粉色：重点算法 紫色：生癖词汇 绿色：引文&未补充公式
An Autonomous Vision-Based Target Tracking System for Rotorcraft Unmanned Aerial Vehicles
摘要
In this paper, an autonomous vision-based track- ing system is presented to track a maneuvering target for a rotorcraft unmanned aerial vehicle (UAV) with an onboard gimbal camera.
To handle target occlusions or loss for real-time tracking, a robust and computationally efficient visual tracking scheme is considered using the Kernelized Correlation Filter (KCF) tracker and the redetection algorithm. The states of the target are estimated from the visual information.
Moreover, feedback control laws of the gimbal and the UAV using the estimated states are proposed for the UAV to track the moving target autonomously.
The algorithms are implemented on an onboard TK1 computer, and extensive outdoor flight experiments have been performed.
Experimental results show that the proposed computationally efficient visual tracking scenario can stably track a maneuvering target and is robust to target occlusions and loss.
本文提出了一种基于视觉的自动跟踪系统，用于跟踪带有机载云台摄像机的旋翼飞行器(UAV)的机动目标。
摘要针对实时跟踪中目标遮挡或丢失的问题，提出了一种基于核化相关滤波(KCF)跟踪和重检测算法的鲁棒性和计算效率高的视觉跟踪方案。目标的状态是通过视觉信息来估计的。
提出了基于估计状态的框架和无人机和云台的反馈控制律，用于无人机自主跟踪运动目标。
这些算法是在TK1舰载计算机上实现的，并进行了大量的室外飞行实验。
实验结果表明，提出的计算效率高的视觉跟踪方案能够稳定跟踪机动目标，对目标遮挡和目标丢失具有鲁棒性。
SYSTEM CONFIGURATION
A DJI Matrice100 is used as the UAV platform, which is equipped with an onboard TK1 computer and a monocular RGB gimbal camera.
An overview of the system configuration is shown in Fig. 2.
The gimbal camera mounted on the UAV platform provides the video stream and internal angles for the onboard computer.
The visual tracking algorithm obtains position of the target on the image plane,which is feedback to the gimbal controller.
In addition,the states of the target are estimated by fusing the inertia measurement unit (IMU) data of the UAV platform and the gimbal.
A switching tracking strategy is performed based on the estimated states.
The high-level controller computes the desired velocities of the UAV,
and the low-level controller controls the attitude correspondingly.
The frequencies of the video stream and the control signal are 30Hz and 10Hz, respectively.
采用DJI矩阵100作为无人机平台，配备车载TK1计算机和单眼RGB万向架摄像机。
系统配置概述如图2所示。
安装在无人机平台上的万向节摄像机为机载计算机提供 视频流和 内角。
视觉跟踪算法获取目标在图像平面上的位置，该位置反馈给框架控制器。
利用无人机平台和云台的融合惯性测量单元(IMU)数据，对目标状态进行估计。
在此基础上，提出了一种基于估计状态的跟踪切换策略。
高阶控制器计算无人机所需速度，
低阶控制器相应地控制姿态。
视频流和控制信号的频率分别为30Hz和10Hz。
A VISUAL TRACKING SCHEME
In this section, a computationally efficient visual tracking scheme robust to target occlusions and loss is presented, which consists of the KCF tracker, the target loss detection and the redetection methods.
Fig. 3 shows the configuration of the visual tracking scheme.
Firstly, the KCF tracker estimates the state of the target.
The status of the target, i.e.
１loss or not, is then detected based on the regression function of the KCF tracker.
２Finally, a redetection method is presented to track the target when it come out off ccullsions.
提出了一种对目标遮挡和丢失具有鲁棒性的高效计算视觉跟踪方案，该方案由KCF跟踪器、目标丢失检测和重检测方法组成。
首先，KCF跟踪器估计目标的状态。
目标的状态，即
１然后根据KCF跟踪器的回归函数检测是否丢失。
２最后当目标离遮挡之后，一个重检测方法用来跟踪目标。
A. KCF Tracker
KCF develops from Circulant Structure of Tracking-by-detection with Kernels (CSK) [17], applying online learning methods to solve tracking problems.
More precisely, it is a machine learning method without any prior knowledge.
At the first frame, the object of interested (OOI) region is chosen manually and the KCF tracker transforms the region into a multi－channel HOG feature descriptor.
A regression function f (z) of OOI region z is initialized by Ridge Regression with HOG descriptor.
For the new frame, f (z) is evaluated on several regions around the last region of OOI.
Finally, the region which has max response of evaluation is considered as the output and applied to update f (z).
To accelerate the matrix computation of Ridge Regression, KCF transforms each channel of HOG feature descriptor into a circulant matrix by cyclic shifting.
It is known that circulant matrix can be made diagonal by Discrete Fourier Transform(DFT) [18].
KCF是由核的基于检测的跟踪循环结构(CSK)发展而来，采用在线学习方法解决跟踪问题。
更准确地说，它是一种没有任何先验知识的机器学习方法。
在第一帧中，手动选择感兴趣的对象(OOI)区域，KCF跟踪器将该区域转换为多通道HOG特征描述符。
利用HOG描述符进行岭回归，初始化OOI区域z的回归函数f (z)。
对于新一帧，f (z)是在OOI的最后一个区域附近的几个区域上求值的。
最后，将评价响应最大的区域作为输出，应用于更新f (z)。
为了加快山脊回归矩阵的计算，KCF通过循环移位将HOG特征描述符的每个通道的描述符转换为一个循环矩阵。
通过离散傅里叶变换(DFT)[18]可以将循环矩阵的对角化。
Thus matrix computation, especially matrix inversion, can be efficiently processed in fourier domain.
Furthermore, a kernel function, which maps the regression function f (z) into non-linear space, is applied in the KCF tracker to promote the performance of tracking.
These solutions are introduced by CSK, and optimized in KCF.
In this way, the process speed and mean precision of KCF have reached 172FPS and 73.2% respectively.
More details of the KCF tracking algorithm may refer to [7].
因此，可以在傅立叶域中有效地处理矩阵计算，特别是矩阵求逆矩阵。
此外，在KCF跟踪器中还应用了一个核函数以提高跟踪性能，将回归函数f (z)映射到非线性空间。
这些解决方案由CSK引入，并在KCF中进行了优化。
这样，KCF的处理速度和平均精度分别达到了172FPS和73.2%。
KCF跟踪算法的更多细节可以参考[7]。
B. Target Loss Detection and Redetection
Various visual trackers have been proposed to tackle illumination variation, scale variation and occlusion problems [19].
However, most of current algorithms can not detect the target in the presence of full occlusions, which are often encountered in tracking mission of UAVs.
In this paper, a simple and efficient redetection method is proposed so that the target can be estimated when it appears again.
针对照明变化、尺度变化和遮挡问题，提出了多种视觉跟踪器。
然而，目前的算法大多不能在完全遮挡的情况下检测目标，这是无人机跟踪任务中经常遇到的问题。
本文提出了一种简单有效的重检测方法，使目标在再次出现时能够被估计。
The status of the target, i.e. loss or not, is estimated by the target loss detection based on the KCF tracker.
The max response of regression function f max (z), denotes the relevance between the OOI region and the target.
When the value of f max (z) is less than a threshold, it implies that the target may be lost.
The redetection works when f max (z) is less than a threshold.
According to flight experiments, the values of f max (z) vary in the range of (0, 0.5) in outdoor environments.
The threshold is experimentally set as 0.17.
目标的状态，即是否丢失，由基于KCF跟踪器的目标丢失检测来估计。
回归函数f max (z)的最大响应，表示OOI区域与目标之间的相关性。
当f max (z)的值小于阈值时，说明目标可能会丢失。
当f max (z)小于阈值时，可以重新检测。
飞行实验表明，f max (z)值在(0,0.5)的范围内变化，室外实验。
实验设置阈值为0.17。
The UAV hovers and starts to search the target when the target loss is detected.
Some classical algorithms [14] scan all pixels in the new frame to search the target, and the computational complexity is high.
In general, the target is moving when it re-appears in the view of the camera.
Hence, the target can be estimated by detecting the moving foreground, instead of searching all pixels in the new frame
当检测到目标丢失时，无人机悬停并开始搜索目标。
一些经典算法[14]扫描新帧中的所有像素来搜索目标，计算复杂度较高。
一般来说，当目标重新出现在相机的视图中时，它正在移动。
因此，可以通过检测运动前景来估计目标，而不是搜索新帧中的所有像素
In the paper, a moving object detector based on the Frame-Difference (FD) method [20] is applied, as shown in Fig.4.
In specific, the detector subtracts the last frame from the current frame to obtain the difference image.
Although the FD method is computationally efficient, it is sensitive to noise, as shown in Fig. 4(b).
The Gaussian blur is hence applied to remove the noise in the difference image.
Consequently, the detector constructs the bounding boxes based on the center of the foreground, as shown in Fig. 4(d).
It is noted that the size of these boxes is the same as the initial OOI region.
Finally, the regions contained by these bounding boxes are evaluated by the regression function.
The region,which has the maximum value of f max (z) and is greater than the threshold, is selected as the position of the target in the new frame.
本文采用基于帧差(FD)方法[20]的运动目标检测器，如图4所示。
具体来说，检测器从当前帧中减去最后一帧，得到差分图像。
FD方法虽然计算效率高，但对噪声敏感，如图4(b)所示。
因此，高斯模糊被用于消除差分图像中的噪声。
因此，检测器根据前景的中心构造边界框，如图4(d)所示。
需要注意的是，这些框的大小与初始OOI区域相同。
最后，用回归函数计算这些包围框所包含的区域。
选择f max (z)最大值的区域，并且f max (z)大于阈值的区域，就作为目标在新帧中的位置。
IV. GROUND MANEUVERING TARGET STATE ESTIMATION
地面机动目标状态估计