PaperlayerView exact revision on arXiv

Source-linked AI summary

Assessing bikeability with street view imagery and computer vision

Koichi Ito, Filip Biljecki

arXiv:2105.08499v3cs.CV

TL;DR

Existing bikeability assessments often rely on conventional or labor-intensive methods, while SVI-based studies have covered limited aspects and have not thoroughly tested CV and SVI together. This paper develops and evaluates a 34-indicator index in Singapore and Tokyo, finding that SVI indicators explain substantially more of the estimated index than non-SVI indicators, while limitations support combining both approaches.

  • Problem

    Prior studies had not developed a comprehensive bikeability index using SVI as a major data source or thoroughly assessed whether SVI and CV could replace traditional techniques.

  • Method

    The study evaluates 34 indicators across five categories using SVI, surveys, OSM, land use, elevation, and air-quality data in Singapore and Tokyo.

  • Results

    An estimated bikeability index was developed, and SVI indicators had an R^2 of 0.85, outperforming non-SVI indicators.

  • Takeaways & Limitations

    SVI and CV can comprehensively assess bikeability within and among cities, while SVI indicators show potential for independent assessment.

  • Takeaways & Limitations

    Random sampling intervals for GSV panoramas may introduce bias and larger data variance, while the index could benefit from broader indicators and improved preference-based weights.

Abstract

from arXiv · show

Studies evaluating bikeability usually compute spatial indicators shaping cycling conditions and conflate them in a quantitative index. Much research involves site visits or conventional geospatial approaches, and few studies have leveraged street view imagery (SVI) for conducting virtual audits. These have assessed a limited range of aspects, and not all have been automated using computer vision (CV). Furthermore, studies have not yet zeroed in on gauging the usability of these technologies thoroughly. We investigate, with experiments at a fine spatial scale and across multiple geographies (Singapore and Tokyo), whether we can use SVI and CV to assess bikeability comprehensively. Extending related work, we develop an exhaustive index of bikeability composed of 34 indicators. The results suggest that SVI and CV are adequate to evaluate bikeability in cities comprehensively. As they outperformed non-SVI counterparts by a wide margin, SVI indicators are also found to be superior in assessing urban bikeability, and potentially can be used independently, replacing traditional techniques. However, the paper exposes some limitations, suggesting that the best way forward is combining both SVI and non-SVI approaches. The new bikeability index presents a contribution in transportation and urban analytics, and it is scalable to assess cycling appeal widely.

1. Introduction

Bikeability research seeks to quantify how cycling is facilitated, but conventional approaches face scalability and coverage limitations. This study tests whether SVI and CV can support comprehensive, fine-scale assessment across cities.

  • Bikeability indexes quantify the extent to which urban conditions facilitate cycling.
  • CV and SVI are evaluated as potentially scalable alternatives for assessing bikeability within and among cities.
  • The study develops a bikeability index with 34 indicators under five categories and implements it in Singapore and Tokyo.
  • The paper combines a literature review, methodological development, and empirical comparison to investigate whether SVI and CV can replace traditional techniques.

2. Related work

Previous bikeability studies increasingly use scalable data sources, but still struggle to combine objective and subjective indicators, capture street-level conditions, and automate comprehensive assessment. SVI and CV address parts of these challenges, yet comprehensive SVI-based bikeability indexing remains underdeveloped.

  • Earlier bikeability studies relied heavily on field surveys, making assessment time- and resource-intensive and difficult to scale.
  • Later studies improved scalability through remote sensing, SVI, and crowdsourcing, but often excluded subjective indicators or retained labor-intensive collection.
  • The literature identifies recurring challenges involving data-collection costs, subjectivity and objectivity, street-level information, and spatial-granularity standardization.
  • CV and SVI support scalable extraction of greenery, infrastructure, traffic, and perception information relevant to bikeability.
  • Only one cited study created an SVI-based bikeability index, and it covered limited aspects such as greenery and enclosure.
  • This paper addresses the gap by developing a comprehensive SVI- and CV-based index while examining their value, independence, scalability, and application barriers.

3. Methodology

The study builds a comprehensive bikeability index from six data sources and 34 indicators across five categories, combining SVI-derived and conventional geospatial measures. It samples Singapore and Tokyo, models perception from SVI features, and aggregates indicators using established scaling and weighting procedures.

  • Index design: Six data sources support 34 indicators across connectivity, environment, infrastructure, perception, and vehicle-cyclist interaction.The sources are SVI, surveys, OSM, land use, DEM, and AQI.
  • Index design: The indicator inventory was compiled from previous bikeability indexes to expand related work and reduce selection bias.
  • Data sources: 21 indicators come from SVI, 10 from OSM, and one each from land use, DEM, and AQI.
  • Sampling: OSM supplied approximately 250,000 Singapore points and 440,000 Tokyo points, from which 7,142 locations were randomly selected for SVI retrieval.The sample size matched the maximum number of Google Street View images available through the monthly free API credit.
  • Feature extraction: Features were aggregated within 500m buffers for contextual densities and vehicle counts, while selected road and pavement indicators used 100m buffers.The 500m vehicle aggregation addressed the unreliability of estimating traffic from a single street-view image.
  • Perception modeling: Perception indicators were predicted from high- and low-level SVI image features using surveyed ratings and LightGBM models.For each city, 400 images were surveyed, producing 800 rated images overall; the data were split 80:20 for training and validation, with 10-fold cross-validation for tuning.
  • Index construction: The composite index conflates previous bikeability indexes and applies category-and-indicator weighting after min-max scaling.The study considers independent, arbitrary, and equal weighting systems from prior work.

4.1. Data collection

The study assembled street-network, point-of-interest, land-use, street-view, and air-quality data for Singapore and Tokyo, harmonizing land-use categories across the two cities.

  • 252,369 Singapore and 450,379 Tokyo street segments were retrieved, alongside city-specific POIs and mass rail transit stations.
  • 5,833 Singapore and 6,181 Tokyo panorama images remained after removing indoor and grey images from 7,142 images.
  • Land-use data were harmonized into residential, commercial, and industrial categories, with other categories excluded.
  • 2020 air-quality data came from 5 Singapore stations and 126 Tokyo stations.

4.2. Extracted indicators and composite index

The composite index combines five bikeability dimensions and integrates objective indicators with survey-based perception measures and computer-vision modeling. Results show contrasting city profiles, while perception prediction and some category distributions constrain reliability.

  • Connectivity: Connectivity used intersections with and without lights plus cul-de-sacs, with Tokyo scoring higher overall than Singapore.Singapore scored much lower on intersections without traffic lights, while the other two indicators had similar distributions.
  • Environment: The environment category combined slope, POIs, land-use mix, AQI, and greenery, building, and water pixel ratios, with Tokyo achieving the higher mean score.Tokyo scored higher for slope, land use, and building pixels; Singapore scored higher in AQI and greenery.
  • Infrastructure: Infrastructure covered road and pavement type, road width, transit facilities, potholes, lighting, cycling facilities, amenities, sidewalks, crosswalks, and curb cuts, with Singapore scoring much higher.Singapore scored higher for pavement, street amenities, and especially utility poles, where Tokyo showed the opposite pattern.
  • Vehicle–cyclist interaction: Vehicle–cyclist interaction used vehicle counts, speed-control devices, on-street parking, and traffic lights or stop signs, producing similar city means with a slightly higher Singapore mean.The category treats fewer and slower traffic as safer for cyclists, and its indicators showed very small variances.
  • Perception: Perception covered seven survey dimensions, with scores strongly correlated at squared R values from 0.58 to 0.79 and generally higher Singapore ratings.Eight unique participants rated each of 800 surveyed images on a 0–10 scale; most responses were between three and eight.
  • Perception: Image features linked beauty, building attractiveness, cleanliness, cycling attractiveness, and spaciousness to terrain, greenery, land use, curb conditions, utility poles, junkyards, and building density.The exploratory analysis found more influential image-classification features than other feature types and did not reproduce all prior findings.
  • Perception: Perception models achieved MAE around 0.65, MAPE around 0.1, and RMSE around 0.8, but every R2 value was below 0.The authors therefore judged the models worse than predicting constant values regardless of input data.
  • Bikeability: Singapore’s bikeability scores were generally homogeneous, whereas Tokyo’s were more heterogeneous; Singapore also had a slightly higher mean and lower standard deviation.

4.3. Comparison between SVI and Non-SVI indicators

The study compares comprehensive, SVI-only, and non-SVI-only indexes to assess whether street-view indicators can substitute for conventional data. SVI-only scores closely tracked the comprehensive index, but several indicators remain better obtained from non-SVI sources.

  • The comparison developed comprehensive, SVI-only, and non-SVI-only indexes and plotted their scores at the same sample points.Connectivity and perception were excluded because each contained only one indicator type.
  • SVI indicators correlated more strongly with the combined bikeability index than non-SVI indicators, with R2 of 0.85 and lower kurtosis.The authors note that the stronger relationship partly reflects the larger number of SVI indicators.
  • Transit facilities, POIs, land uses, slope, and air quality remain difficult, inefficient, or less accurate to estimate from SVI.These measures are more straightforward or accurate to collect from OSM and other established data sources.
  • The authors conclude that combining SVI and non-SVI indicators is more beneficial for assessing bikeability.

5. Challenges and future directions

The study identifies data-quality, sampling, accessibility, and modeling challenges that constrain bikeability assessment, while outlining practical directions for improving data coverage, tooling, and index design.

  • Data quality and sampling: Randomly sampled GSV panoramas may introduce bias and larger variances, especially when sampling intervals are long.Shorter sampling intervals are proposed to reduce possible bias.
  • Data quality and sampling: Vehicle-based SVI may not represent bicyclists’ typical perspectives, particularly on wide roads.The authors note that limited cycling-path coverage means panoramas may still capture much of the actual cycling perspective in Singapore and Tokyo.
  • Data quality and sampling: OSM incompleteness caused traffic-speed and traffic-lane indicators to be eliminated, motivating exploration of alternative data sources.AQI coverage also differed substantially between Singapore and Tokyo, with five stations in Singapore versus 126 in Tokyo.
  • Data quality and sampling: Extremely low indicator variances cannot differentiate sample points, so more diverse city sets are needed for further examination.The issue affects indicators whose values vary little across sampled locations.
  • Perception modeling: Perception modeling may be biased because most survey participants were not residents of the study areas.Future studies could use survey services that specify participant residence, while nonresident respondents may also reduce cross-city familiarity bias.
  • Implementation and future directions: The workflow requires moderately advanced Python knowledge and graphics-processing capabilities, suggesting GUI or API development as a mitigation.These requirements limit accessibility for users without programming expertise or suitable computational hardware.
  • Implementation and future directions: Indicator selection and weighting may need adaptation to cycling purposes, socioeconomic characteristics, cyclists’ preferences, and route-based index designs.The authors also propose expanding the range of SVI indicators and improving indicator weights.

6. Conclusion

The paper addresses the limited and insufficiently critical use of SVI and CV for scalable, comprehensive bikeability assessment. It develops and evaluates a broad index across Singapore and Tokyo, finding stronger explanatory power for SVI indicators while retaining a role for non-SVI data and identifying practical constraints on SVI-only assessment.

  • 6. Conclusion: Earlier SVI- and CV-based studies assessed limited bikeability aspects and rarely examined the technologies across multiple cities critically.This motivates a more comprehensive and comparative assessment framework.
  • 6. Conclusion: The study creates an exhaustive bikeability index using CV-extracted SVI indicators, explores automatable subjective assessment, and compares SVI with non-SVI indicators.It also investigates whether SVI indicators can independently support bikeability assessment.
  • 6. Conclusion: 0.85 R^2 was obtained for SVI indicators versus 0.4 R^2 for non-SVI indicators when explaining the estimated bikeability index.The comparison indicates substantially stronger correlation for SVI indicators, although the usefulness of non-SVI indicators should not be discounted.
  • 6. Conclusion: CV and SVI can comprehensively assess bikeability within and among cities at fine spatial and city-aggregate scales.The resulting index is presented as at least supplementing traditional instruments.
  • 6. Conclusion: SVI-derived indicators explain a large portion of overall-index variance and overshadow indicators computed through orthodox non-SVI mechanisms.The conclusion characterizes SVI indicators as substantially stronger than their non-SVI counterparts.
  • 6. Conclusion: SVI may potentially assess bikeability independently, but practical challenges and greater acquisition difficulty make an SVI-only route not always viable.The authors therefore identify further amelioration of practical issues as necessary.
  • 6. Conclusion: Future work should expand SVI indicators and improve indicator weights according to cyclists’ preferences.The authors frame these changes as ways to further enhance the bikeability index.
Loading 2105.08499v3…