Visual Multiple Target Tracking From a Descending Aerial Platform
Visual Multiple Target Tracking From a Descending Aerial Platform
复制标题
DOI：
10.23919/acc.2018.8431915
复制
发表时间：
2018-06
期刊：
影响因子：
--
通讯作者：
Parker C. Lusk;R. Beard
中科院分区：
文献类型：
--
作者：
Parker C. Lusk;R. Beard
演示了一种实时视觉多目标跟踪器。移动地面目标的测量使用Kanade-Lucas-Tomasi （KLT）跟踪方法生成。基于同形图的图像配准用于将测量值对齐到同一坐标框架中，从而允许检测独立移动的物体。近年来发展起来的递归ransac算法利用视觉测量来估计杂波中的目标。高度相关的调整增加了飞行器下降过程中轨迹的连续性和覆盖范围。该算法不需要操作者的干预，提高了无人机系统的态势感知能力。在gpu和cpu上分析了实时跟踪效率。使用MOTA和MOTP度量来展示和讨论跟踪结果。
A real-time visual multiple target tracker is demonstrated onboard a descending multirotor. Measurements of moving ground targets are generated using the Kanade-Lucas-Tomasi (KLT) tracking method. Homography-based image registration is used to align the measurements into the same coordinate frame, allowing for the detection of independently moving objects. The recently developed Recursive-RANSAC algorithm uses the visual measurements to estimate targets in clutter. Altitude-dependent tuning increases track continuity and coverage during the descent of the vehicle. The algorithm requires no operator interaction and increases the situation awareness of the unmanned aerial system. Real-time tracking efficiency is analyzed on GPUs and CPUs. Tracking results are presented and discussed using the MOTA and MOTP metrics.