Automatic Recognition of PublicTransport Trips from Mobile Device
Sensor Data and TransportInfrastructure Information
Mikko Rinne Mehrdad Bagheri Tuukka TolvanenJaakko Hollmen
Aalto University, Department of Computer ScienceEspoo, Finland
Presentation at the KNOWMe: 1st International Workshop on KnowledgeDiscovery from Mobility and Transportation Systems
September 22, 2017
TrafficSense Research Project (2013–2017)
I TrafficSense — Energy Efficient Traffic withCrowdsensing (2013-2017)
I Research project funded by Aalto Energy EfficiencyResearch Programme (AEF)
I Project desciption:http://aef.aalto.fi/en/research/trafficsense/
I Motivational video:https://player.vimeo.com/video/109910585
Motivation for automatic detection of public
transport usage
I Understand the commuting habits of passengers at largeand over longer periods of time
I Enable compilation of door-to-door trip chains to assistpublic transport providers in improved optimisation oftheir transport networks
I Predictions of future trips based on past activities can beused to assist passengers with targeted information aboutopportunities and problems on the predicted trip chain
I Connecting passenger assistance
Targets of the experiment
I Create a dataset of mobile device sensor measurementsand public transport infrastructure data to test algorithmsfor automatic recognition of public transport trips
I Test initial algorithms
I Recognize both the public transport vehicle type (bus,tram, train, subway) and current line
I Fully automated sensor data collection with no effortfrom the user
I Power saving features to facilitate ’always-on’measurements with minimum impact on mobile devicebattery consumption
TrafficSense system and mobile application
Screenshots from the TrafficSense mobile application.
What was collected?
Mobile device:
I Position (combination ofsatellite, WiFi and Cell-IDmethods)
I Activity (’still’, ’walking’,’running’, ’on bicycle’, ’invehicle’ with confidencepercentages as estimatedby Google Play Services)
I Model of each device
Manual bookkeeping:
I trip start and end times,lines, boarding and exitstops etc.
Infrastructure:
I Static public transporttimetables and routeinformation
I Live samples of publictransport vehiclepositions.
I Measured train stop timesat stations for the date ofthe experiment.
Mobile application power saving state diagram
Sleep
Active
Timer running
No timer
Start timer
Timer expired
Activity not STILL
Activity not
STILL
Activity STILL
Location change
Sensor data filter algorithm
New Location
X min after last queued
Accept
Yes
Accuracy <= Y m
‘Good’ Activity !=
last queued
Yes Yes No
Distance to last queued > Accuracy
No Yes
Imperfections of sampled data
I Activity estimate sometimes incorrect, e.g. passengerriding a bicycle during a tram trip
I Intermittent or missing measurements (power saving,problems in positioning for rail traffic and in undergroundconditions)
I Positioning location hopping between the correct positionand a single distant point
I Live locations not available from all public transportvehicles at the time of the trial
Matching with live vehicle location
distance threshold R
distance r
score max(0, R-r)
user path samples
vehicle path
vehi
cle
path
dt
Matching with static timetables
Boarding stop End stop
Actual trip:
Sampled trip:
Planned trip:
IN_VEHICLE leg’s origin
IN_VEHICLE leg’s destination
IN_VEHICLE
BUS
2.BUS
1.WALK 3.WALK
tPT
tWb tWe
tV
WALK WALK
WALK WALK
tPTb tPTe
Recognition results
Table: Trip matching statistics for all recognition methodscompared to the manually logged trips (“line name” = matchingline name, “line type” = matching line type).
Static Old live New live Combined LoggedBus 8 3 4 9 15Bus (line name) 6 3 4 7 15Tram 8 8 8 9 15Tram (line name) 8 7 7 9 15Train 4 0 0 4 9Train (line name) 4 0 0 4 9Subway 19 9 17 25 58Public transport 40 20 29 48 97Public transport (line type) 39 20 29 47 97
Conclusions
I Combining the live and static methods 60% of bus andtram trips, and 43–44% of train and subway tripsrecognised with correct vehicle type
I Requiring bus line name dropped recognition to 47%
I Joint combined public transport recognition at 48% forthe correct line type
I Level adequate for e.g. public transport disruptioninformation filtering for frequently used lines
I Dataset available for improvements:https://github.com/aalto-trafficsense/
public-transport-dataset
Contact information
TrafficSense research project home page
I http://aef.aalto.fi/en/research/trafficsense/
Author contact information
I Mikko Rinne [email protected]
I Mehrdad Bagheri, [email protected]
I Tuukka Tolvanen, [email protected]
I Jaakko Hollmen, [email protected]