Page 84 - JCAU-7-1

Journal of Chinese
Architecture and Urbanism Machine-simulated scoring of child-friendly streets

to interval distance and time-of-day variations. However, to extract the pixels of streetscape features (Figure 5). This
a consensus has emerged that the methodology produces approach enhanced scene comprehension at both global and
reliable results when the study area is sufficiently large (Kim local scales by up-sampling and concatenating features across
et al., 2021). For example, some previous studies have used various scales, allowing efficient semantic segmentation by
50 m intervals for street greenery research (Lu et al., 2019; Ye accurately categorizing each pixel through the classification
et al., 2019), while others applied sampling points every 50 layer. PSPNet was trained on the ADE20K dataset, which
and 100 m (Law et al., 2020; Yang et al., 2019). In this study, can segment 150 object classes. For this study, we focused on
we selected a 30 m interval to capture SVIs, using Python 10 object classes of streetscape features relevant to children
in conjunction with the Google Maps platform and Google commuting to school, to evaluate streetscape features
API to capture parameters such as geolocation coordinates, contributing to the concept of “child-friendly cities.”
orientation, perspective, and field of view. Each sample point To evaluate PSPNet’s performance on the output
included four images facing south, north, east, and west. dataset, we used the pixel accuracy metric, the simplest
We used detailed street mapping data obtained from classification accuracy measure that calculates the
the Hong Kong GeoData website, which required street percentage of correctly classified pixels in the image.
consolidation, simplification, and topological treatment For each row (each SVI), we calculated the sum of the
of road networks to streamline the SVI data collection correctly classified pixel proportions across all categories:
process. For this study, we selected 29 living streets in
the central area of Sham Shui Po (Figure 4) and excluded Total Accuracy _{ } i = ∑ _{j =1}^{ } _{ }O p ij (I)
expressways and elevated roads to focus on public streets where p_{ij} represents the proportion of correctly
that foster connectivity and socializing for local residents.
classified pixels for category j in row i, and O is the total
To ensure quality in SVI extraction, precise parameters number of object classes.
for perspective and dimensions were set. Python’s urllib Next, we calculated the average of the total accuracies
module was used to download SVIs with strategic overlaps, across all rows (SVIs):
enabling four images from different angles to be merged into
a single panoramic view. Tools like OpenCV and Numpy Pixel Accuracy
facilitated image integration, enhancing the efficiency of = (1 / ) N ∑ _{i =1}^{ } Total Accuracy _{ }N i (II)
the image segmentation processes. Our project adopted the
semantic segmentation method PSPNet (Zhao et al., 2017) where N is the total number of images.

Figure 4. Site location, street segmentation, and example of the street view image collection process. Source: Drawing by the authors

Volume 7 Issue 1 (2025) 7 https://doi.org/10.36922/jcau.3578