Principal Investigator

Anthony Tomasic

Status

Completed

Start Date

Jan. 1, 2017

End Date

Dec. 31, 2018

Project Type

Research Advanced

Grant Program

FAST Act - Mobility National (2016 - 2022)

Grant Cycle

2017 Mobility21 UTC

Visibility

Public

Abstract

Knowing how crowded a bus can be (the bus "load") is a recurring theme among riders with disabilities and those who are busy with other tasks. Waiting for a bus that passes by due to being full can lead to severe repercussions, both in time and personal safety. Failure to board a bus results in riders being stranded for long periods of time. This is especially dangerous in poor weather or lighting conditions and when in unsafe neighborhoods. The problem is compounded for those with disabilities due to increased risk getting lost (cognitive and visual disabilities) and of street crime. 

Almost every public bus system in the country uses automatic passenger counting (APC) systems in some form to detect bus load (in part, because of public transit system reporting requirements). Every APC system has a certain amount of error. In aggregate these numbers are relatively accurate because the undercounting and overcounting error balances out. However, APC systems may have very high error for the tracking of an individual bus. We have discovered cases where empty buses have APC values of plus or minus 50 passengers. This local error means that APC systems are inadequate as an information source for a transit information system to notify users about the crowds on a particular bus.  And in any case, APC systems are almost never real time. They simply gather information and save it on the bus. This information is harvested when the bus returns to the depot.

Our objective is to produce relatively accurate estimates of the availability of seating on a bus (a proxy for knowing the actual number of people on a bus) through the use of sensors on mobile phones - a form of autonomous sensing of the environment. This project seeks to detect via crowdsourcing the bus load (or at least the availability of seats) via the Tiramisu app and evaluate user acceptance and system performance.    

Description

Challenge

Detecting the busload directly on a bus is a difficult task. However, almost every passenger on a bus carries a smartphone. All smartphones contain (i) accelerometers to measure phone movement (and vibration), (ii) a computer, and (iii) telecommunications to the internet. Using accelerometer data, we will attempt to build a  model that correlates phone movement to the user activity states. The states we are trying to detect are: user on a bus versus user not on a bus, user sitting versus standing, user holding the phone versus phone in pocket/bag. Given these states we can detect if a person is standing on a bus. We then hypothesize that people rarely stand on a bus when a seat is available and that the number of people standing is correlated to busload (relative to the bus capacity, which is known).. However, there are many technical challenges in constructing this system.

Work Plan

Our first technical challenge consists of building a model for user activity. While user activity models are widely reported in the industry, our particular combination of user states has not be previously studied. To build a model, we will first construct a prototype system that measure accelerometer activity and allows the user to 'hand label' the user state that they are currently in. We will then deploy a collection of researchers to label the various states and use the prototype to gather the appropriate data. Given this data, we will construct a machine learning model, using appropriate experimental validation techniques. In particular, we will use a training, development, and test sets of data. The test set is used only once when the performance of the system is finally tested. (If the test fails, the test data becomes training data and more test data is gathered.)

An additional task measures the hypothesis that standing people are rare. One approach would be to sample all (approximately 800) buses in the PAC across all trips and count the number of people standing and busload. However, this approach is extraordinarily expensive. Instead, we utilize past APC data to attempt to predict when buses are close to having the busload equal the seating capacity. (Note that packed buses are less useful for our error measurement just as empty buses are less useful.) Targeting these particular bus trips and stops we will again send researchers into the field to measure the number of people standing when a seat is available. 

Finally, we will first model single point in time models to estimate the availability of seats. As a second step we will look at dynamics to improve our models. (For example, buses generally start empty and fill to capacity and then quickly empty at certain critical points.)

The team will focus efforts on methods that support deployment of prototype features into the Tiramisu Transit app. We have the ability to run randomized trials by deploying different feature sets to groups of users (e.g., Tomasic et al, 2014). The project work plan is designed around these two elements and in two phases. An initial implementation and deployment of a more straightforward approach will provide valuable data for a more technically challenging, yet more user-friendly method.

Daisy Yoo, John Zimmerman, Aaron Steinfeld, and Anthony Tomasic. 2010. Understanding the space for co-design in riders’ interactions with a transit service. Proceedings of the 28th international conference on Human factors in computing systems, ACM, 1797–1806. 

Yun Huang, Anthony Tomasic, Yufei An, Charles Garrod, and Aaron Steinfeld. 2013. Energy efficient and accuracy aware (E2A2) location services via crowdsourcing, The 9th International Conference on Wireless and Mobile Computing, Networking and Communications (WiMob). 

National Council on Disability (NDC). 2015. Transportation Update: Where We’ve Gone and What We've Learned. Washington, D.C. Retrieved from http://www.ncd.gov/publications/2015/05042015/ 

Aaron Steinfeld, Leslie Bloomfield, Sarah Amick, Yun Huang, Qian Yang, William Odom and John Zimmerman. In preparation. How location-aware mobile information increases accessibility to transit. 

Anthony Tomasic, John Zimmerman, Aaron Steinfeld, and Yun Huang. 2014. Motivating contribution in a participatory sensing system via quid-pro-quo. In Proceedings of the 17th ACM conference on Computer supported cooperative work & social computing (CSCW '14). ACM, New York, NY, USA, 979-988.

Timeline

March – May 2018: Build prototype for accelerometer labeling, label dataset, build machine learning models. Observational study of riders standing on buses. 

June – August 2018: Additional modeling of user accelerometer data. Additional experimentation. Paper writing.

September - October 2018: Addition of model to Tiramisu.

November 2018: Measurement of in vivo system. 

December 2019: Additional paper writing, documentation, wrap up reporting on project.

Strategic Description / RD&T

Deployment Plan

The team has an existing technology transfer agreement with the lab’s spinout, Tiramisu Transit LLC. The company has received two SBIR awards, one from US DOT and one from the National Institute on Disability, Independent Living, and Rehabilitation Research.

Expected Outcomes/Impacts

Demonstrate feasibility of modeling new user activities via accelerometers with 90% accuracy for user state prediction. Ability to relatively accurately predict in real-time the busload of bus, for buses that have a Tiramisu user.

Expected Outputs

TRID

Individuals Involved

Email	Name	Affiliation	Role	Position
steinfeld@cmu.edu	Steinfeld, Aaron	Robotics Institute	Co-PI	Faculty - Research/Systems
tomasic@cs.cmu.edu	Tomasic, Anthony	LTI	PI	Faculty - Research/Systems
johnz@cs.cmu.edu	Zimmerman, John	HCII	Co-PI	Faculty - Research/Systems

Budget

Amount of UTC Funds Awarded

$100000.00

Total Project Budget (from all funding sources)

$100000.00

Documents

Type	Name	Uploaded
Progress Report	31_Progress_Report_2018-09-30	Aug. 16, 2018, 7:17 a.m.
Publication	Real_Time_Detection_of_Crowded_Buses_via_Mobile_Phones_2.pdf	Jan. 22, 2019, 9:12 a.m.
Final Report	Real_Time_Detection_of_Crowded_Buses_via_Mobile_Phones_Final_Report_caPaOeA.pdf	Feb. 11, 2019, 4:34 a.m.
Progress Report	31_Progress_Report_2019-01-01	March 18, 2019, 10:59 a.m.
Publication	A Long-Term Evaluation of Adaptive Interface Design for Mobile Transit Information	March 21, 2021, 4:50 p.m.

Match Sources

No match sources!

Partners

No partners!