IEEE Geoscience and Remote Sensing Magazine - March 2022 - 128

an Intel NUC8i7HVK, to achieve the UAV tracking tests. The
potential directions are also indicated in this work, guiding
further research into DCFs for UAV tracking.
The main contributions of this work are fourfold, formally
summarized as follows:
◗ Comprehensive review: This work offers a general framework
of DCF-based trackers and summarizes many
state-of-the-art DCF-based trackers according to their
innovations and contributions.
◗ Code library: This work integrates most of the publicly
available DCF-based trackers into one code library. In
addition, our experimental results are also organized for
convenient reference. The code library and experimental
evaluation results are located at https://github.com/
vision4robotics/DCFTracking4UAV.
◗ Experiment evaluation: This work provides exhaustive experiments
of DCF-based trackers on six authoritative
UAV benchmarks, i.e., UAV123, UAV123@10fps, UAV20L,
UAVDT, DTB70, and VisDrone2019-SOT, to demonstrate
their performance under complex scenes and their superiority
against other types of trackers. The code library
and experimental evaluation results are located at https://
github.com/vision4robotics/DCFTracking4UAV.
◗ Onboard test: This work further implements DCF-based
trackers onto a typical CPU-based onboard PC to achieve
real-flight UAV tracking tests, where their real-time capabilities
and robustness under harsh scenes are validated.
RELATED WORKS
OBJECT TRACKING METHODS
According to different representation schemes, object-tracking
methods can be generally divided into two types: generative
[34]-[36] and discriminative [37]-[39]. The following
sections detail both methods.
GENERATIVE METHOD
The main idea of the generative method is to learn a feature
template from the target area in the first frame and to find
the most similar matching area to the template appearance
in the search region of the subsequent frames as the tracking
result. As is known, most of the early object tracking
methods are generative. For instance, Lucas and Kanade
[40] proposed holistic templates, which are based on raw
intensity values. To cope with appearance changes, subspace-based
tracking methods emerged [34], [41], [42]. As
a well-known branch in the generative method, numerous
tracking approaches based on sparse representation have
also grabbed researchers' attention [35], [43]-[45].
Nevertheless, the weaknesses of generative methods are
obvious. The first is that a large number of training samples
requires abundant computational resources, which makes
it difficult to meet real-time requirements on board a UAV.
Then, traditional generative approaches have neglected the
background information, which may help ensure more robust
tracking. Third, the generative methods usually suppose that
128
the appearance of the object will not undergo large changes
in a period while appearance variation happens frequently
in UAV tracking scenes. Recently, the discriminative method
has become common within the visual tracking community.
DISCRIMINATIVE METHOD
Unlike the generative method, the core idea of the discriminative
method (also called the tracking-by-detection method)
is to train a classifier that can distinguish the tracking object
from the background.
Belonging to the discriminative method, support vector
machine (SVM)-based approaches are the first ones to
flourish in the visual tracking community [37], [46]-[48].
More concretely, Avidan [37] constructed pyramids from
the support vectors and used a coarse-to-fine approach in
the classification stage in consideration of large motion
issues. Bai and Tang [48] presented an online Laplacianranking
support vector tracker to robustly locate the object.
Ning et al. [46] proposed a simple yet effective dual linearstructured
SVM to boost tracking efficiency. STRUCK [47],
proposed by Hare et al., achieved great tracking results due
to its kernelized structured-output SVM, which is utilized
to provide adaptive tracking. Although the performances
of the trackers based on an SVM are promising, large-scale
training samples will consume much of the machine's
memory and computing time, greatly limiting their realtime
capabilities on board a UAV.
Another competitive branch in the discriminative method
is based on multiple-instance learning (MIL) [49]-[52].
In [49], Babenko et al. showed that, compared with traditional
supervised learning methods, MIL can enable a more
robust tracker with even fewer parameter tweaks. However,
a serious problem in MIL is the instability of the sample labels.
In other words, if there is a slight change in the training
set (which is frequent in UAV tracking situations), the
output sample labels can undergo drastic changes, ending
up with poor robustness.
It is worth mentioning that among the previous trackers,
Zhang et al. proposed CT [53], which employed a simple yet
efficient appearance method and used a Bayes classifier in
the compressed domain for tracking. The speed of CT is appreciable;
nevertheless, its robustness and accuracy gradually
fell behind the ranks.
With the development of the convolutional neural network
(CNN) in recent years, applying one for object tracking,
which we call a deep tracker, has become a research hot
spot [54]. Specifically, such methods usually train a CNN
specifically for object tracking using a large number of labeled
images offline.
While tracking the object, for one type of deep trackers,
the object's template and search region are input into the
network simultaneously, and the object's location and size
in the search region are predicted directly end to end, e.g., in
[38] and [55]-[60]. Held et al., to name a few [38], have used
a simple feedforward network structure that was amazingly
efficient but still made it difficult to achieve real-time
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2022
https://www.github.com/vision4robotics/DCFTracking4UAV https://www.github.com/vision4robotics/DCFTracking4UAV http://www.github.com/vision4robotics/DCFTracking4UAV http://www.github.com/vision4robotics/DCFTracking4UAV
IEEE Geoscience and Remote Sensing Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2022
