Author: Zhichao Han, Ruibin Zhang, Neng Pan*, Chao Xu, and Fei Gao
Publisher: ICRA 2021
Publish year: 2021
Editor: 柯西

Abstract/摘要

本文给出了在目标行动未知,稠密环境未知的情况下,无人机跟踪的系统解决方案.
本文主要包括两部分:

  1. 目标轨迹预测:利用观测到的目标历史信息,在考虑目标动态约束的情况下预测未来轨迹
  2. 跟踪轨迹规划:传统规划结构,前端采用考虑运动学的搜索方法,后端采用时空优化和无碰撞的轨迹优化方法

    Instruction/介绍

    在无人机跟踪领域,首先要在有限传感器条件下识别目标和周围障碍物,然后规划出可行的安全轨迹.此外,为了处理意外情况,需要高频率的重规划过程.我们提出一个目标意图未知,周围环境未知的无人机跟踪系统框架.
    目标轨迹预测及重定位:
    将目标近似为刚体,假设目标运动速度和加速度有界且连续.对于观测过的历史轨迹使用polynomial regression,然后用Bernstein basis polynomial进行动态约束,生成的轨迹作为目标未来运动预测.此外,针对障碍物遮挡,传感器范围限制,目标运动未知时,无人机难以定位目标的情况,我们设计了目标重定位策略.
    跟踪轨迹生成:
    设计一个考虑当前目标观测和未来运动预测的启发式搜索函数,然后在已搜索的路径上生成安全走廊,后端优化在安全走廊内生成时空优化安全轨迹.
    本文贡献:
    我们将上述方法集成到无人机系统上,采用多个相机提升传感器性能.在真实环境中实验,并与其他方法相比,更有优势.本文贡献如下:

  3. 轻量级且目标运动未知的基于贝塞尔多面体约束的运动预测方法.

  4. 前端:一个考虑目标的运动学搜索方法;后端:时空轨迹优化方法
  5. 结合传感器,形成系统解决方案,并进行评估

image.png

目标运动预测

带约束的贝塞尔预测轨迹

采用贝赛尔曲线描述目标预测轨迹,n阶贝塞尔曲线表示如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图2
将观测到的世界坐标系下的目标位置Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图3及时间戳Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图4存入长度为L的FIFO队列中,其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图5.Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图6时间范围是Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图7等于当前时刻.当获取到新的目标观测,调整历史观测生成新的目标预测轨迹Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图8.预测的时间范围是Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图9.如图3
image.png
历史观测的可信度随时间递减,因此,观测到的时间越靠前,在损失函数中权重越低。添加权重Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图11判断不同时间戳观测的置信度。采用双曲切线tanh(x)计算权重:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图12
函数值随Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图13值的增加快速衰减,可以有效判断不同观测的有效性。整体损失函数如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图14

使目标轨迹和观测距离差最小。为避免过拟合,增加二阶正交项,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图15为该项权重。
image.png
为了保证目标预测轨迹的动态可行性,预测速度和加速度约束为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图17.根据贝塞尔曲线的凸包属性和hodograph属性(详见运动规划课程第六节),运动学约束如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图18
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图19是贝塞尔曲线阶次,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图20是时间尺度。
贝塞尔曲线的二范数正定,所以该问题是带约束的QP问题。我们采用OOQP开源求解器求解该问题。

目标重定位

在无人机丢失目标观测后,需要重新定位目标位置。在多数情况下,无人机快速到达目标最后观测位置,并沿预测轨迹运动可以重新获取到目标。详细步骤如下:

  1. 采用路径搜索算法得到飞机当前位置到目标丢失最后观测位置的最短路径Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图21,然后离散最后预测轨迹。如果预测轨迹上有障碍,采用路径搜索算法生成无障碍路径Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图22.Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图23Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图24构成重定位路径Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图25
  2. 基于Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图26生成安全走廊
  3. 在安全走廊中生成时空优化和无障碍轨迹,作为重定位轨迹。
  4. 在重定位步骤中,如果在重定位轨迹上有新的障碍,则重新生成新的轨迹。

过程如图4:
image.png

安全轨迹生成

目标导向的运动学跟踪路径搜索

运动学搜索方法基于hybrid A算法,主要包括邻节点拓展和启发式函数设计两部分:
*邻节点拓展:

  • 无人机状态向量:Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图28
  • 控制输入:加速度,用Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图29,将加速度离散,得到离散序列Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图30
  • 拓展邻节点的时间序列为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图31

拓展节点公式如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图32

其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图33是上一帧状态向量,给出Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图34可以得到基于运动学的拓展节点。每次拓展步骤可以生成Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图35个拓展节点
损失函数:
每个节点的损失函数为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图36,其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图37表示初始状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图38到当前状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图39的实际消耗,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图40是使搜索更快速的启发式函数。
采用最小化能量-时间损失函数如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图41
因为拓展节点由离散输入Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图42组成,定义其损失函数为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图43,若路径由Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图44个拓展节点组成,则
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图45
启发式损失函数分为两部分:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图46
其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图47是无人机当前状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图48到目标状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图49的距离。其中目标状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图50由跟踪目标的当前状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图51和预测Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图52时刻后状态Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图53的中间量得到,公式如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图54
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图55为权重,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图56由预测轨迹Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图57得到,速度量是Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图58的一阶导数。Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图59
两点边界值问题(OBVP)用于求解最小化动态损失函数的最优路径。定义最小化损失函数为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图60Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图61的OBVP距离,即Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图62,过程描述见下图:
image.png
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图64是时间惩罚函数,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图65是期望拓展时间的和,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图66是权重,该项能够加速搜索,因为它倾于选择当前节点的相邻区域,而非整个状态空间。增加该项通过牺牲得到最优解而降低计算量,但在结果中我们仍然能够找到可行且满足条件的解。

时空优化轨迹生成

在后端,我们采用文章[2]“Z. Wang, C. Xu, and F. Gao, “Generating large-scale trajectories effi-
ciently usingdouble descriptions of polynomials””中的优化方法,在安全走廊中生成时空优化轨迹Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图67
安全走廊Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图68定义如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图69
其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图70表示小方块
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图71
损失函数如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图72
其中平滑项Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图73如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图74
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图75是保证Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图76Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图77相交范围内的对数约束条件,定义如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图78
其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图79为常数,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图80为所有的单位向量,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图81为自然对数。
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图82用于调整轨迹的激进程度,定义如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图83
其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图84为最大速度,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图85为最大加速度。

Results/实验结果

应用详情

四旋翼硬件平台同论文”Teach-repeat-replan”,目标是头戴二维码移动的人.用三个有限视场角的相机检测目标.因为视场角较宽,无需使四旋翼面向目标,而是朝向轨迹切线方向,以保证能够观测到轨迹方向的障碍.系统总运行时间接近20ms,设置整个系统重规划频率为15HZ.
image.png

实验详情

我们在室内未知复杂环境中进行多次实验.室内环境如下图.在该环境中,目标平均速度1.3m/s,飞机能够安全平滑动态的跟踪目标
image.png
image.png
在室外环境的稠密树林中,试验场景如下图,目标运动速度更快,在该场景下,四旋翼最快速度达到3m/s
image.pngimage.png

基准比较

将目标运动预测和跟踪轨迹规划方法与其他方法相比.采用Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图91的仿真环境,随机生成140个障碍.将我们的工作与论文[10]”J. Chen, T. Liu, and S. Shen, “Tracking a moving target in cluttered environments using a quadrotor”做对比.为了比较,给目标的对地真实位置添加高斯噪声,参数为(Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图92),高斯噪声的标准差分别为Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图93用来表示环境中的低噪声,中噪声,高噪声.我们对比目标将来位置Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图94和预测轨迹Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图95的平均距离误差.平均误差计算公式如下:
Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图96
其中Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图97是当前时间,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图98是采样间隔,Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图99是采样步长,每个场景预测超过15000个点.对比结果如下表:
image.png
相比于论文[10]中的方法,本文的方法预测精度更高,计算时间0.5ms,比论文[10]的0.3ms有所增加.

此外,与论文[9]”B. Jeon, Y. Lee, and H. J. Kim, “Integrated motion planner for real- time aerial videography with a drone in a dense environment””对比跟踪轨迹规划方法.
将目标未来轨迹的对地真实位置作为两种规划方法的输入.
定义:
有效跟踪时间:无人机到达与目标距离在Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图101平面小于3m的时间。
跟踪频率Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图102:在整个跟踪的总时间中,有效跟踪时间的占比
设置目标不同的平均速度Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图103和最大速度Fast-Tracker: A Robust Aerial System for Tracking Agile Target in Cluttered Environments - 图104为不同的跟踪场景,在每个场景200个跟踪任务中,对比计算时间和跟踪频率,结果如下表:
image.png
相比于论文[9],我们的方法计算量更小,并且能更好更有效的跟踪目标。结果表明高频的重规划对目标突然加速度等不可预知场景有较好的反应。此外如下图,目标轨迹为黑色虚线,蓝色曲线及扇形是本文规划轨迹和加速度,红色为论文[9]的方法,实验表明我们的方法更节能
image.png

Future/未来研究方向

未来研究方向是在跟踪过程中避开动态障碍,并且将会提供物体检测方法,使系统能够应用与更多场景。