1. A
/ /
I L
The Internet
IoT
.53 No.4 1349–1359 (Apr. 2012)
口アイコンの対として,階段から上
コンに相当する.このような入口と
する.対となるアイコンが水平視野
,画面の上部の端に矢印を表示し,
けると対となるアイコンが存在する
る(図 5).Google Street View と
真を撮影した緯度・経度は厳密に
のつながりを示すアイコンを提示す
は階層間をつなぐ施設とその対応に
ことができる.詳細情報はパノラマ
ウスオーバ前
ウスオーバ後
内につながりがない場合
e the odd end is not visible.
表 1 パノラマ情報の詳細
Table 1 The details of a panoramic information.
名称 型 詳細
panoId text パノラマ写真の識別子
latlng geometory point パノラマ写真撮影地点の緯度・経度
yawdeg float パノラマ写真撮影方位(0 ∼360 ,0 :北)
link panoId text 隣接する panoId
表 2 DB のアイコン情報の詳細
Table 2 The details of a icon information in DB.
名称 型 詳細
id int DB 内の識別子
ontent type int 階層間をつなぐ施設の種類(「0 :上り階段」など)
latlng geometory point 階層間をつなぐ施設の緯度・経度
height double アイコンを表示する高さ
title text 階層間をつなぐ施設の名称と識別子
comment text 階層間をつなぐ施設の名称
user id int コンテンツ作成者の識別子
created at date コンテンツ作成時刻
ビューを用いて,また周辺の大局的な情報は 2 次元地図を
用いることを想定した提案手法のインタフェースを図 6 に
示す.
地上と地下それぞれのパノラマ写真が持つパノラマ情
報を表 1 に,階層間をつなぐ施設に表示するアイコンが
持つアイコン情報を表 2 に示す.パノラマ写真撮影地点
の座標は緯度・経度で記録されており,実装したパノラマ
ビューの球の中心を示す.アイコンにも緯度・経度と高さ
による座標が与えられているため,パノラマ写真撮影地点
との相対的な方位と距離,そして仰角を求めることで描画
できる.このときのアイコンの高さについては,パノラマ
写真の大まかな水平線に合わせて 0 としている.1 つの階
層間をつなぐ施設に対し,地上と地下それぞれのパノラマ
写真に合わせて 2 つのアイコンの座標を計算し,描画して
いる.地上のアイコンは地上のパノラマ写真上のみに,そ
して地下のアイコンは地下のパノラマ写真上のみに表示す
ることで,モード切替えにともない球の中心となる緯度・
経度がずれても,正しい位置にアイコンを描画できる.
図 6 提案手法のインタフェース
Fig. 6 The proposed interface.
-
Smart phone Other vehicle Smart home
Security Operation Center/
Insurance company/…
Internet
Edge server
Wi-Fi/
Bluetooth
4G/5G, LTE
Domain
Controller/IDS
IDS
ECUECU
ECU ECU ECU
In-vehicle
Infotainment
Central
Gateway
Router
ECUECU
ECU
ECU
CAN/CAN FD
Ethernet
CAN/CAN FD
OBD-II port
Switch
Switch
Switch
ECU
ECU
ECUECU
Domain
Controller
Domain
Controller/IDS
In-vehicle network
Ethernet
Hardware Trojan,
Malware
Connected
equipments
Sensing
data
Sensors
Actuators
DoS attack
情報基盤システム学研究室(B206
LAN
LAN
Wireless LAN
H27年度政府総合防災訓練の様子(2015年9月1日)
ネットワーク運
Software Defin
トラフィック解
超高精細画像IP
セキュリティ
マルウェア解析
ペアリング暗号
暗号のGPU実装
ユビキタスネットワーク
Delay Tolerant Network
MANET/VANET
Sensor Network
8. ・
1
・
7
!
Maintenance inspection work before dispatching
Washing car washing and cleaning the bus
Dispatching moving towards the first bus stop
Waiting waiting at the first bus stop
In service bus is in service
Taking a break taking less than 4 hours break
Out of service moving to the garage after the bus finishes service
Refueling refueling at a gas station
At garage parking at the garage
9. ・
8
!
81% 93%
12
11
Takuya Yonezawa, Ismail Arai, Toyokazu Akiyama, Kazutoshi Fujikawa, “Random Forest Based Bus Operation States Classification Using
Vehicle Sensor Data,” 2018 International Workshop on Pervasive Flow of Things (PerFot 2018), IEEE PerCom 2018, pp.819--824, Greece,
March, 2018.
12. Random Forest
Regressor
11
94% 96%
3
Hayato Nakashima, Ismail AraiKazutoshi Fujikawa, “Passenger Counter Based on Random Forest Regressor Using Drive Recorder and
Sensors in Buses,” 2019 International Workshop on Pervasive Flow of Things (PerFot 2019), IEEE PerCom 2019, pp. 561—566, Kyoto, March,
2019.
t paper
Fig. 5: Example of execution with OpenPose
Fig. 6: The sample frame with YOLOv3+Deep SORT
counting precision was not good because it can not be detected
as a human when each part of a body can not be estimated
well. Especially during boarding, due to the position of the
camera, it often shoots the back of the passengers, so the
detection accuracy is not high.
Fig. 2: The sample video of drive recorder
Fig. 3: The sample video of drive recorder of the last paper
Fig. 4: Example of execution with Background Subtraction
video has vibrations from the bus engine, and also the angle
of view is different(Fig. 4). In the last paper, we counted using
passenger’s sideways movement, but in the video of this paper,
there are few sideways movements and there are many moves
of depth. Occlusion tends to occur in the movement of the
depth, and it is difficult to correctly detect each passengers by
using the background subtraction method.
We also tried detecting human beings using OpenPose(Fig.
5). However, since OpenPose presumes parts of a body,
Fig. 5: Example of execution with OpenPose
Fig. 6: The sample frame with YOLOv3+Deep SORT
counting precision was not good because it can not be detected
as a human when each part of a body can not be estimated
well. Especially during boarding, due to the position of the
camera, it often shoots the back of the passengers, so the
detection accuracy is not high.
1) Detecting the human using YOLOv3: We use YOLOv3,
a method for object detection. The method regards object
detection as a regression problem to spatially separated bound-
ing boxes and associated class probabilities. It uses a single
neural network, and the network predicts bounding boxes and
class probabilities directly from full images in one evaluation.
Because it solves as a simple regression problem, processing
speed is fast. In other algorithms, we tried to detect a candidate
region of an object using techniques such as ”sliding window”
and ”Region Proposal”, so we often erroneously detect the
background as an object. In YOLOv 3, such erroneous detec-
tion is half of ”Fast R-CNN” [8] [9]. According to the detected
object, it is possible to obtain the size and the position of the
rectangular outline, the type of the object (person, car, food,
etc.) and the score of the judged object. The score indicates
the accuracy of the judged object in the range of 0 to 1. For
these reasons, we used YOLOv3 with pre-trained model to
detect humans. When the score of human beings is over 0.2,
this method detects humans as the blue box in Fig. 6.
2) Tracking the human using Deep SORT: The proposed
method tracks humans detected by YOLOv3 with Deep SORT.
YOLOv3+DeepSort
+
19. A platform for fixed
point observation!
(Ex. Atmospheric pressure sensed from the
buses on the move in an area of 50m
square)
18
Temporal
Coverage
Frequency
Spatial Coverage
Density
Trash truck
Bus
1010
1011
1012
1013
1014
1015
8:00 8:10 8:20 8:30 8:40 8:50 9:00
Atmosphericpressure[hPa]
Times of day
5
21. D
・ )) )( MI
► L ) N
92.5%(CAN ID 305)
ROC Area
. . .Cell1 Cell2 cell3
LSTM
Cell
CAN Bus
Actual packet
TH
predicted packet
MSE calc.
>TH
. . .
<TH
Anomaly
Benign