pedestrian data mining with object tracking and trajectory
TRANSCRIPT
� 41�� 1� ��������� Vol.41, No.1
2021 � 1 � Systems Engineering — Theory & Practice Jan., 2021
doi: 10.12011/SETP2019-2960 ������: U491 �����: A
������������������ �
� �, ���, ���, � �
( ��� � ����, �� 410073)
� � �������������������������� !"#$,��%&'�()Æ*+� !,-"$. ./'01#$234%&5'6%&��, 7()'0%&8%9�*:#9+;<=>, ?@AB,C8DE%9FG-, H)�I9F=>J�KL%=>MN. -OPQ./R.S�%/R.0TU1)2VW,X, Y3,X�3Z8%9�[4*:#9\�, 5�6.7])9�%*:G-#9+;. ^)9�*:%_8, �./8`9abU1%.7], \,5cde./*:fg9%*:ab:1. 5;<h, =:1�>Ji%\�9�*:,5j?�@G*:��8\�,*:G-.
��� R.S�; /R.0T; *:ab; *:G-; �I9F
Pedestrian data mining with object tracking and trajectoryclustering
SAI Bin, CAO Ziqiang, TAN Yuejin, LU Xin
(College of Systems Engineering, National University of Defense Technology, Changsha 410073, China)
Abstract With the Internet of Things, big data, artificial intelligence making breakthroughs in the field of
security, public video monitoring systems have developed quickly in recent years. The equipment generates
massive amount of unstructured data, through analysis and research on pedestrian trajectory of video
data, it can be found that the hidden behavior patterns contained which have an important research value.
The article uses the multiple object tracking algorithm based on object detection to extract and describe
the pedestrian movement trajectory in the surveillance video of subway station and mall exits, and then
analyzed the trajectory pattern of pedestrians on the basis of trajectory. Aiming at the characteristics
of pedestrian trajectory, a trajectory clustering method based on trajectory similarity was designed and
implemented on the basis of point density clustering algorithm. The results showed that the method can
effectively extract pedestrian trajectories, and extract trajectory patterns from large types of trajectory
data.
Keywords object detection; multiple object tracking; trajectory clustering; trajectory pattern; crowd
behavior
1 AB
��������������, ����������, ����, �����
C�ÆD: 2019-12-31
EF��: (1991–), �, �, ��, �Æ�: �� �� �; ��� (1996–), �, ��, �Æ�: �� �� �; ��� (1958–), �, �, ���, , �Æ�: ���������������; ����: ��(1984–), �, �, , �Æ�: �� �����������.
��GH: �������� (71771213, 91846301, 71790615, 71774168); ��������� (2017RS3040, 2018JJ1034)Foundation item: National Natural Science Foundation of China (71771213, 91846301, 71790615, 71774168); Science and
Technology Program of Hunan (2017RS3040, 2018JJ1034)
IJKL�M: , ���, ���, �. ����������������� ��� [J]. ���������,
2021, 41(1): 231–239.
NJKL�M: Sai B, Cao Z Q, Tan Y J, et al. Pedestrian data mining with object tracking and trajectory clustering[J].
Systems Engineering — Theory & Practice, 2021, 41(1): 231–239.
232 � � � � � � � � � � 41�
�. ����������������������, �����������������. �
������� �� [1] (emergency events database, EM-DAT) ��!, � 1900 � 2019 10 � 31
!�!Æ �""# 10 �#��$�� 18 $, "# 3247 �. %" �#$���%: 2014 12 �
31 &, &#!%&!'""�"��$���� 36 �"#; 2015 9 � 24 &, ''# $Æ!�"�$
��, �( 2236 �"#. �%�", ��������)���#(, ��$ &*������%+
#$( [2].
# �, ��),��, )&��, ),$����, ���'-),'("!%�(&)**� �,
&*%"++�,-���+,-. ,�*-.', %./.�/)..�()/����0/0*0*
*�"���+, 10+2� [3] 11%"�2-0/. 10+2� (multiple object tracking)� (121�*�"�130+3���2,�2�. ��/)..���, YOLO[4] +-.� “one-stage” �
Faster R-CNN[5] +-.� “two-stage” �0+�/0/, �0/4�0+�/�4)�+). ��&�)3
�!33#%�,5�.!, .�0+�/�0+2�0/ (tracking-by-detection) &10+2�-4".1+ /.
04, 5�%./,���*�"���+�010*. 2�, 6567�.352 [6] ��0�12
66�0+�/�0+2��0+#4+2��),����. 87�.78 [7] ��/��*2-93
��"5�6��/�0+2�0/. 6%./ Nikouei[8] ��/�12-9%: CNN 0/, 6478:
;2�85),��", 7%604)4, 56989�:7. 81&;�<;"5, =:9*, <:�<;
.'), #;��<=6&�!3>;�<=.
,���+#(<=, �>./�81?-�� [9,10] #(����9.�+, %" Henderson[11−14]
6?��9.�/��9.���85��#�@�#(<", &:��%Æ@�,%Æ@�4/>, 8
A??��Æ;@A - =B$A4)#>��B;. Helbing[15] C�1/��9.��, &���1D:�
�4/>, 8C��4=B$A�?��9.���B;�B����. E� Helbing, Johansson[16,17] �
,�/��9.����C/�DF�;B;3��+�<=, /�1��9��, 6�����9#+%C�9�D?9�E=9, 86&�$E<;, @ F!��<= [18]. Moussaıd[19] �(8�+F�E��
G<%"G&��@, G��&��A&��+��#(, BH!I1���*=,- [20,21], �>C4
H, 8C/�%9��B;. Porter[22] ��H�?3���'7, /�1 IM ��(), G���I5�
1��9����+F�E���D?@.��, �J?���:I, ��>*160%. ?4�9.
����>��+�%I5�B;HK, 81�3��B; �A�, H �LJ0E (M?�@+ .
�N���*�"OK1�%�PQ �, +CE3��+�E�G</R1,-� �.F, B20
;, ����47%�J�+�)CPQ ����4, �"�G<�AG&!%��PQ"�PQ'B
��@, ���+��, CH�+�/�$ &*�%�$(DE, 8����&F�)%�D3��'
7, ��C3PQ�G<%���A�.�4 [23] 1G<PQ�E�2-H80/,G%&�+�EG<��
", �(�PQ �KL��PQ�4 [24], S67%�J�+�)CPQ%#523PQI5, D8PQ�
4�&K, �J#G<'3!4���+�E.
&�!%.�/)..().F#�0/��..�0*HE, KMI8!3>��)3�5� YOLO
v3[25] 0+�/0/'+��PQ/J�.F, B&%.F#!I Deep SORT[26] 10+2�/, EF1
YOLO v3 C/LM#GJ�<=, LK,'&<;"5>������/�2�, LJ5*�"���
-2J�(+,-, �������PQ,-��>2M�#(. &�40/�IL#, ,���PQ�
'7, IJ DBSCAN �40/ [27] B��C�, �PQ ���N�ONPQMO���4&K�+3
�<=, B&%.F#�!1.�PQ�J)�PQ�40/, ()&)�N 1OP. �@.P, G0/�
���PQ ����4, B�"/J�P�PQ�E, &%.F#, ��2M#(����+'B, ��
�!������+���CH�+���T,.
� 1� , �: ����������������� ��� 233
DBSCAN
O 1 �PQR)*S+,TU-.VW/XMY012O
2 Z34[
2.1 \567YOLO[4,25,28] 1 2015 �7QU�.� Redmon � Farhadi /�, %*Q+ “You Only Look Once”,
��2�H�C�:*, 8Q� YOLO v3 0/� 2018 /�.G�H0/60+�/<=�V+WX<=,
%"RI�134(��>4,Y>�WX. YOLO N(8123R�8:�2SSJ, K��>Z�R
INS��4(�4,Y>��/. 11G+*/�N�V1238:,OT�OL��T�T�:*,
��(U YOLO &6�4)0V �9:I, B),'LW�4��/TM!3> (mAP).
YOLO0/L6RINS$Q�P+93�9�NU. 6$PE�NU'+N[8:�RI, 6�N[
8:,U5�/&K�4O(�(+UP�P,),8E(8(\�QRS�4O(��VI. YOLO v3-416
& COCO[29] �#,'& 29 ms 5V1�/12, B)!3>,'?W& 55.3, �)3�!337%�+PW�:I.
2.2 9]\567:;\5<^=>��/)..&0+�/-4�"]�8, .� Tracking-by-Detection0E�10+2�0/���1
+ /. SORT[3] 0/1& 2016 � Bewley, Ge ��/�, %*Q+ simple online and realtime tracking,
&10+2���"6PBAWX0/�TUD0/&^52$, %8Q4)Y51 260 Hz. 2018 .�
SORT�C�0/ Deep SORTZ/�,&10+2���",&^0+6��;'B,-/J,B(8 CNN
8:&����� ���VR/J'B, WX10/�S[3,/Y12�HK, . 1+ SORT � Deep
SORT &$0 � MOT16 #�&K. Deep SORT ��� SORT T�1�X 45% � ID switches .',
&4J>*�"Y5I5�2�HK.
? 1 _ MOT16 S+@,A`BC��X� MOTA MOTP MT ML ID Runtime
SORT 59.8 79.1 25.4 22.7 781 40 Hz
Deep SORT 61.4 79.1 32.8 18.2 1432 60 Hz
�� YOLO v3 1.�CJNS��0+�/, ��#>J�LM,-, B�,Y3�/0+��2�.
G%, KM&.� YOLO v3 �0+�/0/�.F#, !I Deep SORT 2�Z.
Deep SORT 0/ (#+PBAWX2��TUDLMU3\#. 0+�/0/���0+���/
HE, 60+UP��9,-ZI52�Z, PBAWX%D8�/,-�0+��2��/, U5���
�>2J�6�,-. %), �U52�Z��/&K�0+�/0/��/&K, �KB)R�U4,-,
Y���$<�/, V;0/,H>. G%��V0�/UP��/UPHW�[[CWXY, ��/(
��/(HW�LM�J), D8TUDLM0/���/LM, R�8Z�2�&K. YOLO-Deep SORT
0/���/2��.K\\�N 2 OP.
YOLO �/0/��/&K�!I Deep SORT �2�&K�N 3 OP, #,ZJ] 100, 10005 �
30017 J��� . �[5, &!I Deep SORT 10+2�HE, ,'�13��0+��W[2�BHOKPQ.
234 � � � � � � � � � � 41�
… …
1. 2. 3. 4. 5. 7.6.
O 2 YOLO HDEF,�GaH
b
e
a
f
dc
O 3 YOLO v3 I YOLO-Deep SORT BCJK
3 Lb Hausdroff MNOPQRS
PQ�4�0�&��!G!4�B4,��F��@�'B, ��G<���]PQ�E. &��P
Q�4���", �4��^+��PQ, +EFPQW�J)�V0, KM^8 HausdroffXY�PQ>�
�J)��V0,/�.� HausdroffXY� DBSCAN�40/. DBSCAN12-.�7])��40/,
G0/7%MO ��^_�'3B,%#� �#>"4])!4, _8�PQ�4, `_\��PQ
]�!4�!�I, ��PQ�4.
3.1 TUVc*�"/J5���PQ12`_])W^HDH�(+^H, (+^H�3 �.PPQ�`),
��J,-V0�/JPQ�)W,-. 3EPQ^ [30] +: T = {T1, T2, · · · , Tp}, %" p +PQ�3
. 0+ i �F�PQ.P+: Ti = {(x1, y1, t1) , (x2, y2, t2) , · · · , (xJ , yJ , tJ)}, %" x, y +0+&*�CJ
a;"�Sa(+,-, ti -.J , S)W,-, J +PQ i �`).
3.2 deWWX����&�#)%����D3, G%-3��("�PQ �`)?��B. .���PQ`)�
2��'7, b`^8 HausdroffXY'+PQ�J)�2-)%. HausdroffXY1�U37HW])
�2-)%, 17HWXY�2-.Y_E, ��1DU`PQ��V0H��(\a`)aC, %3
E+:
H (A, B) = max [h(A, B), h(B, A)] , (1)
%", h(A, B) = maxa∈A
minb∈B
‖a − b‖, h(B, A) = maxa∈A
minb∈B
‖b − a‖, � Hausdroff XY,'�!PQHW�]
�), U`PQW� Hausdroff XY�9, -.PQ�J)�4.
3.3 f>Vc,�PQ�4�.', b`� DBSCAN 0/"��>Yb��b>, =IbcPQ�4OPQ�M
OPQ�Yb, cYPQ�4�(, >;c��>3E.
� 1� , �: ����������������� ��� 235
a) MinTys: d31d+bcPQ`dPQ� %ZQ.
b) bcPQ: �Ke2PQ&23XYed�-45, %G]PQ %�� MinTys, YGPQ+bc
PQ. PQ�-4�XY�)%0/�a3�XY@ F3.
c) 4OPQ: 4OPQ�e�bcPQ, 81U�bcPQ�-4.
d) MOPQ: &PQ^", f�e�bcPQb�e�4OPQ�PQ.
.�HausdorffXY�DBSCAN0/,bN�(33U3@ : G4c^ ε�bcPQ89ZQMinTys.
0/f-f�0/ 1 OP.
Algorithm 1 PQ�40/.gh:
��dgh TS = {TR1, TR2, · · · , TRn};_ce` ε i[g\d MinTys;
g]:
����ajgh O = {C1, C2, · · · , Cm};/* � 1 bd: �� Hausdorff cd�X
1: for each TRi, TRj ∈ TS & i �= j do
2: compute Haus(TRi, TRj);
3: if Haus(TRi, TRi) ≥ ε then
4: Nε(TRi)← TRj
5: /* gf TRj �ecgh */
6: end if
7: end for
/* � 2 bd: ������ */;
8: Set clusterID = 0;
9: Mark all trajectory as unclassified;
10: for each TRi ∈ TS & TRi is unclassified do
11: if |Nε(TRi)| > MinTys then
12: for each TRj ∈ Nε(TRi) do
13: if TRj is unclassified then
14: Assign clusterID to ∀TR ∈ Nε(TRi)
15: Add Nε(TRi)−TRi to Queue Q
16: while Q �= ∅ do
17: for each Qi ∈ Q do
18: if |Nε(Qi)| ≥MinTys then
19: for each X ∈ Nε(Qi) do
20: if X is noise then
21: Assign clusterID to X
22: else
23: Add X to QUEUE Q
24: end if
25: end for
26: else
27: Remove Qi from Queue Q
28: clusterID + +
29: end if
30: end for
31: end while
32: else
33: Make TRi as noise
34: end if
35: end for
36: end if
37: end for
236 � � � � � � � � � � 41�
4 ij^_`ak4.1 lmnb
KS�@(8 Sony A7 R III *aZDC f�k, ^893kUhi�0E�U3�h�If�23
+c\f, j3"5�� 20 #e�*�hi, hi)W+ 2019 9 � 3 &. �N 4, &KS�@�"5",
*����'�P&XY�;%23XY�UP,�3Lhi5���0+��i9,ON YOLO��9
0+�/�^_��j. ^5#g>+ 1920*1080, J4> 25, f/ 50M �*�. &PQ/J��4�@
lg, fÆlh+: Windows7 h'��, Inter Core-i7 N�Z (2.80 GHz), 56 16 GB, NVIDIA 940MX(2
GB) WP, Python3.6 '+ighi.
a b c
O 4 opde
4.2 qrTUstD8.� YOLO v3 � Deep SORT 0/�^5�*�����PQ/J, 0/6*�"0+F��
(+��OK, U52�H�(+7, B6#e�B230+�(+jki&2$1+2`PQ. %""5
2^5 3009 `��PQ, "5A^5 1470 `, "5j+ 1632 `. %""52^5 3009 `��P
Q, "5A^5 1470 `, "5j+ 1632 `, O%PQM^8 2.1 g"�B;0/.P. "52"��h
fG#l", H�m��, ��/%�1, (��If%j#j (�hk�f), k\ (\f), j4 (\f)
�l\ (l"mf). "5A�"5jG#+c\f, n#l"��If��, G%���"52, ��/%�9. "5ARIf#j (\f), j4 (\f) �f>j (\f) j3�If, "5jRIf#j (�hk�
f), j#j (\f) �l\ (\f) j3�If. &/J��m#$<:.'%, M,1T2���BOKPQ �. j3"5�/J���&K�N 5 OP.
/J5PQ �E, b`�PQ�$7�Z7��1])N#(, N 6 �P1j3"5�$7�Z7#
>, %"fn+$7#>N, jn+Z7#>N, �N"�[�$n7�Zl7 (#>&*�a;�k`. Deep SORT &0+2���", ��<:��0+oo, <&mS�/5G0+E, 0/lp+1Q�
!���0+, G%�&a;"("$n7�Zl7. +3L/J�PQ �,'V>�.Pe2��&G
!4V>�F�PQ, KM�PQ ���1VI, qmp$n7�Zl7�&*�k4("BHOKJ
9� 100 �PQ, 8ZU5 504, 222, 400 `V>PQ.
(a) (b)
(c) (d)
(e) (f)
O 5 YOLO-Deep SORT ugJTU-.hTVW
(a)
(d)
(b)
(e)
(c)
(f)
O 6 viIwijklmO
� 1� , �: ����������������� ��� 237
4.3 TUxnop�� Deep SORT 0//J�1*�"��0+2�(�"c7(+, &�N�2���", ��0+
2�(�H���_m�"qr9*, ��"c7("rn, �o^���#)�?�.'. N 7 "�V>
PQ%I1X�, G%�(�PQ ���TpN�, /4PQ�J)�V0��4&K!33.
&�N"5"���F�4)L#�o4, BH^5� �+)W^Hn^, OKM^8p�TM
Tp=8�^5�PQ ���TpN�, &�PQTpN����", k�,(N�E�PQ �Ll
pn*�"���,-.
��p�TMTp6G45� �7��TMN�, B-o"c7Q, pm^H"�s�, qqPQ^
H�tI. O t(i) +pnPQ �, Y��p�TMTpE�PQ+:
Ts(i) =1n
(t(i + N) + t(i + (N − 1)) + · · · + t(i)), (2)
%" N +p�)Wsf`).��hi)IL 25J4>��hi, O&p�TMTp��", m N = 25,
�PQ ���TpN�. N 7 +j3"5�TpN�HK� N, fn+V*PQ, jn+TpN�E�
HK. ?3"5"�PQpm1X�, BHu%("�Z)�vF, ,'!3qq���F�tI.
(a) (b)
(e)
(c) (d)
(f)
O 7 -.VWyqrszCJK4.4 tuvw
+1 ��B@ ��4&K�r%, ��V0�4PQI"?PQW�T0la�SJ�4&K [31]:
QMeasure =numclus∑
i=1
(1
2 |Ci|∑x∈Ci
∑y∈Ci
dist(x, y)2)
. (3)
%", QMeasure Q�9.PPQI5PQ�J)�4, �4r%�5. dist(x, y) +PQI"1DU`PQ
� Hausdorff XY. &&^�4SJ$E����@�.F#, ��1Sq@�:*, 8Z33"52"eps=1750, MinTys=5; "5A eps=1650, MinTys=6; "5j eps=1850, MinTys=5. �N 8 OP, +:
*����**&K, %" x s-. eps JQ, y s+ MinTys, z s+ QMeasureQ, QMeasure M,&#;
@ Q.'>JU89Q.
(a) (c)(b)
O 8 -.xyzS{{
238 � � � � � � � � � � 41�
N 9 +?"5"PQ ���4&K�PQI %�#>. %""52U5 6 3�4&K, "5A
5 3, "5j 8 3, MOPQ�9 #>+ 11.7%, 9.9%, 21.25%. G0/,'1T��N��E�PQ �
$/J����PQ�E, �^5��*����], /J5�PQ�E,'!3�na�G!4���PQ�E��+�@. "52+2 (N 10(a)), G0/!3�/J�G"5"�r3 (���t: |wPQ.POLof4u�>3!4�PQ, rwPQ+f>j�j#0�PQI, (RI�hf�����o#j>0u�>3!4�PQI, pw+�a;#0�I5�hf��PQI, xwPQ1a;j4
��hf\^���PQI, vw+a;#0���hk�f�a;jn�PQI.
(a)
(d)(c)
(b)
(e) (f)
O 9 xyBCIlm
(a) (b) (c)
O 10 -.xyzC
5 ^|}
KM�VB�!1�$ !4����PQ/J0/, BH8����+�EG<. G0/��!�
� (#+PQ/J, PQTp�PQ�4j3q�. &PQ/Jq�, & YOLO v30+�/0/�.F#
!I Deep SORT10+2�0/,%HEF1 YOLO v3C/LMG#J,-�<=,/�^��PQ�
%H0/; &PQN�q�, (�^5�pnPQ ���TpN���J)V0; PQ�4q�, C�
.�])��40/�PQ���4BG<�PQ�E. �@&K.P, G0/�%H����*�"�
��0+��PQ/J��4#(, 7%�X�S[3��Ft3. G0/�r��85$ &*��
�����%, �H/J�G<0/��$ &*$0 ���������+0*w7%$(DE.
}~~�
[1] EM-DAT. The international disaster database[EB/OL]. 2012. https://www.emdat.be/.
[2] ��. ���su ssvec� �twt [J]. ���Xu�v�Æ, 2018, 14: 56–62.
Lu X. Application and prospect of big data technology in emergency rescue field[J]. Communications of the CCF,
2018, 14: 56–62.
� 1� , �: ����������������� ��� 239
[3] Bewley A, Ge Z, Ott L, et al. Simple online and realtime tracking[C]// 2016 IEEE International Conference on
Image Processing (ICIP), 2016: 3464–3468.
[4] Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]// Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 779–788.
[5] Girshick R. Fast R-CNN[C]// Proceedings of the IEEE International Conference on Computer Vision, 2015:
1440–1448.
[6] tux, vw�. xyuwxzyyxvw�� [J]. ���y�{�y: A z, 2000, 5(12): 1024–1029.
Liu X D, Su G D. A visual intelligent outdoor surveillance system[J]. Journal of Image and Graphics: A, 2000,
5(12): 1024–1029.
[7] zz, {|�. xy{}���xz, ��Æ����| [J]. �Xu ��, 2007, 24(1): 199–202.
Wan Q, Wang Y N. Research and implementation of detecting and tracking multiple moving objects method[J].
Application Research of Computers, 2007, 24(1): 199–202.
[8] Nikouei S Y, Chen Y, Faughnan T R. Smart surveillance as an edge service for real-time human detection and
tracking[C]// 2018 IEEE/ACM Symposium on Edge Computing (SEC), 2018: 336–337.
[9] Barbosa H, Barthelemy M, Ghoshal G, et al. Human mobility: Models and applications[J]. Physics Reports,
2018, 734: 1–74.
[10] Yan X Y, Wang W X, Gao Z Y, et al. Universal model of individual and population mobility on diverse spatial
scales[J]. Nature Communications, 2017, 8(1): 1639.
[11] Henderson L. The statistics of crowd fluids[J]. Nature, 1971, 229(5284): 381–383.
[12] Henderson L, Lyons D. Sexual differences in human crowd motion[J]. Nature, 1972, 240(5380): 353–355.
[13] Henderson L F. On the fluid mechanics of human crowd motion[J]. Transportation Research, 1974, 8(6): 509–515.
[14] Henderson L, Jenkins D. Response of pedestrians to traffic challenge[J]. Transportation Research, 1974, 8(1):
71–74.
[15] Helbing D. A fluid dynamic model for the movement of pedestrians[J]. arXiv preprint cond-mat/9805213, 1998.
[16] Helbing D. Interrelations between stochastic equations for systems with pair interactions[J]. Physica A: Statistical
Mechanics and its Applications, 1992, 181(1–2): 29–52.
[17] Johansson A, Helbing D, Shukla P K. Specification of the social force pedestrian model by evolutionary adjust-
ment to video tracking data[J]. Advances in Complex Systems, 2007, 10(Supp2): 271–288.
[18] Frank G A, Dorso C O. Room evacuation in the presence of an obstacle[J]. Physica A: Statistical Mechanics and
its Applications, 2011, 390(11): 2135–2145.
[19] Moussaıd M, Helbing D, Theraulaz G. How simple rules determine pedestrian behavior and crowd disasters[J].
Proceedings of the National Academy of Sciences, 2011, 108(17): 6884–6888.
[20] Turner A, Penn A. Encoding natural movement as an agent-based system: An investigation into human pedestrian
behaviour in the built environment[J]. Environment and Planning B: Planning and Design, 2002, 29(4): 473–490.
[21] Batty M. Predicting where we walk[J]. Nature, 1997, 388(6637): 19–20.
[22] Porter E, Hamdar S H, Daamen W. Pedestrian dynamics at transit stations: An integrated pedestrian flow
modeling approach[J]. Transportmetrica A: Transport Science, 2018, 14(5–6): 468–483.
[23] |��, y{. � �: }{��s [M]. 2 |. Æz: u}�~}||, 2012.
[24] ~�{, �. }|����~~ ���w [J]. �~g���, 2015, 30(1): 47–58.
Ji G L, Zhao B. Research progress in pattern mining for big spatio-temporal trajectories[J]. Journal of Data
Acquisition and Processing, 2015, 30(1): 47–58.
[25] Redmon J, Farhadi A. Yolov3: An incremental improvement[J]. arXiv Preprint arXiv: 1804.02767, 2018.
[26] Wojke N, Bewley A, Paulus D. Simple online and realtime tracking with a deep association metric[C]// 2017
IEEE International Conference on Image Processing (ICIP), 2017: 3645–3649.
[27] Murca M C R, Hansman R J, Li L, et al. Flight trajectory data analytics for characterization of air traffic
flows: A comparative analysis of terminal area operations between New York, Hong Kong and Sao Paulo[J].
Transportation Research Part C: Emerging Technologies, 2018, 97: 324–347.
[28] Redmon J, Farhadi A. YOLO9000: Better, faster, stronger[C]// Proceedings of the IEEE Conference on Com-
puter Vision and Pattern Recognition, 2017: 6517–6525. https://doi.org/10.1109/CVPR.2017.690
[29] Chen X, Fang H, Lin T Y, et al. Microsoft coco captions: Data collection and evaluation server[J]. arXiv Preprint
arXiv: 1504.00325, 2015.
[30] }~. w�vw���~~��t���� [D]. Æz: Æz����, 2014.
Kang K. Motion pattern study and analysis from video monitoring trajectory[D]. Beijing: Beijing Jiaotong
University, 2014.
[31] Lee J G, Han J, Whang K Y. Trajectory clustering: A partition-and-group framework[C]// Proceedings of the
2007 ACM SIGMOD International Conference on Management of Data, 2007: 593–604.