IEEE Geoscience and Remote Sensing Magazine - March 2022 - 113

testing, and the remainder for validation. Under these settings,
color multiobject tracker (Ctrack), with the recovery
of long-time disappearing objects in the crowded scenes,
achieves the best tracking performance among all methods
in the VisDrone testing data set in terms of the IDF1, MT,
ML, and FN.
For the UAVDT data set with 50 sequences recorded
in the traffic scenario from UAVs, 60% are for training,
and 40% are for testing. From Table 12, it can be seen
that SORT is superior on most metrics. Although these
results are far from the requirements of practical application,
they can provide feasible direction (e.g., the association
of moving objects) and a reliable theoretical basis
for future research. For the Stanford Drone data set, the
performance gap of the listed baseline method is minimal;
maybe IOUT is slightly better.
ESTIMATION OF COMPUTATION COST
In this survey, all of the reviewed methods have their own
experimental environment, experimental data, and even
source code. Considering that the computation cost is directly
related to the speed, GPU, and backbone model,
we list these three indexes of the UAV-exclusive methods
for these three topics in Tables 13-15. Depending on the
computing power of Nvidia's GPU [224], the computation
cost can be estimated with the backbone network in
the corresponding method.
DISCUSSION AND CONCLUSION
In this article, DL approaches in object and tracking of the
RS field were systematically analyzed according to three
UAV topics, i.e., SOD, VID, and MOT. The conclusions were
drawn as follows.
TABLE 10. PERFORMANCE COMPARISONS OF UAV-EXCLUSIVE AND CLASSIC FOR THE VISDRONE-MOT TESTING
SET, TAKING APS AS THE EVALUATION INDEXES.
METHOD
PAS
Tracker [161]
HMTT [153]
DAN [77]
GGD [155]
Cas_RCNN +
FPNCas_RCNN +
FPN [80]
FRAMEWORK
CenterNet +
IoU
CenterNet +
IoU
RetinaNet +
DAN
FRCNN
Cascade
R-CNN
TRAIN/
TEST
(SEQ)
56/7
56/7
-
IMAGE SIZE AP
608 × 608
608 × 608
50.8
AP_0.25 AP_0.5 AP_0.75
66.1
52.5
28.67 39.05
1,500 × 1,000 13.88 23.19
56/33 3,840 × 2,160 23.09 31.01
-
2,000 × 1,500 28.51 44.76
The best performers are highlighted in bold. PED: pedestrian.
33.8
27.88 19.08
12.81 5.64
22.7
15.55
30.38 10.4
AP_
CAR
AP_
BUS
AP_
TRUCK
AP_
PED
AP_
VAN
EXPERIMENT
DATA
62.7 81.2 43.9 30.3 35.9 VisDroneMOT
44.35
30.56 18.75 26.49 23.19 VisDroneMOT
32.2
8.83 6.61
35.45 28.57 11.9
18.61 3.16 VisDroneMOT
17.2
22.34 VisDroneMOT
35.09
34.58 18.2 -
26.18 VisDroneMOT
TABLE
11. PERFORMANCE COMPARISONS OF UAV-EXCLUSIVE AND CLASSIC FOR THE STANFORD DRONED DATA SET,
TAKING MOTA, MOTP, AND SO ON AS THE EVALUATION INDEXES.
METHOD
IPGAT [159]
CEM [170]
GOG [166]
IOUT [168]
SMOT [167]
SORT [169]
SLSTM [195]
FRAMEWORK
LSTM + CGAN +
Siamese
FRCNN
FRCNN
FRCNN
FRCNN
FRCNN
FRCNN
TRAIN/
TEST
(SEQ)
IMAGE SIZE MOTA MOTP IDF1 IDP
36/24 1,417 × 2,019 3
MT
(%)
ML
(%) FP
36/24 1,080 × 540 99.9 99.9 90 90 99.8 0.13 3
FN
833
IDS FM
EXPERIMENT
DATA
3,395 905 Stanford
Drone
81.8 5.4 47.6 2.7 90.25 972,646 348,495 3,103 5,997 Stanford
Drone
36/24 1,417 × 2,019 98.9 100 86.3 86.7 100 96.2 3
36/24 1,417 × 2,019 99.9 100 93.2 93.2 98.9 1.04 0
66,625 4,928 2,621 Stanford
Drone
2,497
1,170 949 Stanford
Drone
36/24 1,417 × 2,019 99.1 100 91.8 91.9 97.3 1.38 17,212 38,846 2,275 3,926 Stanford
Drone
36/24 1,417 × 2,019 99.5 98.1 95.7 96 98 1.05 20
36/24 1,417 × 2,019 99.3 99.9 89.6 89.6 99.8 0.13 11
32,436 957 952 Stanford
Drone
841
The best performers are highlighted in bold. CGAN: conditional generative adversarial networks; SLSTM: social long/short term memory networks.
MARCH 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
113
3,630 906 Stanford
Drone
IEEE Geoscience and Remote Sensing Magazine - March 2022

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