UW researchers using $25 handlebar sensors find correlation between close passes and past collisions

Images from the ProxiCycle research paper in the Association for Computing Machinery's digital library.

You know that street where car traffic feels more uncomfortable because people tend to pass you too closely? You're not imagining things, it really is more dangerous, new research suggests. A UW research team has developed a cost-effective way to collect on-the-ground data to help identify streets where close passing is common, and after two months of testing they have found a strong correlation between close pass measurements, people's perceptions of cycling safety, and past collisions.

The team-led by Joseph Breda, Keyu Chen, Thomas Plotz and Shwetak Patel-created ProxiCycle, a system that takes proximity data from a $25 sensor attached to a bike rider's handlebars and sends it via a mobile phone to researchers. 15 bike riders participated in the initial two-month test, mostly recruited with the help of Seattle Neighborhood Greenways. They recorded 2,050 close passes from 240 rides, and eight riders accounted for about 80% of the miles. They also conducted a survey and interviewed participants in an effort to compare their perceptions of safety with the data collected. You can read their findings for free in the Association for Computing Machinery's digital library.

In some ways, the findings here are obvious and add to existing knowledge about street designs. Streets without dedicated space for cycling are more dangerous, which is why we build bike lanes. But while someone could sit in a tower downtown and look through maps and design documents to identify streets with designs likely to be dangerous to streets with designs that should be safe, it is important to ground-truth high-level meta analyses. For example, perhaps a street has a bike lane design that is inadequate in key locations but the city's work-planning process gives it a lower priority for improvement because their database shows it already has a bike lane. These sensor measurements are completely agnostic to the type of infrastructure that exists and could perhaps help identify those key problem points.

ProxiCycle data could also be used to improve cycle routing applications to either label segments with a higher likelihood of close passes or route people to routes with fewer close passes.

The major value of correlating close pass readings with past collision history, at least in my view, is that we can now look for locations with those dangerous readings but no or limited collision history and make changes before a collision happens. Of course the city should fix the designs of streets where people are already getting hurt, and we don't need a bike sensor reading to tell us that. But if Seattle is going to reach Vision Zero, we need to also anticipate collisions and take preventative action. It could be a great tool for augmenting a tool like SDOT's innovative Bicycle and Pedestrian Safety Analysis, which attempts to identify street conditions known to be dangerous whether they have a recent collision history or not.

Of course, there are many caveats to discuss. The map published perhaps shows the most obvious issue: The data is richer where participants are riding. It's a UW-based study, and sure enough the data gets heavier around UW campus. The data is much lighter in southeast Seattle and SoDo, where a disproportionate share of the bicycling injuries and deaths occur. The researchers are not claiming that their map is a comprehensive map of bicycle risk in Seattle. Instead they use it to compare their data with collision history to see if their findings are correlated, which they were. I'm not sure how you would get a proper survey of the whole city. Is it even ethical to send someone biking the full length of 4th Ave S for sake of data collection?

The map also seems to show a lot of data points along the Burke-Gilman Trail, presumably a sign that the sensors are picking up close passes with other trail users. People on bikes can more safely and comfortably operate close together, so while you should absolutely be giving other trail users space when passing, a person on a bike passing within three feet feels much different than a person in a car doing the same. The paper's description says they are only trying to measure when a car passes.

One other possible issue is that speed differential is a vital component to safety in addition to proximity. Many Seattle neighborhood side streets are very restrained in width, so I often end up in close proximity to a car. The constrained width also leads to much slower traffic movement, and very often one of us is fully stopped while the other passes. This situation is far safer than a similarly close pass when I'm going, say, 12 mph and someone in a car is doing 30 mph or more. Proximity is an interesting new datapoint to have on hand, but it will always need to be considered in context.

There's also the issue that rider behavior affects the results. For example, someone following vehicular cycling best practices (Cascade Bicycle Club has a program they call Urban Cycling Techniques" that cover these concepts) would be more likely to take the lane" when there is not enough space for safe passing. So their sensor might not pick up a close pass that a more novice rider (or someone who otherwise doesn't feel comfortable taking the lane) might experience if they are trying to stay right to allow passing. In my experience, streets with very skinny lanes often feel safer because there is no question that I should take the full lane and that anyone looking to pass should change lanes to do so. It's those streets with awkwardly wide lanes that are the worst because there isn't really room for safe passing, but it sorts of looks like there might be. This leads drivers to try to squeeze by, which can lead to scary situations. What I'm saying is that there's a lot of nuance to account for in a larger deployment of this tech.

The sensors also only measure a specific type of danger: Sidewipe and rear-end collisions. Turning collisions and parked car door collisions are major safety factors that would likely not be picked up at all. So ProxiCycle data is more appropriate as part of a larger analysis that includes additional considerations.

In some ways, Seattle is a great place to test this tech because it has a wealth of safety data from SDOT's excellent Vision Zero team that ProxiCycle can use to ground-truth their readings. But the best use cases for this data might actually be in places where no agency has done the kind of safety analysis work that SDOT has done. It's may be one of the cheapest ways to get a quick and dirty map of dangerous cycling conditions, and a community organization could take the lead on the effort rather than waiting for, say, a reluctant DOT or engineering department.

Below is a short presentation the team made to a conference in April:

And here's is an even shorter UW video about it (full press release):