template fusion. The core of TSMMR is a mixed Vision Mamba backbone, which combines the long-sequence modelling efficiency of state-space models with self-attention blocks to capture global context. Furthermore, we devise a text-guided cross-drone redetection strategy that leverages semantic similarity to re-localise lost objects by referencing drones with higher tracking confidence. Experimental results on the MDOT dataset demonstrate that TSMMR achieves a success rate of 58.4% and a precision of 75.1%, outperforming existing state-of-the-art multi-drone trackers. This work underscores the potential of integrating state-space models with semantic cues to advance robust collaborative perception in aerial systems. Our source code is available at: https://github.com/JerryBryant24/TSMMR.