Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception

Hongsen Cao; Chia-Yen Chiang; Mona Jaber; Meihui Wang; Yasmin Fathy

cs.CVNov 5, 2026

Edge-Assisted Crowdsensing for Explainable Cyclist Safety Perception

Authors: Hongsen Cao, Chia-Yen Chiang, Mona Jaber, Meihui Wang, Yasmin Fathy

Organizations: School of EECS, Queen Mary University of London, London, UK · School of Public Health, Imperial College London, London, UK · Department of Engineering, University of Cambridge, Cambridge, UK

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-model detection pipelines scale to continuous monitoring of perceived safety along a route. We 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, and an explainable boosting machine (EBM) combines them with OpenStreetMap road infrastructure and on-bike kinematics to predict a crowdsourced 1-9 safety score. Extraction belongs at the network edge: the fastest self-hosted model needs 2.49 s per frame and scoring takes 0.4 ms on a CPU. On 1,362 annotated frames from two London routes, the EBM regressor reaches 72.2 percent binary accuracy and a mean absolute error of 1.011 under five-fold cross-validation. It is within one accuracy point of black-box ensembles. Its per-feature shape functions give exact attribution: physical lane separation is worth +0.97 rating points, a painted lane +0.70 and a construction site -0.65. Planners can therefore price an infrastructure change in rating points, and the results enter cost-benefit analysis directly. We also report inter-annotator agreement (alpha = 0.4695), an attribute audit, and 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 1.4 accuracy points, making continuous monitoring practical.

Figures & tables

Fig. 1: Three-stage pipeline: data preparation, VLM-based traffic description at the network edge, and EBM-based safety modelling with interpretable analysis.

Attribute	Claude Sonnet	GPT-4o	Grok-2-Vision	Baseline
is_cycling_lane_present	68.9	55.8	83.6	67.0
is_cycling_lane_separation_present	83.8	55.0	86.3	80.4
is_intersection	38.6	63.4	53.5	85.6
is_construction_site_present	59.9	57.6	58.2	—
immediate_crash_risk_from_traffic	50.4	60.8	53.9	—
is_high_traffic_density	62.3	67.1	64.9	—

TABLE I: VLM comparison on human-audited attributes (% accuracy) and per-token price, with the always-absent baseline for the three attributes that carry corpus-wide labels.

Fig. 2: Mean safety score for each combination of cycling infrastructure (column groups: cycling lane present, cycling lane separation) and adverse condition (row panels: construction site, crash risk from traffic, intersection). Each cell gives the mean score and the proportion of observations in brackets. The no-infrastructure, no-hazard cell of each column group is the baseline, and every other cell adds its signed change from it.

Interpretable
Model	Acc	F1(W)	Rec(U)	Rec(S)	MAE
EBM regressor (ours)	0.722	0.719	0.815	0.623	1.011
EBM classifier (ours)	0.711	0.710	0.799	0.633	—
Logistic regression	0.665	0.664	0.761	0.580	—
Decision tree	0.669	0.662	0.836	0.520	0.999
Black-box reference

TABLE II: Model comparison for binary safety classification under random five-fold cross-validation. Differences of about one point at this sample size do not establish an ordering; see Section IV-C .

Fig. 3: Global feature importance from the EBM, as the mean absolute weighted score. Every main-effect term outranks every pairwise interaction term.

Model	Lat. (s)	VRAM	Frames/GPU h	Lane / Sep.
Qwen3-VL 4B	2.74	9.2 GB	1312	0.854 / 0.816
Qwen3-VL 8B	2.99	17.8 GB	1205	0.819 / 0.816
Qwen3-VL 32B	8.94	67.2 GB	403	0.767 / 0.822
InternVL3 8B	2.49	16.1 GB	1444	0.877 / 0.825
Llama-3.2 11B	3.63	22.3 GB	992	0.796 / 0.793
Lat.: mean latency per frame. VRAM: peak GPU memory.

TABLE III: Self-hosted extraction on a 309-frame stratified subset (A100-80GB, bf16, batch 1, 1024 px long edge). Accuracies are on cycling lane and lane separation, the two attributes the audit retains.

Fig. 4: Sampling interval against accuracy under spatial blocking, with the uplink each interval implies relative to 5 s capture on the same axis. Shading marks the fold-to-fold spread of the 5 s baseline. The size-matched control keeps the 5 s cadence while using the same number of frames as the thinned setting. Bars span folds.

Interpretable ML

Interpretable machine learning develops models and explanations that make predictions, evidence, and decision processes understandable to people.

See all 316 papers

Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction
May 21, 2026Luís Maria Perdigão, Miguel Costa, Carlos Santiago +1Learning to Rank
Institute for Systems and Robotics (ISR-Lisboa), Instituto Superior Técnico, Lisbon, Portugal. · Technical University of Denmark, Kongens Lyngby, Denmark.
Caption-Mediated Perceived-Safety Estimation for Pedestrian Routing
Sep 29, 2026Simon Parkinson, Paloma Liu, Wei Zheng +1VLM Evaluation Explainable Artificial Intelligence
School of Computing and Mathematics, Manchester Metropolitan University, Manchester M1 5GD, U.K. · University of Huddersfield, Huddersfield HD1 3DH, U.K. · The University of Manchester, Manchester M13 9PL, U.K.
BikeScenes: LiDAR Semantic Segmentation for Bicycles
Oct 29, 2025Denniz Goren, Holger CaesarAutonomous Driving Perception 3D Semantic Segmentation

May 21, 2026J/K move · Enter open · S save

Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction

Luís Maria Perdigão, Miguel Costa, Carlos Santiago, Manuel Marques

Institute for Systems and Robotics (ISR-Lisboa), Instituto Superior Técnico, Lisbon, Portugal. · Technical University of Denmark, Kongens Lyngby, Denmark.

Learning to Rank

Open paper AskPDF arXiv

Cycling delivers significant public-health and environmental benefits, yet its uptake in cities is often limited by perceived safety. When street environments appear unsafe, individuals are less likely to cycle, making perception a key barrier to adoption. Recent work has shown that pairwise comparisons of street-view images provide a scalable way to learn subjective safety judgments. However, existing approaches do not explicitly model human visual attention, which plays a central role in how humans perceive safety. We propose an Eye-Tracking-Guided Perceived Cycling Safety framework (EG-PCS) that integrates gaze data into a pairwise learning pipeline based on vision transformers. By supervising the model's attention mechanism with eye-tracking signals, we encourage alignment between learned attention maps and human fixation patterns. Experiments show that gaze-guided models achieve similar ranking performance compared to state-of-the-art approaches while producing attention maps that more accurately reflect human visual attention behavior. Our results demonstrate that incorporating eye-tracking information enhances both predictive accuracy and interpretability in perception-based urban analytics.

Abstract

Interpretable ML

Explore similar work

Learning to See Like Humans: Gaze-Aligned Cycling Safety Prediction

Caption-Mediated Perceived-Safety Estimation for Pedestrian Routing

BikeScenes: LiDAR Semantic Segmentation for Bicycles