BikeButler: A Personalized, Context-sensitive Bike Routing Tool using Open Data and VLM-based Analyses of Street View Imagery

Paper Title

BikeButler: A Personalized, Context-sensitive Bike Routing Tool using Open Data and VLM-based Analyses of Street View Imagery

Publication Info

• Topic area: Personalized bike routing using context-sensitive algorithms and multimodal data.
• Keywords: Bike routing, personalization, context-sensitive, Vision Language Model, OpenStreetMap, Street View imagery, urban cycling, bikeability, GIS data.

Background and Problem

• Problem / challenge: Existing bike routing tools like Google Maps and Strava lack dynamic context sensitivity, fine-grained customization, and transparency in route recommendations. They fail to account for subjective bikeability factors and dynamic contexts such as commuting versus recreational biking.
• Significance: Urban cycling promotes sustainability, health, and well-being. Tools that improve route personalization and transparency can encourage cycling adoption and enhance user experience.
• Motivation and related work: Prior research has explored bikeability indices, infrastructure preferences, and computer vision for urban analysis. However, gaps remain in integrating subjective and objective bikeability factors into customizable, interactive tools. Existing tools like Cyclopath and RideWithGPS provide limited interactivity and lack context-aware routing.

Solution

• Proposed approach: BikeButler, a personalized, context-sensitive bike routing tool that combines OpenStreetMap (OSM), open government data, and Vision Language Model (VLM)-based analyses of Street View imagery (SVI) to generate and customize bike routes based on user-defined preferences.
• Novelty:
  1. Integration of multimodal data sources (OSM, government elevation data, VLM-based SVI analysis) to assess bikeability across eight factors.
  2. Interactive route generation and refinement using custom profiles and segment-based voting mechanisms.
  3. Algorithmic transparency through detailed visualizations, including SVI previews, elevation charts, and color-coded bikeability scores.
• Procedure and key techniques:
  • Users create bikeability profiles with sliders for preferences like bike lanes, slope, vegetation, and surface quality.
  • Routes are generated using a preference-weighted routing algorithm based on Dijkstra’s algorithm, scoring street edges for bikeability.
  • Routes are segmented and visualized with color-coded scores, SVI previews, and elevation charts.
  • Users iteratively refine routes by adjusting profiles or voting on segments.

Results

• Concrete findings:
  • BikeButler produced routes with lower speed limits, better vegetation, and smoother surfaces compared to Google Maps routes (e.g., 11% better vegetation in Scenario 1, 30% in Scenario 2).
  • Participants created 187 routes during the study, averaging 12.5 routes per participant.
  • VLM achieved human-level performance in scoring vegetation and surface quality but underperformed in detecting bike lanes and estimating widths.
• Advantage over baselines: BikeButler routes were more aligned with user preferences, offering greater customization and transparency compared to Google Maps. Participants selected routes with fewer bike lanes but more residential streets and better overall bikeability.
• Experiments / evaluation:
  • User study with 16 participants exploring three tasks: two scenario-based and one open-ended.
  • Technical evaluation of VLM performance compared to human labelers for vegetation, surface quality, and bike lane detection.
  • Comparison of BikeButler routes to Google Maps routes for the same origin-destination pairs.
• Limitations and future work:
  • Limited to a single city (Seattle) and small sample size (16 participants).
  • Reliance on high-quality data from OSM, government sources, and GSV.
  • Need for deployment studies to evaluate real-world usage and biking outcomes.
  • Future work includes expanding to other cities, improving VLM accuracy, and integrating live GPS navigation.

Summary

BikeButler introduces a personalized, context-sensitive bike routing tool that leverages multimodal data sources to generate and customize routes based on user-defined preferences. The tool demonstrated significant advantages over existing commercial tools like Google Maps, producing routes with better alignment to user priorities such as safety, comfort, and bikeability. A user study highlighted the tool’s ability to support iterative route refinement and enhance transparency through visualizations like SVI previews and elevation charts. While limited to Seattle and reliant on high-quality data, BikeButler showcases the feasibility of context-aware bike routing and provides a foundation for future research and deployment at scale.