TL;DR
This paper introduces an edge-assisted crowdsensing system that uses Vision-Language Models (VLMs) to translate cyclist handlebar video into structured traffic data, which is then scored for perceived safety using an Explainable Boosting Machine (EBM). By offloading heavy VLM inference to the network edge and using interpretable models, the system provides city planners with a scalable, transparent way to "price" infrastructure improvements in terms of rider safety points.
What Problem Does This Paper Address?
Cyclist safety is usually measured through crash statistics, but these are reactive and don't capture perceived safety—how safe a rider feels. Perceived safety is a primary driver of whether people choose to cycle, yet it's notoriously difficult to measure at scale.
Current approaches either rely on expensive, non-scalable surveys or "black-box" computer vision pipelines that don't explain why a scene is rated a certain way. Furthermore, streaming high-resolution video from thousands of bikes to a central server for analysis is a bandwidth and privacy nightmare. This paper seeks a middle ground: a system that is scalable, privacy-conscious, and provides actionable, explainable insights for urban planners.
Methods
The authors propose a three-stage pipeline:
- Data Preparation: Handlebar-mounted cameras (GoPros) capture street-level frames, which are fused with OpenStreetMap (OSM) infrastructure data and on-bike kinematics (speed and acceleration).
- Traffic Description (The Edge Stage): Instead of sending raw video to the cloud, the system uses a VLM at the network edge. A single prompt (e.g., to Grok-2-Vision or Qwen3-VL) transforms a frame into a compact JSON vector of attributes like
is_cycling_lane_presentoris_high_traffic_density. This "Mobile Edge Generation" (MEG) approach drastically reduces bandwidth. - Safety Modelling: An Explainable Boosting Machine (EBM) takes these attributes and predicts a safety score (1–9). EBMs are a type of Generalized Additive Model that uses gradient boosting to learn "shape functions" for each feature. This makes the model's logic entirely transparent.
Datasets
- London Cyclist Dataset: 1,362 GPS-tagged frames collected from two routes in London (Route A near Queen Mary University and Route B in Farringdon).
- Crowdsourced Annotations: 7,045 ratings on a 1–9 scale provided by 99 regular cyclists via a custom platform.
- OSM Integration: Frames were map-matched to the OpenStreetMap road network to include context like road class.
Key Insights and Ideas
- VLMs as "Swiss Army Knife" Extractors: The authors show that a single VLM prompt can replace a complex stack of specialized object detection models. However, they also perform an "attribute audit," finding that while VLMs are great at identifying bike lanes, they tend to over-report intersections.
- Interpretability Without Sacrifice: The EBM matched the accuracy of black-box models like Random Forests and XGBoost within one percentage point. This proves that for this domain, you don't need to sacrifice transparency for performance.
- Quantifying Infrastructure: Because the EBM is additive, planners can see exactly how much a feature is "worth." For example, physical lane separation adds +0.97 to the safety rating, while a construction site subtracts -0.65.
- The 20-Second Sweet Spot: The researchers found that sampling one frame every 20 seconds (and re-encoding it) allows a fleet of 1,000 riders to be served at 54 Mbit/s with only a marginal (1.4 point) drop in accuracy.
Key Results
- Performance: The EBM regressor achieved 72.2% binary accuracy (safe vs. unsafe) and a Mean Absolute Error (MAE) of 1.011 on the 1–9 scale.
- Efficiency: Scoring a frame with the EBM takes just 0.4 ms on a CPU, making it suitable for deployment even on low-power devices.
- Edge Feasibility: The fastest VLM configuration (InternVL3 8B) processed frames in 2.49 seconds on an A100 GPU, fitting within a 16.1 GB VRAM footprint.
- Bandwidth: Re-encoding images to 1024px and using a 20s interval reduced the data payload by over 90x compared to raw 5s capture.
Limitations and Future Work
- Geographic Scope: The study is limited to two routes in London. Cross-city generalization remains a key hurdle for future research.
- Temporal Context: The current VLM processes frames independently. The authors suggest that "contrastive prompting" across consecutive frames could help the model understand the flow of traffic better.
- Rater Diversity: The current model reflects the judgments of frequent cyclists. Expanding the annotator pool to include occasional riders or non-cyclists would provide a more inclusive view of perceived safety.
Summary updated