Summary
This paper presents an edge-assisted crowdsensing system for continuously monitoring cyclists' perceived safety along routes. Handlebar-mounted cameras capture street-level frames, a vision-language model (VLM) converts each frame into structured traffic attributes via a single prompt, and an explainable boosting machine (EBM) fuses these attributes with OpenStreetMap road infrastructure data and on-bike kinematics to predict a crowdsourced 1-9 safety score. Computation is placed at the network edge: the fastest self-hosted model requires 2.49 s per frame while scoring takes only 0.4 ms on CPU. Evaluated on 1,362 annotated frames from two London routes under five-fold cross-validation, the EBM regressor achieves 72.2% binary accuracy and a mean absolute error of 1.011, within one accuracy point of black-box ensembles. Per-feature shape functions provide exact attribution: physical lane separation contributes +0.97 rating points, painted lanes +0.70, and construction sites -0.65, enabling planners to price infrastructure changes directly in rating points for cost-benefit analysis. The authors also report inter-annotator agreement (alpha = 0.4695), an attribute audit, and sensitivity analysis. With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s with only 1.4 accuracy points lost, making continuous monitoring practical.
Paper Overview
Research Area: CV
Authors: Hongsen Cao, Chia-Yen Chiang, Mona Jaber, Meihui Wang, Yasmin Fathy
Published: 2026-11-05
arXiv: 2611.08441
Abstract
Cyclist fatality rates remain high despite policy support for active travel, and autonomous vehicles add complexity to mixed traffic. Neither crash statistics nor multi-modal detection pipelines scale to continuous monitoring of perceived safety along a route.
The authors present an edge-assisted crowdsensing system:
- Handlebar cameras supply street-level frames
- A vision-language model (VLM) turns each frame into structured traffic attributes through a single prompt
- An explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1-9 safety score
Why the edge? Extraction belongs at the network edge: the fastest self-hosted model needs 2.49 s per frame, while scoring takes 0.4 ms on a CPU.
Results
On 1,362 annotated frames from two London routes, the EBM regressor reaches:
- 72.2% binary accuracy and a mean absolute error of 1.011 under five-fold cross-validation
- Within one accuracy point of black-box ensembles
Its per-feature shape functions give exact attribution:
| Feature | Effect on rating |
|---|---|
| Physical lane separation | +0.97 rating points |
| Painted lane | +0.70 rating points |
| Construction site | -0.65 rating points |
Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly.
Validation and Scalability
The paper also reports:
- Inter-annotator agreement (alpha = 0.4695)
- An attribute audit
- A sensitivity analysis linking attribute error to accuracy
With re-encoding, a 20 s sampling interval serves 1,000 riders at 54 Mbit/s at a cost of only 1.4 accuracy points, making continuous monitoring practical.
---
*Source: arXiv 2611.08441, auto-collected 2026-11-08.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636942