Scholarlink

We present a lightweight, geometry-driven multi-object tracker for UAV sense-and-avoid on embedded platforms. Independently moving features are segmented via robust background motion (USAC/MAGSAC) and HDBSCAN; clusters are converted to boxes and tracked with a constant-velocity Kalman filter and IoU–Hungarian association. On a 14,250-frame X-Plane11 simulation evaluated with HOTA/TrackEval, our method improves association accuracy (AssA) by +95.8% over a YOLOv8 baseline and +116% over an AVOIDDS baseline under the same protocol, while running in real time on Jetson Orin NX. Qualitative EO/IR results show robustness to very small targets ( \(\le 8\times 8\)  px).

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

A Geometry-Driven Approach to Detecting and Tracking Multiple Aerial Objects

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jen-Jui Liu,
  • Curtis P. Evans,
  • Randal W. Beard

摘要

We present a lightweight, geometry-driven multi-object tracker for UAV sense-and-avoid on embedded platforms. Independently moving features are segmented via robust background motion (USAC/MAGSAC) and HDBSCAN; clusters are converted to boxes and tracked with a constant-velocity Kalman filter and IoU–Hungarian association. On a 14,250-frame X-Plane11 simulation evaluated with HOTA/TrackEval, our method improves association accuracy (AssA) by +95.8% over a YOLOv8 baseline and +116% over an AVOIDDS baseline under the same protocol, while running in real time on Jetson Orin NX. Qualitative EO/IR results show robustness to very small targets ( \(\le 8\times 8\)  px).