1. Crowd Density Estimation for Public Transport Vehicles
Marcus Handte, Umer Iqbal, Stephan Wagner, Wolfgang Apolinarski, Pedro Marrón (UDE)
Eva Munoz, Santiago Martinez (ETRA I+D)
Sara Izquierdo, Mario González (EMT Madrid)

2. Motivation
Existing information systems for public transport focus on timetables and routes
This is probably the most important piece of information but …
… in urban settings the same destination can be reached over multiple „similar“ connections
e.g. same bus line leaving every 3 minutes
e.g. two stops very close to destination
Question: how to select most comfortable trip
Idea: discriminate options based on number of passengers
Focus: how to determine this crowd-level automatically

3. Design Goals Create a system to determine the crowd-level achieving
Sufficient accuracy: provide a meaningful (i.e. representative) estimate, not necessarily perfect but good enough
Full automation: no additional work for existing support personal (e.g. driver or guards) and no work for the passengers
Low cost: hardware cost should be minimal to be feasible to deploy at a city scale (e.g. city of Madrid operates > 2000 busses)
Low latency: to support real-time operation (what is the current crowd-level), the latency of the solution must be low (~ few minutes)
Low privacy impact: non-intrusive system to ensure that solution is acceptable for passengers

4. Approach
Use of passive monitoring of WIFI-enabled devices carried by the passengers (i.e. phones)
Probe requests generated as part of active WIFI scans
Deployment of low-cost monitoring devices in vehicles
Real-time reporting of crowd-level via (existing) 3G connection
Use of (existing) GPS tracking to determine bus location
Attribution of crowd-level to specific route segment (i.e. from stop - to stop) using network information and GPS position

5. I‘ve seen this already, what‘s new?
Passive WIFI monitoring has been done before, e.g. …
A. B. M. Musa and J. Eriksson. Tracking unmodified smartphones using wi-fi monitors. In Proceedings of the 10th ACM Conference on Embedded Network Sensor Systems, SenSys ’12, pages 281–294, New York, NY, USA, ACM.
But, crowd-level estimation in vehicles is different because
the WIFI monitoring devices are moving through space
we are only interested in devices that are in the vehicle
Thus, we have to filter out WIFI devices of persons …
… walking next to the bus
… at the bus stop waiting for the next bus
… sitting in their backyard …

6. Crowd Density Estimation
Use of sliding window to count devices seen „consistently“
Experiment
Installation of 1 WIFI monitor in a bus in the city of Madrid
Collection of raw probe requests over 14 day period
Results
Almost 50% of the devices are only seen once
Most devices are seen again after (at most) 3 minutes
3 min. window, 1 min initialization

7. Vehicle Tracking
Goal: attribution of crowd-level to route segment (i.e. from stop – to stop) based on (imprecise) GPS position
Basis: modeling of routes as polylines consisting of stops and intermediate points (e.g. road intersections)
Approach: computation of shortest paths from GPS position to all polyline segments
Attribution to route segment with shortest path

8. Implementation
Development of inexpensive WIFI monitor using OpenWRT and TP-Link 3020 router (<50€)
Computes and sends number of passengers every 30 seconds
Development of 3 web services to
get the routes of the bus network
get the polyline for each route
get the route and gps position of a vehicle
Development of simple map-based crowd-level visualization {
”Id ”:4281 ,
”LineId ”:17 ,
”Loc”: {
”Lat”:− ,
”Lon”:− },
”Route”:33342 }

9. Evaluation Discussion Experiments
Full automation  no manual intervention required
Low cost  less than 50€ per vehicle for the WIFI monitor
Low privacy impact  local processing on the WIFI monitor, only crowd-level is stored (not the MAC addresses)
Experiments
Deployment of 3 WIFI monitors in 3 busses in Madrid for 3 weeks
Low latency  3 minutes + transfer time (usually < 1 minute)
Sufficient accuracy  comparison with manual counting shows a consistent ~20% detection rate

10. Conclusions
Existing information systems for public transport focus on timetables and routes
We argue that crowd-level information
can enable passengers to optimize their trips in cities
can be gathered with low cost in an automated and non-intrusive fashion using passive WIFI monitoring
Next steps
Integration with the bus navigator application developed by the GAMBAS FP7 project

11. Questions? average crowd level over 3 weeks
thicker line  higher crowd level