1.
MOBIUS: Smart Mobility Tracking
with Smartphone Sensors
1Open University of The Netherlands 2Maastricht University
3rd December 2020
Daniele DI MITRI1 Khaleel ASYRAAF1
Kevin TREBING2 Stefano BROMURI1
daniele.dimitri@dipf.de - @dimstudio

2.
The authors
Daniele DI MITRI
dimitri@dipf.de
Khaleel ASYRAAF
khaleel.asyraaf@ou.nl
Stefano BROMURI
stefano.bromuri@ou.nl
Kevin TREBING
kevin.trebing@
student.maastrichtuniversity.nl
Center for Actionable
Research of the Open University
MOBIUS: Mobility Tracking w. Smartphone Sensors 2

3.
Outline presentation
• The authors
• Background
• Related Work
• Functional requirements
• System Architecture
1. Data Collection
2. Data Storing
• Collected dataset
3. Data annotation
4. Data Analysis
• Results
• Discussion
• Conclusions
MOBIUS: Mobility Tracking w. Smartphone Sensors 3

4.
Background
• MOBIlity for Urban Sustainability (Mobius) is a project part of the
SafeCity programme of the Open University of The Netherlands
• The municipality of Heerlen monitors citizens journeys
throughout the city using paper based questionnaires
• MOBIUS idea:
• replace questionnaire with a mobile application (android)
• classify the transportation mode using Neural Networks (NN)
• collected labels asking user to self-report using the mobile
apps when the journey happens
MOBIUS: Mobility Tracking w. Smartphone Sensors 4

5.
Related Work
• cell phone signal coverage and classified "stationary", "walking",
and "driving”
• 80%-85% accuracy (Anderson, 2006; Sohn, 2006)
• proximity to bus-stops and rail-lines as well as GPS data
• accuracy of 93% (Stenneth, 2011)
• Only accelerometer of the cell phone (Hemminki, 2013; Liang, 2019)
classify "stationary", "walk", "bus", "train", "metro", and "tram"
• achieve 80% and 95% accuracy, respectively.
• Using a Bi-LSTM (Zhao, 2019) with a MLP using acclerometer and
gyroscope readings on the classes "stationary", "walk", "run",
"bike", "bus", and "subway”
• achieve 92% accuracy
MOBIUS: Mobility Tracking w. Smartphone Sensors 5
Our classifier
uses a similar
CNN architecture
as (Liang, 2019).

6.
Functional requirements
(FR1) Data collection
• Long periods data collection of accelerometer,
gyroscope, and GPS sensor data.
• Data collection should be a feature in the
background
• Accelerometer and gyroscope data sampled at
60Hz (60 updates per seconds)
• GPS coordinates sampled every 10 seconds
• Data gathering should not drain the battery
(FR2) Data storing
• Sessions to be stored into compressed format
(FR3) Data annotation
• Self-report the mode of transportation
• Activate a toggle for each mode of transportation
• The user-annotations should be editable
(FR4) Data analysis
• suitable representation for machine learning
(FR5) Privacy
• User should be able to deactivate the GPS
coordinate tracking
• Classification should rely on Acceleroemter and
Gyroscope
MOBIUS: Mobility Tracking w. Smartphone Sensors 6

7.
MOBIUS: Mobility Tracking w. Smartphone Sensors 7
System
Architecture

8.
(1) Data Collection: MOBIUS client
Android mobile application with 2 flows of data collection:
• Sensor data
• Accelerometer gyroscope - 60Hz (60 updates per second)
• Global Positioning System (GPS) - every 10s. Can be
deactivated
• User starts/stop the collection which runs in as deamon
service in the background
• Self-reported annotations:
• “Walking”, “Running”, “Biking”, “Train/bus”, “Car”.
• User toggles the annotations at start and ends of their
journeys
• At the end of the day, the user uploads the compressed
data in the server
MOBIUS: Mobility Tracking w. Smartphone Sensors 8
Source code: https://github.com/khaleelasyraaf/Mobius_Client

9.
(2) Data Storing: Mobius Server
MOBIUS: Mobility Tracking w. Smartphone Sensors 9
• Each session is stored in a zip folder
• When the user hits the ‘Upload’ button, the zip
folder is transfered from the Cliebt to the Mobius
Server
• The Mobius server is a RESTful web service
implemented with Python Flask
• Each session folder is composed by three CSV files
contaning the data from: (1) acc & gyroscope, (2)
GPS, (3) annotations
• The files are later transformed into MLT-JSON
format to be compatible with the annoation tool
(next slide)
MLT-JSON data format
CSV data format
Source code: https://github.com/HansBambel/Mobius_server

10.
(3) Data Annotation: Visual Inspection Tool
MOBIUS: Mobility Tracking w. Smartphone Sensors 10
Source code: https://github.com/dimstudio/visual-inspection-tool

11.
Collected dataset & pre-processing
MOBIUS: Mobility Tracking w. Smartphone Sensors 11
• The research team (4 users) collected ~47 hours recordings
• The dataset was pre-processed and used as input for training
the Neural Networks
• We only consider Acc. and Gyr. (x,y,z)
• A sliding window approach was used (size 512 = 8.7s)
• It resulted into 168476 samples
• We removed gravity from the Accelerometer
• Applied smoothing at each sensor stream
• We also tried using only the Acc. magnitude

12.
(4) Data Analysis: SharpFlow
Compared approaches:
1. Convolutional Neural Network
(CNN)
2. Bi-directional Long Short Term
Memory (Bi-LSTM)
3. Dummy classifiers
1. Most frequent class
2. Dependent on class
distribution
MOBIUS: Mobility Tracking w. Smartphone Sensors 12
Source code: https://github.com/dimstudio/SharpFlow
Grafic Representation of the CNN

13.
Results
MOBIUS: Mobility Tracking w. Smartphone Sensors 13

14.
Discussion
• Gyroscope sensor seems to have a minor performance
impact, thus excluding it will extend battery's life and
storage space
• Bi-LSTM did not achieve over 90% as in the literature,
due to smaller time window (2.56s) with a sampling rate
of 50Hz unlike our longer time (8.7s)
window was sampled at 60Hz
• The CNN outperforming the LSTM approach due to cyclic
motions are easy to spot by the spatially invariant
kernels of the convolutions
MOBIUS: Mobility Tracking w. Smartphone Sensors 14

15.
Conclusions
• We introduced Mobius, a system for Smart Mobility Tracking
with Smartphone Sensors
• Mobius can automatically classify transportation mode using
only Accelerometer and Gyroscope data
• We tested Mobius using 47 hours of recorded data
• We achieved an overall 89% using a CNN
• Scalable client-server architecture which can accommodate
data from multiple clients
• User in the loop can self-report mode of transport
• We released all our code Open Source on GitHub
• We are aware that tracking apps pose privacy concerns due
to possible surveillance
MOBIUS: Mobility Tracking w. Smartphone Sensors 15

16.
Future applications
• We plan to develop an Android Wear (smartwatch) version of the
Mobius client app
• Add more modalities such as the heart-rate
• The approach proposed by Mobius can be used for Human Activity
Recognition:
• Daily Life Activities
• Sports performance monitoring
• Physical rehabilitation Tasks
• The trained models can be embedded in the android app to avoid
uploading in the server
MOBIUS: Mobility Tracking w. Smartphone Sensors 16

17.
Thank you!
Any question?
Code:
github.com/dimstudio
Connect:
dimitri@dipf.de
@dimstudio