Biking in Michigan presents fundamental challenges, given underserved roads and seemingly endless winter seasons. While spring brings more cyclists out in droves to enjoy the return of warmer weather, both drivers and cyclists spent the entire year dodging potholes.
As drivers increasingly share the road with cyclists this time of year, the research Shan Bao and Fred Feng are spearheading takes on new meaning. Prof. Shan Bao, an associate professor and her colleague Prof Fred Feng, an assistant professor, both in the Industrial and Manufacturing Systems Engineering (IMSE) department at UM-Dearborn, have been gathering data and applying it for the last four years.
In a nutshell, their research aims to make bicycling safer.
Researchers are challenged to design self-driving algorithms that safely interact with other road users, needing a lot of scenario data to develop the right algorithms to run simulation after simulation. Bao says, “It’s complicated to predict a cyclist’s intention from a machine algorithm point of view. You need a lot of testing using simulations to make sure cars can yield to cyclists and pedestrians, but you are limited to simulations generated by engineers, and their data creates a bottom up approach to add a variety of real-world scenarios.”
Their data allows researchers to examine challenging conditions for cyclists and run simulations on a variety of bicycle and vehicle interactions. Various related projects, funded by the Toyota Research Institute, will one day help autonomous vehicles safely share the road with cyclists.
Studying bike rides in a ‘natural’ setting
In the past few years much progress has been made in self-driving technologies, yet Feng and Bao are quick to point out that there are still great challenges to be solved. With safety critical, they query if a car’s design is making the vehicle safer and safer, how can we extend that level of safety and protect other road users?
Says Bao, “The goal of this multi-year project is to understand the behavior of road users in terms of drivers and cyclists in the real world. One of the critical challenges is that self-driving cars need to share the existing infrastructure with other non-motorized road users such as bicyclists and pedestrians.”
One way to help is to collect naturalistic data of people riding bicycles in their everyday trips on real world roadways and use this da
ta to create guidelines, supports, and test scenarios to develop artificial intelligence algorithms for self-driving cars.
All this to say they tricked out a lot of bicycles and cars with cameras and gave them to people to use in their everyday lives. With no specific instructions on how they are supposed to drive or ride, they used the equipment however they saw fit and the sensors recorded their natural drive/bike behavior.
“We collected data for two years and have over 5,000 miles of bike riding data in a wide variety of scenarios. We plotted the GPS data for all the subjects and replicated an almost exact map to the city of Ann Arbor,” says Feng.
Examining both points of view
Bao is also an associate research scientist at the University of Michigan Transportation Research Institute (UMTRI), which has completed multiple significant naturalistic driving studies, gathering millions of miles of driving data. All the sensors and data acquisition systems on the bikes in these bike studies were instrumented and completed by the engineering group at UMTRI.
Researchers put radar, cameras and Mobileye sensors in the vehicle and can detect bicycles or pedestrians while driving. Next steps were to mine the collected naturalistic driving data to see how drivers interact with bicyclists. Research from that data looks at crash data sets to identify the most severe crash scenarios on the public roadways. Spoiler alert: it was drivers overtaking bicyclists, the leading fatal crash scenario.
They specifically examine the ‘corner cases’ – situations that don’t happen often but when they do they can be fatal. Leveraging data on the corner cases, they can build simulation models to test the rare, severe and unexpected vehicle/bicycle interactions.
Using data from both the driver’s point of view and the bicyclist’s vantage point allows self-driving algorithms to map rich, complex scenarios that help move autonomous vehicle safety technology forward.
That’s where Bao and Feng’s research comes into play.
Unique yet replicable studies
The naturalistic cycling study is one of the first studies of its kind, and Feng and Bao are glad to find more and more studies on the safety of bicyclists now underway. The data collected from the drivers’ perspective when overtaking a bicyclist is particularly useful, and this is one of the first studies using this methodology. The particular advantage from the vehicle’s perspective is that you can see exactly what is going on with the driver and have the entire continuous history of the driver in each scenario, allowing researchers to look into issues such as driver distraction and improper maneuvers.
“As a part of the analysis, we looked at driver distractions and found an alarming 8% of the overtaking occured when the driver was visibly distracted right before overtaking the bicyclist.” shares Feng.
Getting attention for their research
The driving data looking at how drivers overtake bicyclists gained acclaim in a top academic journal and in Forbes. Particularly interesting to model were factors like how lane markings, traffic conditions, driver behavior and the number of lanes in a road affected how much a driver decides to move over when overtaking a bicyclist.
“It’s interesting by itself to better understand how a driver makes decisions when coming up to bicyclists. These data may also potentially serve as a benchmark for self-driving algorithms,” shares Feng.
The research team has been busy presenting at conferences and has several additional papers underway that summarize their findings.
Around the corner
Bao and Feng also hope to continue the application of corner cases in simulation, developing better algorithms, and eventually move on to testing. “When we say autonomous vehicles are safe on the road, they need to be safe in all kinds of perspectives.” says Bao.
There’s also more data to collect. The naturalistic cycling study took place only in Ann Arbor. Feng states, “Some of the findings may not be directly applicable in other city settings. Riding in Manhattan is obviously different from riding in a college town like Ann Arbor. We hope to see if we can scale up to other places and collaborate with researchers in other locations to collect a wider variety of data.”
The research team sees applications for this data in other aspects of mobility as well. Shares Feng, “I believe that bicycling is a viable and important mobility option. When we talk about the biggest challenges we are facing as a society, such as climate change, traffic fatality, congestions, air and noise pollution, bicycling is a very promising and sustainable option. For the long term, I hope we can contribute more research on making the streets safer for people riding bicycles.
Bao adds, “We have observed that current technology has troubles in detecting small objects like bicycles, scooters or wheelchairs. In terms of a moving vehicle, accuracy of sensors are much less accurate for pedestrians or these types of modes of transportation. Our data will hopefully help with that.
“There’s a well-known theory “safety in numbers”. In the context of road safety, when more people choose to use non-motorized travel modes such as bicycling or walking, drivers get more used to seeing and interacting with them, so the per person risk will actually get lower.” continues Feng.
To ride or not?
After collecting and analyzing thousands of miles of bicycle/vehicle interaction data, both Bao and Feng are still happy to hop on a bicycle. Feng is no stranger to bikes; they are his primary mode of transportation.
“I, and many people I know, ride bikes on a regular basis. Our findings on distracted driving are unsettling. Eight percent of the overtaking are by a distracted driver – from a bicyclist’s perspective, that translates to being passed by a distracted driver every thirteen times they are passed.”
Research partners make a big difference
Bao says, “We’ve had support from Toyota Research Institute (TRI) from the very beginning, in three or four different projects related to this topic. We feel fortunate and excited by this opportunity to work with TRI. From our perspective what’s valuable and interesting is that TRI is focused on developing self driving cars that are safe for other autos and different types of road users. Cars need to see different types of road users, not just pedestrians and bicyclists but skateboards, wheelchairs and more. We are glad to work with a company that wants to make sure technology is safe for cars and for other road users. It is important for the public to trust the vehicle, and this technology is critical to develop – and very challenging.”
There are a lot of mutual benefits from working with companies like TRI. Bao explains, “We have enjoyed working with people from TRI. They have provided a lot of support and guidance for our research. They’re experts from industry so they bring up the practical needs in the technology development, and their involvement is a very important part of success of the project. Our team also has filed one patent with one of our TRI liaisons based on the findings of our TRI project.”