Hardware Analysis: E-Scooter Safety

Client: Transloc/Ford Motor Company, in partnership with King County Transit & Cornell University

Role: Ford PI

Research Question

Shared e-scooters were expanding rapidly across U.S. cities, but city and campus stakeholders had legitimate safety concerns and little real evidence to act on. Existing safety research relied on hospital injury records, which describe outcomes but not the riding conditions or behaviors that caused them.

Spin/Ford needed to understand what actually happens in the moments before an e-scooter crash: which infrastructure, rider behaviors, and environmental conditions raise real-world risk, and by how much?

My Role

As Ford PI, I contributed to overall study design and led the rider survey element. I contributed to all data interpretation and co-authored the resulting papers. VTTI led hardware data collection and video/data reduction. Spin operated the e-scooter fleet.

Approach

The study paired a rider survey with hardware data collection using a naturalistic method, a technique from large-scale vehicle safety research where data is collected passively under real riding conditions instead of in a controlled test. VTTI's team coded every safety-critical event, plus a randomly sampled set of 1,601 control segments, against a shared protocol covering infrastructure, behavioral, and environmental risk factors. We then used a case-cohort design to calculate odds ratios, comparing how often each risk factor showed up in crash/near-crash events versus normal riding conditions. This isolated which factors actually drove risk, not just which ones happened to be present.

Data Sources

Ford partnered with VTTI and Spin to instrument 50 e-scooters in a 200-scooter campus fleet with VTTI's proprietary microDAS, an onboard system built specifically for this study. The six-month deployment produced 3,500 hours of data across 8,500+ trips, the largest naturalistic e-scooter safety dataset collected to date. Data collection methods included:

Forward-facing video, capturing riding behavior and road interactions without filming the rider, to protect privacy
GPS location and speed data
Three-axis accelerometer data, capturing vibration and impact events
Algorithmic detection of over 2,000 candidate safety-critical events (crashes and near-crashes), each confirmed through manual video review

What We Found About Users

Infrastructure was the dominant risk factor. Two-thirds of all safety-critical events were precipitated by infrastructure conditions, more than double the combined share attributed to other road users or rider behavior.

Surface type mattered enormously. Riding on grass carried 28 times the risk of riding on asphalt or concrete; loose surfaces like gravel or dirt carried a similar 28-times increase. Even routine sidewalk-to-roadway curb cutouts more than doubled risk.

Transitions between surfaces were riskier than the surfaces themselves. Moving between pavement and grass increased risk 38-fold. Moving between gravel/dirt and grass increased it nearly 60-fold.

Risky rider behavior sharply increased danger. Aggressive or trick riding increased crash risk nearly tenfold. Riding in a group nearly doubled it.

Low light was a meaningful, if less dominant, factor. Riding outside full daylight conditions was associated with 5.5 times higher risk, despite the deployment's operating hours being restricted to daylight for most of the study.

No vehicle collisions occurred in the dataset. Conflicts with pedestrians, other scooters, and bicycles accounted for a meaningful share of near-crashes, most of which were successfully avoided through evasive action.

Decisions this work informed

The findings gave Virginia Tech, municipalities, and e-scooter operators quantified guidance instead of general safety intuition. They supported specific recommendations: prioritize paved, dedicated e-scooter infrastructure over shared-use paths and off-road terrain, invest in surface maintenance given how much risk degraded pavement introduced, and target educational outreach at the specific behaviors and surface transitions the data showed mattered most.

Separately, and outside the scope of this study's own testing, Spin used related insights from the deployment to inform later product features, including an in-app drunk-riding detection capability and speed throttling for a rider's first few trips.

Limitations

The study took place on a pedestrian-dense university campus with e-scooter speeds governed to 12 mph (4 mph in high-pedestrian zones) and no permitted riding after dark for most of the deployment. Vehicle traffic exposure was lower than a typical urban deployment, which likely explains the complete absence of vehicle collisions in the dataset. The infrastructure and behavioral findings are well-supported, but their magnitude, and the absence of any vehicle-conflict risk, may not generalize directly to denser urban environments with mixed traffic.