IEEE Geoscience and Remote Sensing Magazine - March 2022 - 150

for UAV123, UAV123@10fps, UAV20L, and VisDrone2019SOT,
whose original attributes are the same. Their attributes
of CM and VC are classified as the VC. The original
attributes of FM, LR, and IV are the same as in this work.
POC and FOC are classified as OCC. For DTB70, their
original attribute, fast CM, is classified as VC, and OCC is
classified as OCC in this work. For UAVDT, their CR, SO,
IV, and LO attributes correspond to VC, LR, IV, and OCC,
respectively. Table 5 lists the correspondence between the
original attributes in the six benchmarks and the new attributes,
as well as the serial number contribution of each
benchmark to each new attribute. Above the horizontal
line is the sequence number of each benchmark under
the new attribute, and below is the original attribute(s)
of each benchmark corresponding to the new attribute.
Based on this work, the tracking results of each tracker
under the new attributes were also drawn into 10 plots
containing success rate and precision plots, as depicted in
Figure 7(a)-(e). Note that the calculation method adopted
is used to average by sequence, that is, to count the sequences
containing specific attributes in all the sequences
in the six benchmarks, and the arithmetic average of the
results of sequences involved is used as the final result. Figure
7(a)-(e) shows that the performance of the trackers under
specific attributes is obviously worse than the overall
performance in each benchmark, and the challenges that
each tracker does well in are also different. Overall, the
ARCF-HC tracker [24] performs the best in FM and VC, AutoTrack
[22] is good at VC and IV, and ECO-HC [95] ranked
first in OCC issues.
Figure 8 presents the capability comparison of the three
best-performing trackers under different challenges, i.e.,
the ARCF-HC [24], AutoTrack [22], and ECO-HC trackers
[95]. As far as each attribute is concerned, the performance
TABLE 5. THE CORRESPONDENCE BETWEEN THE ORIGINAL
ATTRIBUTES (UNDER THE LINE) IN THE SIX BENCHMARKS
AND THE NEW ATTRIBUTES AS WELL AS THE SERIAL NUMBER
CONTRIBUTION OF EACH BENCHMARK TO EACH NEW
ATTRIBUTE.
ATTRIBUTES
BENCHMARK
UAV123
CM VC
70
UAV123@10fps
UAV20L
CM VC
109
VisDrone2019-SOT
CM VC
41
DTB70
UAVDT
Total sequences
FCM
30
CR
336
FM LR
X
X
91
X
23
SO
116
OCC
20
LO
237
POC FOC IV
17
X
28
IV
247
FM LR
28
20
POC FOC IV
79
55
CM VC
16
FM LR
28
FM LR
13
7
30
POC FOC IV
54
POC FOC IV
13
73
18
VC
70
FM LR
28
30
OCC
54
IV
73
of the three trackers under the two attributes, OCC and
FM, is worse than that of IV, LR, and VC, indicating that the
two major tracking scenarios of OCC and FM are currently
more challenging in UAV tracking. As far as the three trackers
are concerned, when comparing the success rates, the
three have a similar overall performance in VC; the ECO
tracker [95] has a higher success rate in IV, LR, and OCC;
and the ARCF-HC tracker [24] does better in OCC. The
ARCF-HC and AutoTrack trackers both do better in LR
compared to the ECO-HC tracker. In the precision comparison,
there is little difference among the three in IV,
while the overall precision of the ARCF-HC [24] and AutoTrack
trackers [22] is significantly higher than that of
the ECO-HC tracker [95]. The ARCF-HC, AutoTrack, and
ECO-HC trackers are more accurate under FM, LR, and
OCC, respectively.
AGAINST DEEP TRACKERS
To better demonstrate the superiority of the DCF-based
trackers using handcrafted features in the UAV tracking
scenes, various deep trackers are selected, including those
adopting a basic CF framework but using CNN features, i.e.,
the ASRCF [73], ECO [95], CFWCR [64], MCCT [89], MPCF
[65], DeepSTRCF [66], CoKCF [101], IBCCF [102], and
HCFT [67], and also the end-to-end CNN trackers, i.e., the
UDT+ [98], SiamFC [57], CFNet_conv2 [99], TADT [100],
UDT [98], and DSiam trackers [58].
Remark 17: All of the deep trackers used GPU acceleration
in the experiment, and all of the DCF-based trackers
using handcrafted features, e.g., the AutoTrack tracker, were
evaluated on a single CPU using one core only.
Remark 18: According to Table 4 and focusing on the
AUC and DP, the AutoTrack tracker [22], which achieved
the most top-three scores and the most first-rank scores in
six benchmarks, is considered the most outstanding DCFbased
tracker. This section chooses AutoTrack to demonstrate
the DCF-based trackers' superiority against the deep
trackers in UAV tracking.
Table 6 lists their DP and tracking speed under benchmark
UAVDT [20]. As is demonstrated in their tracking
performance, even with low-cost handcrafted features,
AutoTrack tracker [22] still prevailed over most of the deep
trackers, achieving both excellent precision and considerable
tracking speed.
According to the experimental results, even compared
with the deep trackers, the handcrafted DCF-based trackers
still maintain their strong competitiveness in accuracy and
possess real-time performance, which further confirms that
the DCF-based trackers using handcrafted features, e.g., the
AutoTrack tracker [22] and so forth, are the best choice for
UAV-based aerial tracking.
FAILURE CASES AND CHALLENGES
Although DCF-based methods have exhibited their superiority
in UAV tracking, there still exist tracking challenges
that have not been well addressed. This section analyzes
150
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2022
IEEE Geoscience and Remote Sensing Magazine - March 2022

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