Estimating pedestrian volume using Street View images: A large-scale validation test

https://doi.org/10.1016/j.compenvurbsys.2020.101481Get rights and content

Highlights

  • Pedestrian volume was assessed with street view images and machine learning.

  • Automated pedestrian detection was validated against field observation.

  • The street connectivity and pedestrian volume affects the level of accuracy.

  • The quality, size, and collection time of street view image also affects accuracy.

Abstract

Pedestrian volume is an important indicator of urban walkability and vitality. Hence, information on pedestrian volumes of different streets is indispensable for creating healthy, pedestrian-oriented cities. Pedestrian volume data have traditionally been collected through field observations, which has many methodological limitations, e.g. time-consuming, labor-intensive, and inefficient.

Assessing pedestrian volume automatically from Street View images (SVIs) with machine learning techniques can overcome such limitations because this approach offers a wide geographic reach and consistent image acquisition. Nevertheless, this new method has not been rigorously validated, and its accuracy remains unclear.

In this study, we conducted a large-scale validation test by comparing pedestrian volume extracted from SVIs with the results from field observations for more than 700 street segments in Tianjin, China. A total of 4507 sampling points along these street segments were used to collect SVIs.

The results demonstrated that using SVIs with machine learning techniques is a promising method for estimating pedestrian volumes with a large geographic reach. Automated pedestrian volume detection could achieve reasonable (Cronbach's alpha ≥0.70) or good (Cronbach's alpha ≥0.80) levels of accuracy. It is worth noting that various factors of SVIs and street segments may affect the accuracy. SVIs with higher image quality, larger image size, and collection times closer to the targeted periods produced more accurate results. The automated method also worked better in areas with high pedestrian volume and high street connectivity.

Introduction

Recent developments in urban big data and machine learning have provided us new data sources and an interdisciplinary approach to understand urban phenomena (Ruppert, 2013). For instance, in the past, information on perceived of built environment characteristics (e.g., urban greenery and aesthetics) or residents' travel behavior (e.g., walking, cycling) were often manually collected through surveys or field observations. Recent urban big data have provided effective means to collect such environmental and behavioral data on a large geographical scale. Hence, urban big data can advance urban studies, especially in the areas of walkability and healthy cities (Rzotkiewicz, Pearson, Dougherty, Shortridge, & Wilson, 2018).

Rapid global urbanization over recent decades has led to a fundamental change in people's lifestyle and a rapid expansion of urban populations. The United Nations estimates that nearly 70% of the world's population will live in cities by 2050 (United Nations, 2018). Moreover, urban residents have experienced rapid declines in physical activity levels, with roughly one-third of urban adults in the world being physically inactive (Hallal et al., 2012). Compelling evidence demonstrates that regular physical activity, such as walking and cycling, has an array of health benefits, including the reduced risk of obesity and chronic illness, and improved physiological and psychological health (I.-M. Lee et al., 2012; Sallis, Floyd, Rodriguez, & Saelens, 2012; Wang et al., 2019).

As the most common form of physical activity, walking can be easily incorporated into daily life, and it has additional environmental and social benefits, such as reducing private vehicle use, mitigating traffic congestion and air pollution, and encouraging social interaction (Giles-Corti et al., 2013; Lu, Xiao, & Ye, 2017; Nazelle et al., 2011; Yin, 2017). Many seminal urban planning theories aims to facilitate walking and pedestrian activities, by well-designed city image (Lynch, 1960), streets (Jacobs, 1961), and public open spaces (Whyte, 1980). Recent planning theories, such as smart growth and neourbanism, explicitly aim to promote walking and active living through various design strategies, such as mixed land use, compact development, well-connected streets, and the provision of pedestrian destinations (Durand, Andalib, Dunton, Wolch, & Pentz, 2011; Giles-Corti et al., 2013).

To provide sustained support for researchers and planners who aim to create walkable and healthy cities, it is necessary to assess pedestrian volume, and other walking behaviors constantly. Assessing pedestrian volumes in different streets or areas can help researchers to evaluate walkability and to discern how built environment characteristics affect walking behaviors (Ewing & Clemente, 2013). In the past, pedestrian volume has been typically collected with pedestrian counts on sites. However, field observation is inherently subject to significant limitations, such as high demand for manpower and cost, and small study areas. Street View image services from companies such as Google, Baidu, and Tencent provide high-resolution, geocoded, streetscape images in many global cities (Rzotkiewicz et al., 2018). Researchers can use Street View images (SVIs) in conjunction with machine learning techniques to extract new built environment characteristics and behavioral data (Shapiro, 2017; Ye et al., 2018; Ye, Zeng, Shen, Zhang, & Lu, 2019). For example, this automated approach has been used to assess levels of social and physical disorder (Badland, Opit, Witten, Kearns, & Mavoa, 2010; Rundle, Bader, Richards, Neckerman, & Teitler, 2011), the presence or absence of public facilities (Clarke, Ailshire, Melendez, Bader, & Morenoff, 2010; Kelly, Wilson, Baker, Miller, & Schootman, 2013; Wilson et al., 2012), and the degrees of urban greenness (Li et al., 2015; Lu, Sarkar, & Xiao, 2018). Recent research has suggested that SVIs is a promising and effective alternative for assessing pedestrian volume (Yin, Cheng, Wang, & Shao, 2015). To our knowledge, the accuracy of this new method has not been rigorously tested yet. Therefore, it remains unclear whether SVIs and machine learning techniques can offer a reliable approach to assess pedestrian volume with acceptable accuracy.

In this study, we conducted a comprehensive validation test by comparing the automated pedestrian volume extracted from SVIs with field observation for 701 street segments in Tianjin, China. More importantly, we identified the optimal parameters of SVIs and street features to achieve the highest level of agreement with field observation data. The SVIs parameters we considered included image source, collection time, and image quality. The street features included street length, pedestrian volume, and street connectivity. Our study contributed to the development of an innovative and efficient method to collect pedestrian volume data for any location covered by SVIs worldwide. More importantly, this study can help researchers and planners to create healthy cities and promote physical activities by identifying which areas have high or low pedestrian activity, and by highlighting which built environment characteristics facilitate or hinder walking in the long run.

Section snippets

The importance of walking

The last few decades have witnessed a substantial decline of engagement in physical activity by urban residents, and this decline has serious effects on the population's health (Koohsari, Badland, & Giles-Corti, 2013; I.-M. Lee et al., 2012). A global survey indicated that 31% of the world's urban population fails to meet the recommendation of 150-min moderate to vigorous physical activity per week (Hallal et al., 2012). Many developing countries such as China and India have undergone rapid

Study area

This study was conducted in Tianjin, China, one of four municipalities directly administered by the central government. It is a large city in northern China, with a population of 15.5 million in 2015 (Liu, Gao, & Wang, 2018). Tianjin remains a monocentric city, and this study focused on several parts of the downtown area (Fig. 1). The “street segment” was used as the unit of analysis, which was defined as the portion of street between two adjacent street intersections. The lengths of these

Results

The descriptive statistics are presented in Table 1. Most of the sampling street segments were covered by Tencent service in 2014. However, the coverage of Baidu Street View was more inconsistent in different years. A total of 4507 sampling points from 701 streets were used in this study. As a result of SVIs availability, 4840, 3440, 4134, and 7376 images were collected from Baidu 2013, 2015, 2016, 2017 datasets respectively. A total of 9012 images were collected from Tencent 2014 dataset.

Discussion

Previous studies have verified the feasibility of using SVIs to conduct environmental audits, such as audits of general neighborhood environments (Charreire et al., 2014; Rundle et al., 2011), street greenery (Lu, 2018), and open skies (Yin & Wang, 2016). However, most of these studies have focused on static environment features, rather than dynamic information such as pedestrian volumes, which constantly fluctuate over time and space. In this study, we sampled 701 street segments in Tianjin

Conclusion

The traditional method of assessing pedestrian volume with field observations is often time-consuming, labor-intensive, and with limited study areas. In this study, we have investigated the accuracy of an automated method for assessing pedestrian volume. This method uses Tencent and Baidu SVIs with machine learning techniques, to overcome the limitations of field observations. Our results demonstrated that overall, the new method provided acceptable or good levels of agreement with field

Declaration of Competing Interest

None.

Acknowledgements

The work described in this paper was fully supported by grants from the National Natural Science Foundation of China (No. 51778552), the Research Grants Council of the Hong Kong Special Administrative Region, China (No. CityU11666716).

References (92)

  • H.W. Kohl et al.

    The pandemic of physical inactivity: Global action for public health

    The Lancet

    (2012)
  • M.J. Koohsari et al.

    (re)designing the built environment to support physical activity: Bringing public health back into urban design and planning

    Cities

    (2013)
  • I.-M. Lee et al.

    Effect of physical inactivity on major non-communicable diseases worldwide: An analysis of burden of disease and life expectancy

    Lancet

    (2012)
  • X. Li et al.

    Assessing street-level urban greenery using Google street view and a modified green view index

    Urban Forestry & Urban Greening

    (2015)
  • Y. Lu et al.

    Urban density, diversity and design: Is more always better for walking? A study from Hong Kong

    Preventive Medicine

    (2017)
  • Y. Lu et al.

    The effect of street-level greenery on walking behavior: Evidence from Hong Kong

    Social Science & Medicine

    (2018)
  • Y. Lu et al.

    Associations between overhead-view and eye-level urban greenness and cycling behaviors

    Cities

    (2019)
  • S.W. Ng et al.

    Why have physical activity levels declined among Chinese adults? Findings from the 1991–2006 China health and nutrition surveys

    Social Science & Medicine

    (2009)
  • N. Owen et al.

    Understanding environmental influences on walking: Review and research agenda

    American Journal of Preventive Medicine

    (2004)
  • D.C. Parra et al.

    Perceived environmental correlates of physical activity for leisure and transportation in Curitiba, Brazil

    Preventive Medicine

    (2011)
  • M. Purciel et al.

    Creating and validating GIS measures of urban design for health research

    Journal of Environmental Psychology

    (2009)
  • A.G. Rundle et al.

    Using Google street view to audit neighborhood environments

    American Journal of Preventive Medicine

    (2011)
  • A. Rzotkiewicz et al.

    Systematic review of the use of Google street view in health research: Major themes, strengths, weaknesses and possibilities for future research

    Health & Place

    (2018)
  • K.R. Smith et al.

    Walkability and body mass index density, design, and new diversity measures

    American Journal of Preventive Medicine

    (2008)
  • R. Talavera-Garcia et al.

    Q-PLOS, developing an alternative walking index. A method based on urban design quality

    Cities

    (2015)
  • R. Wang et al.

    The relationship between visual enclosure for neighbourhood street walkability and elders’ mental health in China: Using street view images

    Journal of Transport & Health

    (2019)
  • J.S. Wilson et al.

    Assessing the built environment using omnidirectional imagery

    American Journal of Preventive Medicine

    (2012)
  • B. Xie et al.

    Healthy aging with parks: Association between park accessibility and the health status of older adults in urban China

    Sustainable Cities and Society

    (2018)
  • L. Yin

    Street level urban design qualities for walkability: Combining 2D and 3D GIS measures

    Computers, Environment and Urban Systems

    (2017)
  • L. Yin et al.

    Measuring visual enclosure for street walkability: Using machine learning algorithms and Google street view imagery

    Applied Geography

    (2016)
  • L. Yin et al.

    “Big data” for pedestrian volume: Exploring the use of Google street view images for pedestrian counts

    Applied Geography

    (2015)
  • S.A. Adams et al.

    The effect of social desirability and social approval on self-reports of physical activity

    American Journal of Epidemiology

    (2005)
  • D. Ahmetovic et al.

    Mind your crossings: Mining GIS imagery for crosswalk localization

    ACM Trans. Access. Comput.

    (2017)
  • Z. Asadi-Shekari et al.

    Disabled pedestrian level of service method for evaluating and promoting inclusive walking facilities on urban streets

    Journal of Transportation Engineering

    (2013)
  • C. Babb et al.

    Institutional practices and planning for walking: A focus on built environment audits

    Planning Theory & Practice

    (2015)
  • H.M. Badland et al.

    Can virtual streetscape audits reliably replace physical streetscape audits?

    Journal of Urban Health

    (2010)
  • R. Benenson et al.

    Ten years of pedestrian detection, what have we learned?

  • J.R. Bethlehem et al.

    The SPOTLIGHT virtual audit tool: A valid and reliable tool to assess obesogenic characteristics of the built environment

    International Journal of Health Geographics

    (2014)
  • R.C. Brownson et al.

    Environmental and policy approaches for promoting physical activity in the United States: A research agenda*

    Journal of Physical Activity and Health

    (2008)
  • R.C. Brownson et al.

    Measuring the built environment for physical activity: State of the science

    American Journal of Preventive Medicine

    (2009)
  • A.M. Chudyk et al.

    Agreement between virtual and in-the-field environmental audits of assisted living sites

    Journal of Aging and Physical Activity

    (2014)
  • J.W. Curtis et al.

    Using google street view for systematic observation of the built environment: Analysis of spatio-temporal instability of imagery dates

    International Journal of Health Geographics

    (2013)
  • P. Dollar et al.

    Pedestrian detection: A benchmark

  • P. Dollár et al.

    Fast feature pyramids for object detection

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2014)
  • A. Duany et al.

    Suburban nation: The rise of sprawl and the decline of the American dream

    (2000)
  • M.J. Duncan et al.

    Perceived environment and physical activity: A meta-analysis of selected environmental characteristics

    International Journal of Behavioral Nutrition and Physical Activity

    (2005)
  • Cited by (60)

    View all citing articles on Scopus
    View full text