1. Evan Bialostozky September 16, 2009
Development of a Mode Detection Algorithm for GPS-Based Personal Travel Surveys in New York City
Evan Bialostozky
September 16, 2009
This presentation will be about the project on which I spent the largest portion of my time during my 9/11 Memorial internship, the development of a mode detection algorithm for GPS-based personal travel surveys in New York City. I should warn you that I could easily speak for 10 hours on this subject, so I will do my best to make these 10 minutes as clear and as comprehensive as possible.

2. Quick History of Personal Travel Surveys
ongoing shift from “traditional” paper diaries to GPS-based surveys
advantages:
easy, precise collection of travel time,
distance, route choice
disadvantages:
trip purpose?
travel mode?
The best place to start is with a brief history of personal travel surveys.
The usual method since the 1970s for gathering survey respondents’ travel data has been a paper diary. The volunteer is asked to fill out a form with information about a randomly-assigned weekday: where they traveled, at what time, why they went there, how they got there. But researchers have found that human memory isn’t so accurate: people forget about some trips, round off numbers, misunderstand instructions, etc.
GPS technology has been available for easy and precise use by the public since Travel survey researchers quickly realized the great benefits that could come from re-thinking the design of these surveys. Surveys across the globe have begun to use only data collected by GPS devices, eliminating paper diaries entirely. A GPS logger is loaned to a survey respondent for a period of time; the volunteer carries the small device around with them during all of their trips; and the logger is returned to the surveyors, where the data is downloaded. Respondent burden using this method is minimal.
GPS is an incredibly precise way to collect time and location data. Information requested from survey respondents such as travel time, distance, and route choice can now be collected with tremendous accuracy. The downside is that two variables crucial to travel demand modelers, trip purpose and travel mode, are not readily apparent with GPS.

3. Objectives of Algorithm
determine mode used from raw GPS data
consider 5 modes:
car
bus
subway
commuter rail
walk
account for potential GPS signal distortion in high-density New York City
This project therefore looked to develop a computerized methodology using GIS functionalities to determine travel mode based solely on basic raw GPS data, time and location. The 5 modes listed here were considered: car, bus, subway, commuter rail, and walking.
A complicating factor in New York City is the high-density land use of many neighborhoods. Skyscrapers create what are known as urban canyons in areas such as Lower Manhattan and Midtown. Two different issues arise in these areas that combine to create an urban canyon effect.

4. Urban Canyon Effect: Reason 1
GPS devices work by receiving signals from satellites that orbit the Earth. However, if a GPS device has a limited view of the sky, the number of satellites that the device can “see” probably decreases.

5. Urban Canyon Effect: Reason 2
Additionally, urban canyons increase the likelihood that some signals picked up by the GPS device bounce off the surrounding buildings on their way down, as shown by the purple lines in this graphic. Both of these problems lead to decreased data accuracy.
There are methods to minimize urban canyon effects, the specifics of which I don’t have time to go into today. But despite implementing these methods, the results showed a noticeable decrease in mode detection accuracy in neighborhoods with many urban canyons, as I will explain later.

6. Step 1: Division of Data into Trips
The determination of travel mode from GPS data can be boiled down into 3 steps. The first step is the division of a day’s worth of data into trips. What exactly defines a trip? A trip is travel between two activities. And how can one identify activity nodes? GPS data is generally collected at a regular interval, say every 5 seconds. It makes sense then that activity nodes will appear visually as clusters of points. Activities, as opposed to the travel between them, take place in one location: your home, your workplace, the grocery store.
In this map, it is very clear: I bought lunch during the top cluster, then walked down Amsterdam to the park at the bottom cluster to eat my sandwich.
So by finding clusters and only analyzing the moving points between them, one only analyses the travel portion of the day’s data.

7. Step 2: Division into Trip Segments
Assumptions:
underground travel when 2 consecutive points are more than 120s and 250m apart
walk segment at every modal transfer
Not all trips are taken with only one travel mode, however. Say to get from my home to NYMTC I walk, then take a bus, then take a subway, then walk some more. This 1 trip must be divided into 4 trip segments in order to correctly identify all the modes used.
New York City’s transportation network actually provides a lot of help in this regard. GPS devices cannot calculate location underground, as it is impossible to receive a satellite signal. So no data is collected during an underground subway trip. There are points as you descend the steps at your first station, then there is a significant gap both in terms of time and location, then there are points once you come up to street level again after getting off. So the assumption can be made that an underground travel segment occurred if two consecutive points are more than 120 seconds and 250 meters apart. This divides trips into underground segments and aboveground segments.
What happens, though, if you switch modes aboveground? Let’s say you take a bus to an elevated train station in the Bronx. The second general assumption made is that there is at least a short walk segment at every transfer. As we will see on the next slide, a walk segment must be at least 60 seconds long. It is very rare to be able to switch modes in less than 60 seconds in New York. In this Bronx example, by the time you get off the bus, climb the stairs up to the station, go through the turnstile, wait on the platform for the next subway, and get on the subway, it will probably be a few minutes of slow-moving travel.
[QUICK ]

8. Step 2: Division into Trip Segments
Characteristics of walk segments:
at least 60s long
maximum speed ≤ 10km/h
average speed ≤ 6km/h
And this slow speed is the most distinctive characteristic of a walk segment. It is relatively easy to pick out walk segments from a day’s GPS data. And for each walk segment, we can assume that the travel before it used one of the 4 non-walk modes, as did the travel after it.
Now we have all travel divided into trip segments, and each segment is temporarily identified as walk, non-walk aboveground, or non-walk underground. Mode detection now continues with the non-walk segments. This occurs in 3 sub-steps.

9. Step 3a: Aboveground Subway/Rail Detection
The easiest type of segment left to identify is aboveground subway or rail, so that is done first. GPS data is matched up to GIS layers of subway and commuter rail lines within the 5 boroughs. The accuracy of GPS data makes this very straightforward, as can be seen in this map. The survey respondent clearly was traveling on the LIRR in Queens. It is obvious that the travel was not on a parallel street.

10. Step 3b: Car vs. Bus How to distinguish a bus from a car?
A bus segment:
begins and ends near a bus stop
travels only along bus routes
has a maximum speed lower than 55mph
has a maximum acceleration lower than 1.5m/s2
Once we have identified aboveground subway and rail, we are left with on-street travel that is not walking. This raises the tricky question of how to distinguish a bus from a car based only on time and location. The differentiation is quite fuzzy, but there are some clues that can be used.
A bus segment must start and end near a bus stop. This is checked against a GIS layer of all the City’s bus stops.
It must travel only along streets that are bus routes. Again, the check is made with the appropriate GIS layer.
According to the MTA, New York City buses are physically incapable of speeds faster than 55mph or acceleration faster than 1.5m/s2.
Yes, it is certainly possible to drive one’s car slowly and along a bus route from a parking spot next to a bus stop to another parking spot next to a bus stop – but usually a segment with these characteristics will be onboard a bus.

11. Step 3c: Signal Gaps
Finally, there are the underground segments. Here, the clue comes from the location of the GPS points before and after the gap. If they are near subway stations, then the segment can be identified as subway.
Similarly, a trip through the Brooklyn Battery Tunnel, for instance, can be identified as car or bus.
For those of you who know West Harlem, this map clearly shows the appearance and disappearance of GPS signals as you travel aboveground, belowground, and back up again. And yes, this map has been rotated from the norm to fit better on the screen. This survey respondent traveled on the 1 train from City College [point] to 86th St [point]. But the 1 train, because of the strange topography here, comes aboveground between about 135th St and 122nd St, hence the brief re-appearance of GPS points here [point].

12. Results
79.1% success rate
urban canyon effect causes lower success rates in high-density neighborhoods
Data from two small pilot surveys were run through the algorithm, with a 79.1% overall success rate of correctly identifying the travel mode on a particular segment.
While there were many interesting conclusions found after looking at the results in more detail, I will mention one particular result that relates back to the urban canyon effect discussed earlier. The success rate for those respondents who traveled to Lower Manhattan was notably lower than for those who did not. Urban canyons cause errors in location data, so tests that help identify mode such as checking the proximity to a subway station are not nearly as reliable downtown, leading to greater error in the final results.

13. Benefits for NYMTC Future regional household travel surveys:
more accurate
possibly multi-day data
more cost-effective
NYMTC deploys a regional household travel survey approximately once every 10 years. In the past, this has been done using the traditional paper diary method, though the survey coming up in a few months will contain a small GPS component. In the future, surveys that are entirely GPS-based might become feasible by using automated algorithms such as this one. Advantages include that the data collected would be more accurate. Also, it would be more feasible to collect more than one day’s worth of data from a household – carrying a GPS device around for an extra day or two is a minimal additional burden to the volunteer, while now, household travel surveys are almost always one day only because of the large burden of the paper diary.
It is also quite possible that a GPS-based survey would be more cost-effective, especially when considering the value gained from the more accurate data. Prices of basic GPS loggers continue to decrease, and the cost of mailing a few devices on loan to a household would probably be less expensive than the high cost of labor now required for phone calls to follow up with respondents and for cleaning data from inaccurate paper diaries.
Using algorithms such as this one, especially with continued improvement of their accuracy, would greatly benefit NYMTC and its travel demand model.

14. Acknowledgments NYMTC/UTRC Hongmian Gong (Hunter College, CUNY)
Jorge Argote (NYMTC)
I quickly want to thank NYMTC and UTRC for sponsoring this program and making this invaluable year possible. And special thanks also to Hongmian, my advisor at Hunter, and Jorge, my advisor here at NYMTC.

15. Questions?
I don’t know if there is time for questions now, but I’ll be glad to talk to anyone after the other presentations. Thanks!