Principal Investigator

Anthony Rowe

Status

Completed

Start Date

Jan. 1, 2016

End Date

Dec. 31, 2016

Project Type

Research Applied

Grant Program

Private Funding

Grant Cycle

Old Projects

Visibility

Public

Abstract

Last year there were more than 52,000 people seriously injured in the United States due
to bicycle accidents. Bicycle transportation represents a healthy and clean mechanism
to improve both transportation efficiency and parking in urban areas. Traditional bicycles
are now even being augmented with electric drive systems to decrease the barrier of
adoption for commuters. Unfortunately, as bicycles become more capable and hybrid
and electric vehicle usage increases the ability for cyclists to perceive traffic danger is
decreasing. In this project, we explore how vehicle-to-vehicle (DSRC) communication,
differential positioning technologies and sensor-based trajectory estimation techniques
can be used to warn drivers about potential collisions with cyclists. This requires
developing rider models that can help anticipate the complex motion patterns of a cyclist
or pedestrian as compared to that of a vehicle. The algorithms developed as part of this
project can help augment future DSRC warning systems that will be deployed in both
cars as well as smartphones carried by cyclists. Our work up to this point has indicated
that poor GPS performance in urban spaces will drastically hinder the performance of
collision warning systems especially for less predictable targets like cyclists and
pedestrians. In the next phase of this project, we intend to look more closely at
alternative localization techniques based on emerging time-of-flight ranging technologies
and city building geometry data to help augment GPS systems.    

Description

The goal of this project is to use DSRC communication in conjunction with on-bike sensors to model the
trajectory of cyclists as an input to in-car collision warning systems. This is challenging for three main
reasons: (1) cyclists have multiple distinct riding modalities, (2) high-speed differential positioning
technology is still either expensive or in-accurate especially in urban environments and (3) warning
systems must intelligently trigger alarms to avoid driver drop-out. After the first year of the project, we
have an experimental platform that allows us to collect rich data about a bicycle’s motion with respect to a
vehicle both of which have ground-truth positions. We have initial data that shows trajectory that relies on
high-end sensors can be used for reliable collision warning in areas where differential GPS is available.
The focus of the next two years of this project will be to (1) improve trajectory tracking given more
affordable sensors and (2) exploring alternative mechanisms for precise localization in urban
environments.

It is difficult to predict the trajectory of cyclists because they often transition between riding with traffic,
riding along road shoulders or even riding or walking on sidewalks like pedestrians. In the last year of this
project, we experimented with an RTK GPS system that was only effective in the most ideal of
circumstances and was not practical in urban (or many suburban) environments. GPS in its current form
even with differential approaches will not be an adequate positioning technology based on the limited
number of visible satellites at any one time. Buildings cluttering the horizon further exacerbate poor GPS
performance. Even using state-of-the-art differential GPS technologies at ideal times of the day, we
found that it is nearly impossible to maintain locks with enough satellites to do lane-level localization. As
an extension to our previous work, we plan to more rigorously look at the problem of multi-modal sensing
to localize cars, pedestrians and bikes. To help augment GPS, we propose evaluating if RF time-of-flight
ranging technologies in conjunction with 3D models of urban spaces can be integrated into existing
DSRC infrastructure to provide highly accurate and robust outdoor localization in GPS-reduced areas.
We currently have an experimental bicycle platform shown in Figure 1 that is able to capture and log:
pedal cadence, bike speed, bike angle, rider position, helmet position, differential GPS, automobile
dynamics and ranging estimates from multiple RF sources between a vehicle and the bicycle. We have
also defined and collected benchmark data for a set of scenarios that allow us to quantitatively evaluate
different trajectory tracking and data fusion algorithms that could be used to evaluate collision-warning
systems shown in Figure 2. Initial algorithms that utilize differential GPS and basic inertial measurement
sensors show that we can predict cyclist and car locations reliably up to 2-3 seconds into the future.

In the next phase of the project, we intend to explore if a significant level of trajectory prediction is
possible without relying heavily on differential GPS. We propose using ranging technologies that combine
fixed beaconing (located in traffic lights) and peer-to-peer automobile beaconing. RF time-of-flight (TOF)
systems are becoming standard in both WiFi as well as UWB chipsets and provide a promising option for
outdoor positioning systems. RF TOF will almost certainly be possible with next generation DSRC radios
that historically inherit functionality from similar WiFi chipsets. Unfortunately, ranging alone is not enough
in urban spaces that are cluttered with buildings and large vehicles that interfere with most standard
trilateration approaches. We believe that 3D models of buildings and maps in cities can be integrated
with trilateration solvers to not only mitigate errors due to multi-path, but also constraining the possible
solution space leading to improved accuracy. Working with both Bosch Research and Samsung, we
have shown that this type of geometry-aware localization provides significant gains when applied to
indoor spaces constrained by floor plans. With companies like Google, Apple and Uber mapping outdoor
environments, there should be enough information available about 3D building and road geometry to
apply similar techniques to outdoor spaces. With connected vehicles, we can also leverage ranging data
between automobiles in a mesh network to act as temporary anchor points. Since automobiles have
more constrained motion paths and better odometry, they should provide a crowd-sourced plethora of
more reliable position references. One could even imagine fusing in WiFi and cellular data to further
improve location accuracy in the same manner in which smartphones use multiple sources.

Timeline

Year 2:
• Collect data with UWB TOF ranging beacons on bike, car and placed in infrastructure
• Improve current trajectory tracking algorithms to be less reliant on differential GPS
• Develop range-based geometrically constrained solver that utilizes 3D data from Google Earth
• Develop better automobile logging system that includes a CAN-bus interface and UWB in order to
support the crowd-sourcing of mobile localization anchors
Year 3:
• Develop and deploy prototype phone app that can collect cycling characteristics
• Integrate ranging system into Pittsburgh DRSC deployments
• Evaluate system on real roads

Strategic Description / RD&T

Deployment Plan

1.Potential partners include Bosch Research and Samsung Research. Both will provide cost-share
to explore tracking as well as geometrically constrained trilateration.
2. Various cycling groups (like BikePGH) and the new bicycle share program in the area would be
great partners for testing any mobile phone based data collection. We have had incredible
support and many volunteers reach out to us about being test candidates once our platform
becomes more refined.

Our current deployment plan will be to collect data on our heavily instrumented bike and car in
controlled environments. We eventually intend to test the bicycle system with existing DSRC
radios deployed near CMU’s campus on a group of commuting cyclists. Our mobile phone data
collection application will be available to a wider audience to help capture route and phone inertial
data to help build cyclists trajectory model that can feed into our collision warning algorithms.

Expected Outcomes/Impacts

1) Trajectory dataset consisting of onboard sensors with differential position to vehicle
given differential GPS readings as ground truth.
2) Collision warning benchmarks.
3) Improved Automobile Data Logging System
4) Urban geometry-constrained RF TOF trilateration solver for localization
5) Prototype in-car alert system that works with transponder equipped bicycle.

Expected Outputs

TRID

Individuals Involved

Email	Name	Affiliation	Role	Position
agr@ece.cmu.edu	Rowe, Anthony	ECE	PI	Faculty - Adjunct

Budget

Amount of UTC Funds Awarded

$84998.00

Total Project Budget (from all funding sources)

$84998.00

Documents

Type	Name	Uploaded

Match Sources

No match sources!

Partners

No partners!