Scholarlink

This paper addresses the challenges of segmenting fixed-wing Unmanned Aerial Vehicles (UAVs) and generating rotated bounding boxes in complex aerial scenes. Fixed-wing UAVs are particularly important due to their widespread use in long-endurance missions, where accurate orientation estimation is critical for navigation. Existing UAV tracking datasets lack aerial perspectives and rotation annotations, which provide crucial information of UAVs’ flying conditions, i.e., acceleration and orientation. To bridge this gap, we enhance the UAV2UAV dataset with pixel-level masks and rotated box annotations using a semi-automatic labeling approach, creating a benchmark tailored for segmentation. We also propose an improved PCA-SiamMask model that combines ellipse fitting and principal component analysis (PCA) to refine rotation angle estimation. Compared to baseline methods, our model improves region similarity by up to 2.7% and boosts rotated box AUC by 1.8%, demonstrating greater robustness to pose changes. Ablation studies further confirm the value of fixed-wing-specific training data and the proposed angle optimization strategy, offering practical advances for aerial target segmentation and tracking.

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

Rotated Box-Aware Segmentation for Fixed-Wing UAVs: A PCA-Enhanced SiamMask Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chengwei Li,
  • Yunsong Bai,
  • Ruhai Long,
  • Yong Wang,
  • Shunan Wu,
  • Zhigang Wu

摘要

This paper addresses the challenges of segmenting fixed-wing Unmanned Aerial Vehicles (UAVs) and generating rotated bounding boxes in complex aerial scenes. Fixed-wing UAVs are particularly important due to their widespread use in long-endurance missions, where accurate orientation estimation is critical for navigation. Existing UAV tracking datasets lack aerial perspectives and rotation annotations, which provide crucial information of UAVs’ flying conditions, i.e., acceleration and orientation. To bridge this gap, we enhance the UAV2UAV dataset with pixel-level masks and rotated box annotations using a semi-automatic labeling approach, creating a benchmark tailored for segmentation. We also propose an improved PCA-SiamMask model that combines ellipse fitting and principal component analysis (PCA) to refine rotation angle estimation. Compared to baseline methods, our model improves region similarity by up to 2.7% and boosts rotated box AUC by 1.8%, demonstrating greater robustness to pose changes. Ablation studies further confirm the value of fixed-wing-specific training data and the proposed angle optimization strategy, offering practical advances for aerial target segmentation and tracking.