1 of 40

/SafeRoad

2022 Graduation Project

Class1_Team6

201735805 Changhun Kang

201735858 Donghun Lee

201935078 Jeongmin Oh

SafeRoad

2 of 40

Brief description

/01

Data & Model

/02

About application

/03

Marketing

/04

/ TABLE OF CONTENTS

/05

Appendix

SafeRoad

3 of 40

/Brief Description

/01

SafeRoad

4 of 40

SafeRoad

5 of 40

/서울시 안심이 앱

/ MARKET ANALYSIS (1)

/ Safer Way mobile

/ 블록 버스터즈

SafeRoad

6 of 40

/ MARKET ANALYSIS (2)

SafeRoad

7 of 40

/ MOTIVATION

It’s hard to

operate phone in

urgent situation!

Could you do it

for me?

SafeRoad

8 of 40

Detect anomal behavior

Easy report

Police station map

/ FUNCTION

NO NEED!

SafeRoad

9 of 40

/Data & �Model

/02

SafeRoad

10 of 40

/ COLLECTING METHOD

  • Use smart phone built-in sensor (Accelerometer/Gyroscope)
  • Assuming that the smartphone is naturally placed in the pocket
  • Position the screen toward the body at the upside-down state

SafeRoad

11 of 40

/ COLLECTED DATA (1)

  • Data from daily life (walking/stair-up/stair-down/standing)
  • Read sensor value every 0.05 seconds => Data size : 200,000
  • Sampling : Window size 40 lines(≈2.5 sec) & sliding 20 lines

🡪 10,000 samples

​

SafeRoad

12 of 40

/ COLLECTED DATA (2)

  • Filtering : Kalman filter* 🡪 eliminate the noise data
  • Normalization : Z-score** normalization

* Kalman filter

Row data

After filtering

** Z-score normalization

SafeRoad

13 of 40

/ MODEL (1)

Anomaly detection

Unsupervised learning

LSTM AE

Difficult classification

- The frequency of abnormal behavior is very low and there are many variables

Anomaly detection

- Finding abnormal patterns (anomalies), objects through data

Supervised inappropriate

- There are many variables that can occur while walking

- It is difficult to collect abnormal data

Mixed algorithm

- Algorithm that mixes LSTM and Auto Encoder

- LSTM : deal with sequence data

- Auto encoder : learn characteristics of normal data

SafeRoad

14 of 40

/ MODEL (2)

/Auto encoder

- Algorithm with the function that reconstructs the input data

- Compare the difference between the reconstructed result VS the learned normal feature

- Compare the difference with a predefined threshold

🡪 it is judged as abnormal data when the threshold is exceeded

SafeRoad

15 of 40

/ MODEL (3)

/LSTM

- Solve RNN’s long-term dependencies (time gap to get the information you need)

  • The repeat module has 4 interactive layers
  • Takes or discards partial values ​​from previous values

SafeRoad

16 of 40

/ MODEL (4)

/LSTM AE

- Replace cells in Autoencoder network with LSTM cells

LSTM

X

/Our Model

LSTM

LSTM

LSTM

Dense

SafeRoad

17 of 40

/ KERAS MODEL

/Keras model

/Model summary

SafeRoad

18 of 40

/ TRAIN INFO

/Train info

/Train environment

Batch size

Train epoch

Callback

16

50

7

CPU

RAM

GPU

AMD Ryzen 5 3500X 6-Core Processor 3.60 GHz

16.0 GB

GeForce GTX 1660 SUPER

SafeRoad

19 of 40

/ TRAINING LOSS

SafeRoad

20 of 40

/ SET THRESHOLDS

  1. Accelerometer X
  2. Accelerometer Y
  3. Accelerometer Z
  4. GyroSensor X
  5. GyroSensor Y
  6. GyroSensor Z

Calculate and save thresholds

SafeRoad

21 of 40

/ DETECTION

Threshold

Average of

6 sensors threshold

>

Anomaly

Average of mean squared error

Difference between input and output values

SafeRoad

22 of 40

/ MODEL ACCURACY

Detection Accuracy : Anomaly Behavior

SafeRoad

23 of 40

/About Application

/03

SafeRoad

24 of 40

/ INTRO SCREEN

GPS permission

SMS permission

CALL permission

SafeRoad

25 of 40

/ HOME SCREEN (1)

Longitude & Latitude to Address 🡪 ‘Geocoder’

current location

SafeRoad

26 of 40

/ HOME SCREEN (2)

police station

Mark the nearest police station 🡪 ‘Place API web service’

SafeRoad

27 of 40

/ MENU (1) - DETECTION

SafeRoad

28 of 40

/ MENU (2) – CALL

Firestore

SafeRoad

29 of 40

/ MENU (3) – USER INFO

Write to Firestore

SafeRoad

30 of 40

/ DEMO VIDEO

https://youtu.be/A03MYqSFPww

SafeRoad

31 of 40

/Marketing

/04

SafeRoad

32 of 40

/ SWOT

  • Intuitive icons and simple manipulations
  • Applications constantly analyze user behavior
  • Improve service by referring to feedback from other apps
  • Different reporting method from other apps
  • If the city runs its own app, it can implement the "safe return home" service together
  • Difficult to consider different usage scenarios
  • Behavioral analysis may be incorrect

SafeRoad

33 of 40

/ 4P strategy

Application that detects user's abnormal behavior

by using sensors on the smartphone

Free to use the application

Available on Google Play Store

Ads exposure to various platforms and make public as open source

/Product

/Price

/Place

/Promotion

SafeRoad

34 of 40

/ PROMOTION

/Kakao Talk

혼자 걷는 밤길,

불안할 땐?

/Instagram

SafeRoad

어둡거나 인적이 드문 골목길을 걸을 때 …

/GitHub Open Source

SafeRoad

35 of 40

/ IN-APP ADVERTISEMENT

  • App needs to be used quickly in emergencies
  • Ads should never inconvenience the use of apps
  • Always located at the top of the screen
  • Imprinting the location to the user
  • Preventing accidental pressing

/Banner ads

SafeRoad

36 of 40

/Appendix

/05

SafeRoad

37 of 40

/ PROFESSOR & MEMBERS

Prof. Youngmin Oh

youngminoh@gachon.ac.kr

Changhun Kang

chkangsc@gmail.com

Data processing

Donghun Lee

ldh02091877@gmail.com

Model training

Jeongmin Oh

ojm5155@gmail.com

App design

SafeRoad

38 of 40

/ URL

/GitHub

/YouTube

/Slack

* Click the icon

SafeRoad

39 of 40

/ REFERENCE

​

​

SafeRoad

40 of 40

/ THANK YOU!

SafeRoad application of Team 6;

2022 Graduation Project

SafeRoad