Bicycle fatalities persist as a public health challenge even as cities invest in cycling infrastructure and autonomous vehicles complicate street dynamics. Traditional approaches to understanding cyclist risk, relying on aggregate crash data or complex detection systems, fail to capture the granular safety variations that riders experience across different routes and conditions.
A new research project tackles this measurement gap through an architecture that turns handlebar camera footage into actionable safety insights. According to arXiv, computer scientists from multiple institutions have developed an edge-assisted crowdsensing platform that processes real-time street imagery to generate numerical safety assessments between 1 and 9.
How the System Works
The pipeline operates in three stages. First, cameras mounted on handlebars capture street-level video. A vision-language model then processes each frame through a single natural language prompt, extracting structured attributes about traffic conditions, infrastructure, and environmental factors. Finally, an explainable boosting machine combines these visual features with road data from OpenStreetMap and motion sensors from the bicycle itself to produce a safety score.
The critical innovation lies in computation placement. Rather than sending video to cloud servers, attribute extraction happens at the network edge using locally-hosted models. The fastest approach requires 2.49 seconds per frame, while scoring itself takes only 0.4 milliseconds on consumer CPU hardware. This design reduces bandwidth requirements and latency dramatically.
Performance and Interpretability
Testing on 1,362 hand-annotated frames from two London cycling routes, the explainable boosting model achieved 72.2 percent binary accuracy with a mean absolute error of 1.011 point across five-fold cross-validation. Importantly, performance remained within one accuracy point of black-box ensemble methods, suggesting that the tradeoff for interpretability is modest.
The interpretability advantage proves substantial for urban planners. The model generates precise attribution scores for each infrastructure feature. Physical lane separation contributes plus 0.97 rating points to perceived safety, while painted-only lanes add plus 0.70. Construction sites subtract 0.65 points. These quantified impacts enable planners to calculate infrastructure changes directly into cost-benefit analyses rather than relying on abstract safety metrics.
Scaling to Continuous Monitoring
A sensitivity analysis confirms the system tolerates typical data quality variations. With frame re-encoding optimizations, the researchers estimate that a 20-second sampling interval can serve 1,000 simultaneous riders using only 54 megabits per second of bandwidth, incurring just 1.4 accuracy points of degradation.
Researchers also documented inter-annotator agreement metrics (alpha equals 0.4695) and conducted attribute audits to validate that the extracted features correspond to real-world conditions rather than model artifacts.
The work suggests a path toward continuous, scalable safety monitoring that respects privacy by processing data locally rather than centralizing video streams. As cities expand cycling networks and grapple with shared roadspace between bikes, cars, and autonomous vehicles, tools that quantify perceived safety at route level could inform infrastructure priorities more directly than retrospective crash analysis alone.



