Detecting older pedestrians and aging-friendly walkability using computer vision technology and street view imagery

Abstract

As an emerging and freely available urban big data, Street View Imagery (SVI) has proven to be a useful resource to examine various urban phenomena in human behavior, the built environment and their interactions. However, due to technical limitations, previous studies often focused on general pedestrians and ignored certain population subgroups such as older adults. In this study, we develop an innovative method for detecting older pedestrians using SVI. We adopted transfer learning to train a model which can accurately detect older pedestrians on SVI with an accuracy of 87.1%. Using Hong Kong as a case study, we created a dataset consisting of 72,689 street view panoramas and detected 7,763 older pedestrians and 29,231 non-older pedestrians. We further visualized the distribution of detected older pedestrians and found a significant spatial discrepancy between older pedestrians and residential population of older adults. To account for this spatial discrepancy, this study proposed a novel index to assess pedestrian demand and walking environment based on the ratio of the number of pedestrians and the residential population. We also found pedestrian demand assessed with this index has a stronger correlation with the built environment compared with population-level travel survey. This novel approach can be used to assess pedestrian demand for older adults, as well as aging-friendly walking environment.

Full text

Detecting older pedestrians and aging-friendly walkability using computer
vision technology and street view imagery
Dongwei Liu, Ruoyu Wang, George Grekousis, Ye Liu, Yi Lu*
This draft has been accepted by Computers, Environment and Urban Systems
Abstract:
As an emerging and freely available urban big data, Street View Imagery (SVI) has proven to
be a useful resource to examine various urban phenomena in human behavior, the built environment
and their interactions. However, due to technical limitations, previous studies often focused on
general pedestrians and ignored certain population subgroups such as older adults. In this study, we
develop an innovative method for detecting older pedestrians using SVI. We adopted transfer
learning to train a model which can accurately detect older pedestrians on SVI with an accuracy of
87.1%.
Using Hong Kong as a case study, we created a dataset consisting of 72,689 street view
panoramas and detected 7,763 older pedestrians and 29,231 non-older pedestrians. We further
visualized the distribution of detected older pedestrians and found a significant spatial discrepancy
between older pedestrians and residential population of older adults. To account for this spatial
discrepancy, this study proposed a novel index to assess pedestrian demand and walking
environment based on the ratio of the number of pedestrians and the residential population. We also
found pedestrian demand assessed with this index has a stronger correlation with the built
environment compared with population-level travel survey. This novel approach can be used to
assess pedestrian demand for older adults, as well as aging-friendly walking environment.
Keywords: Walkability; aging friendly; street view imagery; human attributes recognition; transfer
learning; walking; pedestrian demand

1. Introduction
With rapidly aging populations in many countries, more and more governments and researchers
have recognized the importance of building age-friendly cities. Researchers have found that regular
physical activities can significantly increase older adults’ life expectancy (Lee et al., 2012) and
reduce the risk of having chronic diseases such as cardiovascular disease (Smith et al., 2007),
coronary heart disease (Manson et al., 1999), type 2 diabetes (Aune et al., 2016), breast cancer (Wu
et al., 2013), and colon cancer (Boyle et al., 2012). As the most common form of physical activity
among older adults, walking has therefore attracted considerable attention (Chodzko-Zajko et al.,
2009; Pahor et al., 2014). Due to their declined physical abilities and mobility associated with aging,
older adults are more sensitive to the surrounding built environment than young adults (Chen et al.,
2022; Feng, 2017; Ghani et al., 2018), so it is important to investigate walking environment for
older adults.
Researchers often employed walkability to assess walking environment, which is defined as
the synergy of certain built environmental factors, including density, diversity, and design, which
can promote and support walking (Forsyth, 2015). For example, the World Health Organization
(WHO) has summarized 88 essential indicators for evaluating aging-friendly cities, of which 12
features relate to the built-environment factors for pedestrians (World Health Organization, 2007).
Some researchers used the Walk Score, a rating system calculated based on the distance to the
nearest amenities such as hospitals and schools, to measure walkability (Carr et al., 2010). However,
these measurements focus only on the availability of walkable infrastructure and walking
opportunities, and may not reflect actual pedestrian demand on streets (Chen et al., 2020; Dhanani
et al., 2017) An area with high walkability may not necessarily have higher pedestrian activity, and
similarly, an area with low walkability may still have many pedestrians.
Indeed, the association between built environment factors and walking behavior tends to be
intertwined and dependent on the social and urban context. For example, higher urban density often
promotes walking behavior in low or medium density cities. However, such association tends to be
insignificant in high-density cities such as Hong Kong (Cerin et al., 2013; Kamada et al., 2011).
Additionally, empirical studies have found that the walking behaviors of disadvantaged populations
tend to be less responsive to built-environment factors, and even demonstrate opposite responses to
the expected effects of these factors (Adkins et al., 2017; Forsyth et al., 2009; Frank et al., 2008;
Huang et al., 2022; Lovasi et al., 2008). Therefore, it is also important to obtain pedestrian demand
on streets (Chen et al., 2020). This approach may arguably more accurate and straightforward to
assess pedestrian activity than walkability-based methods.
There are two common approaches to collect fine-grained pedestrian demand data. The first
one is field observation (Brownson et al., 2009). But it is costly, time-consuming and inefficient.
The second approach is to infer pedestrian activity from population-level travel surveys (Schwartz,
2000). However, such inferring may be inaccurate because people tend to underreport short walking
trips or short walking leg of a trip (Chen et al., 2022). In recent years, researchers have begun to use
Street View Imagery (SVI) along with computer vision techniques to estimate pedestrian volume in
a large area such as a whole city, because of its efficiency and cost-effectiveness. However, due to
technical limitations, most existing research only uses SVIs to detect pedestrians, but not to classify
pedestrians by age (Chen et al., 2020; Yin et al., 2015).
Therefore, the main purpose of this study is to develop a novel method to detect older
pedestrians using SVI and a non-facial human attributes recognition algorithm. We also analyze the

spatial distribution of older pedestrians in a whole city. This study contributes to existing knowledge
in four aspects. First, we established a large SVI-based dataset with labels for older pedestrians.
Second, we developed a novel method for detecting older pedestrians using SVI and non-facial
human attributes recognition approach. Third, we examined the spatial distribution of older
pedestrians in an entire city. Fourth, we used the ratio of the detected pedestrians and the number of
residents to evaluate walking environment.
2. Related works
2.1 Measurement walking behaviors of older adults
Although numerous studies focus on walking behavior and other physical activities of older
adults, most of them have concentrated on North America and Europe where older people have
different lifestyle and walking habits compared to East Asia. For example, North American and
European studies (Bennie et al., 2013; Harvey et al., 2015; Rezende et al., 2014) found that older
adults spent less time on walking compared with younger people, while Asian (Hong Kong and
Japan) studies found the opposite outcomes (Hui et al., 2001; Tsunoda et al., 2012). The methods to
measure walking behaviors also differ across studies. Most studies focus on individual walking
behavior (IWB) of older people, such as daily walking time, walking frequency, and walking
distance (Mendes de Leon et al., 2009; Moniruzzaman et al., 2015; Shigematsu et al., 2009; Van
Holle et al., 2016). Some studies focus on collective walking behavior (CWB) such as urban vitality
and pedestrian volume (Chen et al., 2020; Lee et al., 2017; Sung et al., 2015). Traditionally, IWB
can be obtained by questionnaire (Barnett et al., 2017) while CWB can be assessed by field
observation (Yin, 2017). However, these approaches are costly and time-consuming, and unsuitable
for large-scale studies. Recently, researchers employed urban big data such as smart card data (Long
& Thill, 2015) and mobile signal data (Du et al., 2017) to assess both IWB and CWB. However,
such data still cannot differentiate street-level pedestrian behaviors, because such data sources tend
to have low spatial resolution (e.g., 10 m or 100 m).
2.2 SVI as a novel data to measure urban environment and walking activity
The recent proliferation of Street View imagery (SVI), rapid advances in computer vision
technology and the soaring computing power have created great opportunities for measuring street-
level built-environment and human activities. Some researchers focused on auditing features of the
urban environment such as buildings (Ogawa & Aizawa, 2019), street greenery (Liu et al., 2023;
Lu, 2019) and sidewalks (Ning et al., 2022), while the others used SVIs to predict people’s
subjective perception of the urban environment such as safety (Wang et al., 2019) and aesthetics
(Luo et al., 2022).
On the other research front, some researchers have begun to use SVIs to directly quantify
pedestrian volumes. For example, some researchers have started to pay attention to potential of SVIs
to assess pedestrian volume. Yin et al. (2015) developed an approach to automatically extracting
pedestrian counts on Google SVIs using deep learning technology. They validated the reliability of
the proposed method across 200 street segments in Buffalo, NY, Washington, D.C., and Boston, MA,
USA and found it can produce consistent results with both manual count with Google SVIs and field
count. Chen et al. (2020) has further validated the robustness of using SVIs to estimate pedestrian
volume over 700 street segments in Tianjin, China, by comparing with field observation data (Chen
et al., 2022).

2.3 Non-facial human attributes recognition (NHAR) and crowd analysis
Non-facial human attributes recognition (NHAR) which aims to recognize, describe, and
understand human attributes from images without facial information, has attracted much attention
in computer vision field in recent years (Wang et al., 2022). The face is the most distinctive part of
human and provides an invaluable data source for computer vision algorithms (Thom & Hand, 2020).
However, human faces in SVIs are intentionally blurred to protect privacy (Deng et al., 2014).
Therefore, it is necessary to detect human attributes from non-facial human parts, such as whole
body (Hidayati et al., 2017) or clothing (Xiang et al., 2020). Pedestrian attributes recognition (PAR)
(Wang et al., 2022) is one of such emerging techniques with promising results. It mainly relies on
data derived from people’s posture and gesture. Existing studies have shown that PAR along with
transfer learning (Weiss et al., 2016) and data augmentation (Shorten & Khoshgoftaar, 2019) can
efficiently recognize many different human attributes from images.
Because some pedestrians walk in groups, detecting individuals from a large group of
pedestrians presents additional technical challenge. A new technique known as crowd analysis, may
address this issue (Wu et al., 2010). Crowd analysis is usually applied in crowd behavior analysis
(Saxena et al., 2008), people counting (Liang et al., 2014), anomaly detection (Husni & Suryana,
2010) and people tracking (Rodriguez et al., 2011). Recently, it has been used for pedestrian
detection and pedestrian volume estimation based on SVIs (Chen et al., 2020; Yin et al., 2015).
3. Methodology
3.1 Study area
As one of the densest and most urbanized cities in the world, Hong Kong houses more than 7
million residents in a land of only 1,100 km2. Hong Kong also witnesses an increasing aging
population, presenting various challenges for the region. The number of older adults aged 65 and
over is expected to increase from 1.45 million (19.1% of the total population) in 2021 to about 2.37
million (31.1%) in 2036 (Census and Statistics Department, 2020). According to the government,
Hong Kong will become one of the cites with the highest percentage of older residents in the world
by 2050 (Census and Statistics Department, 2010). In this study, we selected the whole Hong Kong
region as our study area and focus on all streets covered by Google Street View (Figure 1).

Figure 1 The area of Hong Kong region and the distribution of SVI sampling sites. There are 70,021
SVIs and 11,467 of them have detected pedestrians.
The entire territory of Hong Kong consists of three parts: Hong Kong Island, Kowloon and the
New Territories. The dense urban region in Hong Kong is centralized on Hong Kong Island and
Kowloon, which collectively cover a mere 13.8% of the total area, yet house 50% of the population.
Conversely, the New Territories predominantly comprise country parks and rural areas, but with
over 3 million population concentrated in new towns. On Google Maps, there are 70,021 SVI sample
points for the entire Hong Kong area, covering 31,971 street segments and 199 Tertiary Planning
Units (TPUs). Among them, 11,467 SVI sample points were detected to have pedestrians.
3.2 The overall study design
Figure 2 shows the overall workflow of older pedestrian detection and data analysis. First, we
detected and cropped pedestrians from SVIs for the entire Hong Kong area using the pre-trained
You Only Look Once (YOLO) v5x6 model (Redmon & Farhadi, 2017). Then, we classified these
pedestrians into two groups (non-elderly vs. older adults) using Resnet50. Resnet50 was pre-trained
with the modified dataset RAP and PA -100K and fine-tuned labeled SVIs. In this way, we obtain
the pedestrian volume with age information for each street in Hong Kong covered by SVIs.

Figure 2. Workflow of the proposed model. (a) Sampling points along the road centerline; (b)
Retrieve SVIs from right and left direction; (c) Detect and crop pedestrians from SVIs by YOLOv5;
(d) Classify pedestrians on age groups by ResNet50.
3.3 Dataset for model training
3.3.1 Pretraining dataset
To improve the accuracy of the model, we used the Richly Annotated Pedestrian dataset (RAP)
(Li et al., 2018) and the PA -100K (Liu et al., 2017) dataset for pretraining. RAP (Richly Annotated
Pedestrian) and PA-100K are datasets specifically curated for human attribute recognition, and

person re-identification tasks in computer vision research. These datasets are generally designed for
academic research and are meant to facilitate training and benchmarking deep learning.
RAP dataset contains 41,585 pedestrian images captured by surveillance cameras in various
public places like streets, parks, and shopping malls. The images come from the CASIA Office of
Turing Robotic Intelligence and the Harbin Institute of Technology in China. The dataset is richly
annotated with various pedestrian attributes such as age, gender, clothing, accessories, and
occlusions.
PA-100K dataset consists of 100,000 pedestrian images, making it one of the largest and most
comprehensive pedestrian attribute recognition datasets. The images were collected from multiple
sources, including surveillance cameras from many countries worldwide. Although specific
countries are not explicitly mentioned, it is safe to assume that the dataset maintains diversity in
terms of ethnicity, clothing, and backgrounds. It is annotated with 26 attributes like gender, age,
clothing type, hairstyle, and accessories.
We combined these two datasets into one. Each image contains a person tagged with multiple
tags, including age and gender. Age in RAP and PA -100K was divided into five classes, including
less than 16, 16-30, 31-45, 46-60, and over 60 years old. In this study, we focused on pedestrians
who are over 60 years old, so we combined the other four classes into one class: 60 years old or less.
The final combined dataset contains 1,654 images with older adults and 1,654 images with others.
To avoid imbalance between the number of images with older adults and people in other age groups,
we selected all 1,654 images with older adults and randomly selected 1,654 images with people in
other age groups from the combined dataset. We then performed data augmentation on these two
newly selected datasets (NSD).
Since each pedestrian in SVIs is face masked, we masked the same areas of the images in RAP
and PA -100K to reduce the difference between the training dataset and the target dataset (Figure
3).
Figure 3. (a), (b), (c) are non-older pedestrian masked with face in RAP and PA-100K, and (d), (e)
are older pedestrian masked with face in RAP and PA-100K.
3.3.2 SVIs dataset for fine-tuning
Pedestrians in SVIs have different distortions, distributions, and background information that
do not match RAP and PA -100K. To fit the classification model to the context of SVIs, a fine-tuning
process for the pre-trained model is needed. We downloaded SVIs for all sample points in Hong
Kong from Google Street View, maintaining a 50-meter separation between each pair of sampling

locations. Each sample point has four SVIs with a direction of 0°, 90°, 180°, and 270°. All
downloaded images have a size of 1024*1024 pixels and were taken between 2018 and 2019. From
them, we selected 2,000 SVIs in different scenes (downtown, suburban, highway, etc.) to create a
dataset for fine-tuning. Three trained research assistants participated in pedestrian labeling. Two of
them performed the labeling while the third one reviewed their results. In SVIs, each pedestrian's
face is masked, so we determined age based on features other than the face. If a pedestrian has
obvious aging characteristics such as white hair and a strong hunchback, he or she is classified as
an older adult. Accurately determining the ages of pedestrians remains challenging. Therefore, our
research assistants only classify them based on their perceived age group, differentiating between
older and non-older adults without specifying their precise age. We also discarded the pedestrians
50-m away from a SVI sampling point because it is challenging to classify age group with a small
image. Finally, a total of 1,546 pedestrians were labeled, including 708 older adults and 838 non-
older individuals (Figure 4).
Figure 4. (a), (b), (c) are non-elderly in SVI, and (d), (e) are older adults in SVI.
3.4 Detector and classifier
The target of the object detection model used in this study is to detect pedestrians from SVIs
and classify them into two age groups. Object detection models can be divided into two types based
on their pipelines. One is one-stage models, such as You Only Look Once (YOLO) series (Redmon
et al., 2016) and single shot multibox detector (SSD) (Liu et al., 2016). The other is two-stage
models, which segregates detection and classification processes, such as Faster R-CNN, which
offers better performance than one-stage models (Ren et al., 2015). In this study, we performed the
detection and classification process separately to improve the performance of the model. First, we
used the pre-trained YOLO v5x6 (Redmon & Farhadi, 2017) to detect pedestrians from SVIs and
cropped them down. The 0.5mAP (mean average precision at IoU 0.5) for the pretrained YOLO
v5x6 to detect pedestrians in SVIs is 87.7. Next, we classified the cropped pedestrians with our best
fitted model. After our pilot study, we found that Resnet50 (Koonce, 2021) offered the best tradeoff
in terms of classification performance, run-time, and memory consumption.
3.4.1 YOLO model
As the most popular one-step object detection model series, YOLO series has outperformed
other object detection models in terms of accuracy and speed. The YOLO series also has advantages
in detecting objects of different sizes and overlapping objects. In terms of effectiveness and stability,

we selected YOLO v5, the fifth versions of YOLO series, as the detector for our approach. From
the YOLO v5 family, we selected the YOLOv5x6 model with the largest size, which was pretrained
on the COCO dataset (Lin et al., 2014) and achieved 72.0 0.5mAP for the validation dataset.
The basic architecture of YOLOv5 is shown in Figure 5. The whole network consists of five
sections: Input, Backbone, Neck, Head, and Output. The backbone section functions as the feature
extractor and transfers input images into feature maps. The neck section receives feature maps from
the backbone and combines these features into logic groups for detection. The head section is also
called the detection section. It outputs vectors containing the probability for each class, the position
and the size of each object.
Figure 5. Overview of the architecture of YOLO v5
3.4.2 Resnet 50 model
Classifying pedestrian age groups from SVIs requires deep CNN due to high intraclass variance
and low interclass variance. Training deep neural networks is challenging due to the vanishing
gradient problem (Habibzadeh et al., 2018) and the degradation problem (Wichrowska et al., 2017).
To address these challenges, the Residual Network (ResNet) was developed. The fundamental part
of ResNet is batch normalization. Batch normalization modifies the input layer to improve the
performance of the network and reduce the shifting of covariates. Another key area is identity
connectivity, which helps ResNet's network mitigate the vanishing gradient problem.
In this study, we used the ResNet50, which is a variant of ResNet model and can handle the
input images with height, width as multiple of 32 and 3 as channel width. The output of ResNet50
is the probability of each class, and we selected the class with higher probability as the classified
result for each pedestrian cropped from SVIs.

Figure 6. Overview of the architecture of ResNet50
3.5 Model training and evaluation
First, we randomly split the NSD dataset: 75% for training dataset, 25% for validation. Second,
we employed YOLO v5x6 to crop each pedestrian from labeled SVIs to create our customized
dataset. Third, we again randomly split the customized data: 70% for training, 20% for validation,
and 10% for test.
Next, we conducted three groups of experiments. In the first group, we trained models with
only NSD dataset. In the second group, we trained models with only the training and the validation
dataset of customized SVIs. In the third group, we pretrained the models using the NSD dataset and
fine-tuned models with training and validation dataset of customized SVIs. All three groups of
models are tested by the test dataset of customized SVIs. The results of the three groups of
experiments are shown in Table 1. The result showed that the model pre-trained in the NSD and
fine-tuned in cropped SVIs performed the best with an accuracy of 87.1%. Please note that the age
categorization of cropped SVIs is determined by research assistants based on their perceived age
group. It only differentiates between older and non-older adults without specifying their precise age.
In this study, the attribute of AgeAbove60 in NSD is aligned with the classification of older adult in
Cropped SVIs.
Table 1. Comparison of model performances through different training dataset
Models
Attribute
Training dataset
Test dataset
Highest accuracy
Resnet50
Older adult
Cropped SVIs
Cropped SVIs
82.2%
Resnet50
AgeAbove60/ Older adult
RAP and PA-100K
Cropped SVIs
76.8%
Resnet50
AgeAbove60/ Older adult
RAP and PA-100K/Cropped SVIs
Cropped SVIs
87.1%
4. Pedestrian detection and recognition in Hong Kong
4.1 Geographic distribution
The final trained model detected 35,353 pedestrians from 70,021 SVIs in Hong Kong,
including 7,375 older pedestrians and 27,978 non-older pedestrians. We mapped and visualized the
number of detected older pedestrians and the proportion of older pedestrians among all detected
pedestrians within different spatial units in Hong Kong.
More older pedestrians were detected in dense urban areas of Kowloon and Hong Kong Island
(Figure 7a and Figure 7b). In terms of proportions, a high proportion of older pedestrians among
all pedestrians were found in TPUs distributed around Hong Kong Island and new towns in New

Territories (Figure 7c and Figure 7d).
Figure 7. The geographic distribution of detected older pedestrians in Hong Kong. (a) number of
older pedestrians in TPUs; (b) number of older pedestrians in road segments in the core urban area
(including most part of Hong Kong Island and Kowloon); (c) the proportion of older pedestrians
among all detected pedestrians in TPUs; (d) the proportion of older pedestrians among all detected
pedestrians in road segments in the core urban area.
4.2 Spatial mismatch between older pedestrians and older residents
To analyze any potential spatial mismatch between detected older pedestrians and older
residents, we conducted two comparative analyses. We collected resident population in a specific
area in 2020 from WorldPop (worldpop.org). This dataset estimates population residing in each
100m*100m grid using census data in 2020 and a random forest model (Stevens et al., 2015). In
urban areas, a walk of 500 meters or less to the nearest amenity is a desirable distance, so researchers
typically use 500-meter square grids to study walking behavior and walkability (Dovey & Pafka,
2020). Therefore, we aggregated the data of both detected pedestrians and WorldPop into 500 m*
500 m square grid.
First, we compared the number of detected older pedestrians and the population of older
residents in each grid. We classified all grids with high or low (H or L) values of detected older
pedestrians and older residents, according to the median value of the two variables respectively.
Accordingly, we classified the grids into four groups: H/H (high number of older pedestrians and

high population of older residents), H/L, L/H, and L/L (Figure 8 and Table 2&3). In Table 3, we
conducted t-tests on built environment factors for four pairs of comparisons (H/H grids vs. non-H/H
grids, L/H vs. non-L/H, H/L vs. non-H/L, L/L, vs. non-L/L in Figure 8).
H/H grids are mainly located in high density residential area of Kowloon and Hong Kong
Island such as Sham Shui Po, Yau Ma Tei, and new developed town centers in New Territories. They
have higher population density and proportion of residential area, and shorter distance to MTR
station and to city center, compared with other grids. L/H grids are mainly concentrated in major
commercial hubs or mixed used areas such as Tsim Sha Tsui, Central District. They have higher
proportion of commercial area, lower population density and proportion of residential area, and
shorter distance to city center. H/L grids are mainly located in residential areas (e.g., Waterfall Bay,
Tsing Shan Tsuen) scattering in the perimeter of H/H grids. They have higher population density,
and proportion of residential area, lower proportion of commercial area, and shorter distance to
MTR stations. The L/L areas are scattered in the suburban areas or low-density residential areas of
Hong Kong Island (e.g., Stanley), Kowloon (e.g., Yau Tong), and the New Territories (e.g., Sai
Kung), often far away from city center. They have longer distance to MTR stations and city center,
lower population density, proportion of residential area and proportion of commercial area.
Figure 8. Classification of urban areas into four groups, according to the population of older
residents and the number of detected older pedestrians on SVIs
Table 2. Features and representative examples of four groups of areas in Figure 8. Source: Google
Inc.
Type of area
Features
Representative
location
Representative aerial view

High number of older
residents and high
number of detected older
pedestrians（H/H）
High density
residential area
Sham Shui Po,
Yau Ma Tei
Low number of older
residents and high
number of detected older
pedestrians （L/H）
High density
commercial &
mixed used area
Tsim Sha Tsui,
Central District
High number of older
residents and low number
of detected older
pedestrians （H/L）
Medium density
residential area
Waterfall Bay,
Tsing Shan Tsuen
Low number of older
residents and low number
of detected older
pedestrians （L/L）
Suburban and
mountainous
area
Stanley, Sai Wan
Shan, Yau Tong
Table 3 The t-value of the t-test of built environment factors for four pairs of comparisons (H/H
grids vs. non-H/H grids, L/H vs. non-L/H, H/L vs. non-H/L, L/L, vs. non-L/L in Figure 8).
Population
density
Commercial
proportion
Residential
proportion
Distance to
MTR
Distance to city
center
H/H vs. non-H/H
10.02**
0.29
4.87**
-3.87**
-2.55**
L/H vs. non-L/H
-4.41**
4.99**
-3.20**
-0.10
-1.84**
H/L vs. non-H/L
4.41**
-1.50*
2.52**
-1.35*
1.07
L/L vs. non-L/L
-10.02**
-2.91**
-4.29**
5.19**
3.24*
*: p < 0.05; **: p < 0.01.
5 Walking index
5.1 Definition
We believe that the spatial mismatch between the pedestrian demand and residential population
can reveal the degree of walking attractiveness, i.e., to degree to which an area is conductive to walk.
The distribution of residential population has a direct impact on collective walking behavior, e.g.,
pedestrian demand. Given other conditions are constant, an area with higher residential population
will have higher pedestrian demand. Conversely, assuming equal population density between two
areas, differences in pedestrian demand may reveal a difference in walking attractiveness.
Furthermore, walking attractiveness may be different for certain groups of people, such as older or
disabled pedestrians (Adkins et al., 2017; Forsyth et al., 2009; Frank et al., 2008; Lovasi et al., 2008).
Therefore, we constructed a new walking index using the ratio of observed pedestrian demand of a
given group and the residential population of that group with the following equation:
 
= 


 (1)
where 
 represents the number of pedestrians with attribute a in the i-th geographic unit, 


represents the total residential population of people with attribute a in the i-th geographic unit.

The formula can be directly applied to older adults. The walking index for older adults (OWI)
could be measured by the ratio of number of older pedestrians and population of older residents in
an area. Similarly, it can also be used for other age groups, certain population subgroups, or all
people.
To obtain the total population in customized units, we again collected data from WorldPop in
2020. We calculated the walking index for both older adults (OWI) and all people (WI) for the whole
Hong Kong.
5.2 Distribution of walking index for older people (OWI)
We calculated the spatial distribution of OWI based on three spatial units: TPUs, 500 × 500 m
rectangular grids and 200 × 200 m rectangular grids respectively. The results are visualized in
Figure 9(a)-(c). Of the ten TPUs with the highest OWI, three are in the northeast corner of the
Central and Western District, four are in Tsim Sha Tsui and one is in Causeway Bay, which are the
city center of Hong Kong where many shopping malls are located. Far away from city center, Sai
Kung Peninsula also has a high OWI, which remains untouched by urbanization and is also a popular
place for hiking. Of the ten TPUs with observed pedestrians and the lowest OWI, nine are scattered
in New Territory.
Figure 9. Spatial distribution of OWI. (a) OWI at TPUs in Hong Kong; (b) OWI at 500 × 500 m
rectangular grids in Hong Kong; (c) OWI at 200 × 200 m grids for the main area of Kowloon and
Hong Kong Island.
5.3 Spatial mismatch between WI and OWI
This study mapped both WI and OWI in each 500 × 500 m rectangular grid in the same way.
We classified all grids into high or low values of these two variables, according to their respective
median value. Accordingly, we divided the grids into four groups: H/H (high WI and high OWI),
H/L (low WI and high OWI), L/H (low WI and high OWI), and L/L (low WI and low OWI) (Figure
10 and Table 4). In Table 5, we conducted t-tests for built environment features for four pairs of
comparisons (H/H grids vs. non-H/H grids, L/H vs. non-L/H, H/L vs. non-H/L, L/L, vs. non-L/L in
Figure 10).
H/H areas are mainly concentrated in compact high and middle density residential areas such
as Central District and Mong Kok. They have higher proportion of commercial area, lower
proportion of residential area, population density, and shorter distance to city center, compared with
other areas. L/H areas scatter around H/H areas especially in neighborhoods inhabited by affluent
residents, such as the Mid-Levels. They have shorter distance to MTR stations. H/L areas are often
dispersed throughout suburban and mountainous area with adequate green space. They have longer

distance to MTR stations and city center, lower population density, proportion of residential area
and the proportion of commercial area. L/L areas are mainly located at high and middle-rise
residential area in New Territories and peripheries of Kowloon and Hong Kong Island especially in
some new towns. They have higher population density and proportion of residential area, lower
proportion of commercial area, and shorter distance to MTR stations,
Figure 10. Classification of urban areas into four groups, according to the proportion of older
residents and the proportion of older detected pedestrians on SVI.
Table 4. Features and representative examples of four groups of areas in Figure 9. Source: Google
Inc.
Type of area
Features
Representative
location
Representative aerial view
High WI and high
OWI （H/H）
Compact High and
Mid-rise residential
area
Central District,
Mong Kok
Low WI and high
OWI （L/H）
Open High and Mid-
rise residential area
Mid-levels,
Quarry Bay
High WI and low
OWI （H/L）
Suburban and
mountainous area
Lantau Island,
Kam Tin

Low WI and low
OWI （L/L）
High and Mid-rise
residential area in
New Territories and
fringe of Kowloon
and Hong Kong
Island
Sha Tin, Chai
Wan
Table 5 The t-value of the t-test of built environment factors for four pairs of comparisons (H/H
grids vs. non-H/H grids, L/H vs. non-L/H, H/L vs. non-H/L, L/L, vs. non-L/L in Figure 10).
Population
density
Commercial
proportion
Residential
proportion
Distance to
MTR
Distance to
city center
H/H vs. non-H/H
-2.41**
4.58**
-2.25**
0.70
-2.14**
L/H vs. non-L/H
0.45
-0.83
0.18
-1.67**
-0.63
H/L vs. non-H/L
-4.38**
-1.48*
-1.90**
3.64**
2.72**
L/L vs. non-L/L
5.11**
-2.95**
3.55**
-2.16**
0.66
Note. *: p< 0.05; **: p<0.01.
5.4 The validation of WI and OWI
To validate the new walking index, we used data from the Hong Kong Travel Characteristics
Survey (TCS) conducted by the Transport Department of the Hong Kong government in 2012. The
data from TCS were acquired from a large representative sample of 101,385 residents in Hong Kong.
The respondents were asked to provide individual information (e.g., age, gender, occupation,
address) and trip information (e.g., trip mode, trip frequency) during the last 24 hours up to the
surveying time.
We ran two multivariate linear regressions to predict the WI (Model 1) and OWI (Model 3)
and as independent variables (Table 6). As a comparison, we conducted two more models to predict
frequency of walking trips of all respondents (Model 2) and older respondents (Model 4) with built
environment factors. The unit of analysis is TPU for all models.
Built environment factors include population density (Barnett et al., 2017), street intersection
density (Cerin et al., 2017), land-use diversity (Thornton et al., 2017), bus stop density (Christiansen
et al., 2016), and the coverage of Mass transit rail (MRT) stations (Fenton, 2005). We geolocated all
respondents in QGIS based on their dwelling locations. Population density was defined as the
resident population per unit of area and obtained from Census and Statistics Department of Hong
Kong. Street intersection density was defined as the number of street intersections per unit area. The
land use diversity was defined as the entropy score of land use distribution and calculated as
(−1) ∑(ln())/ln ()
, where  is the share of specific land use and n is the number of land
use types. Bus stop density was calculated as the number of bus stops per unit of land area. The
coverage of Mass Transit Railway (MTR) stations was defined as the percentage of land covered by
the 500 meters buffer of MTR stations. Both dependent variables and independent variables were
converted into Z scores to obtain standardized coefficients, which allow us to compare the effect
sizes of different factors in predicting the outcome.
Table 6. Results of multivariate regression model of OWI and average walking frequency per older
respondents from TCS with built-environment factors as independent variables.
Model 1
Model 2
Model 3
Model 4
Dependent variable
WI
Walking frequency
of all respondents
OWI
Walking frequency of
older respondents
Coef. (SE)
Coef. (SE)
Coef. (SE)
Coef. (SE)

Street intersection density
0.201 (0.069) **
-0.013 (0.011)
0.185 (0.059) *
-0.007 (0.005)
Population density
-0.292 (0.062) ***
-0.004 (0.011)
-0.202 (0.071) **
-0.004 (0.005)
Bus stop density
0.350 (0.083) ***
0.012 (0.005) *
0.347 (0.086) ***
0.024 (0.005) ***
MTR station 500m coverage
0.283 (0.062) ***
0.040 (0.010) ***
0.213 (0.064) **
0.006 (0.004)
Land use diversity
0.175 (0.059) **
0.016 (0.011)
0.222 (0.061) ***
0.014 (0.004) **
Adjusted R2
0.532
0.006
0.507
0.006
Note. *: p< 0.05; **: p<0.01; ***: p<0.001
Compared with average walking frequency from TCS, both WI and OWI have higher
association with built-environment factors. Hence, our approach may better predict overall walking
demand than traditional travel survey. It is worth noting that individual travel survey data may
exhibit fluctuation due to the individual attributes of respondents. As such, the WI and OWI present
a superior alternative to evaluate collective pedestrian activity.
6 Discussion
6.1 methodological contribution
Previous studies have shown SVI to be a valuable data source for automatic and large-scale
urban environment audit, with many focusing on static features such as buildings, roads, and
greenery. Dynamic elements, including vehicles, pedestrians, and bicycles, have not been
adequately studied primarily because of their volatility over time. Therefore, researchers typically
measure pedestrian volumes and walking behavior through field observation or questionnaires.
Field counts are time-, labor-, and cost-intensive, which is impractical for a large area (Lee & Talen,
2014). Self-reported data are prone to recall bias and may be difficult to be geolocated (Saelens &
Handy, 2008).
Recently, some researchers have suggested that estimating pedestrian demand using SVI is
feasible. The number of pedestrians at a given location at a given time can be estimated from a
single SVI image. As the sample size increases exponentially (e.g., hundreds of SVIs at different
locations and/or times in an area), the pedestrian volume estimate becomes closer to the true value
for that area, although the estimate for a particular location or time may be very incorrect (Richards,
1961). Recent empirical studies have demonstrated the consistency of pedestrian volume between
the estimate by SVI and the field audit, which reached 0.87 Cronbach's alpha under certain
circumstances (Chen et al., 2020). This opened new possibilities for assessing pedestrian activity
at a scale, depth, and scope inaccessible to traditional assessment methods.
In this study, we extend previous studies by developing a novel approach to assess pedestrian
age group in SVI. First, we used the existing pedestrian dataset RAP and PA -100k as part of the
training samples, which contains the age information to pre-train the models. Second, we created
our own training dataset based on SVI to refine the pre-trained model. After the two-stage training,
the accuracy of the model tested on pruned SVIs reached 87.1%. Using this approach, we are able
to identify older pedestrians in SVI. Our efforts can help create healthy cities and aging-friendly
communities, which are critical for an aging society. It is also possible to use our approach to assess
pedestrians of other age groups or specific attributes (e.g., female, or people using wheelchair).
More importantly, our approach also offers a unique advantage over traditional research studies.
To assess the impact of the built environment on travel behavior, traditional travel survey and
research focused on built environment features around home. However, on average, Hong Kong
people spend 40% of their waking time far away from home (Census and Statistics Department,
2013). It has been pointed out the uncertainty in identifying the spatial areas that exert influences
on individual behaviors may hinder our understanding of environment-behavior link (Kwan, 2012).

Therefore, to better understand the impact of built-environment context on walking behavior, we
use the actual observed pedestrians to estimate walking demand in a given area.
Conventional environmental behavior research has typically employed travel surveys to obtain
data on pedestrian walking patterns, specifically focusing on individual walking behavior. However,
this study utilizes WI/OWI as a metric for assessing collective pedestrian activity on streets. We
demonstrated that WI/OWI has a stronger correlation with the built environment features compared
with individual walking behavior, which is in line with one previous study (Jiang et al., 2021). This
also suggests that WI/OWI may serve as a more reliable measurement for evaluating walking
environments, as opposed to travel surveys which are susceptible to the influence of personal
attributes. Therefore, such an approach can advance the field of research by providing a much-
needed spatial match between the built environment and collective pedestrian activity.
6.2 Spatial distribution of older pedestrian and OWI in Hong Kong
Using our new approach, we measured the spatial distribution of older pedestrians in Hong
Kong. Several results are worth noting. First, there is an obvious spatial mismatch between older
pedestrians and population of older residents. High number of older residents and high number of
detected older pedestrians（H/H）grids concentrated in high-density compact development areas
which are characterized by high-rise buildings and mixed land uses. These areas provide older
residents with easy access to vital amenities, including healthcare, transportation, and commerce,
situated within a walkable range, particularly addressing their mobility constraints (Burton, 2000).
Low number of older residents and high number of detected older pedestrians（L/H）grids have
built environment features that are similar to those of H/H areas, but which have a high
concentration of commercial destinations and fewer residential buildings. As a result, such built
environments attract a large flow of older pedestrians living in other areas to fulfill utilitarian needs,
such as going to a restaurant or visiting a doctor. These areas, such as Tsim Sha Tsui, Central District,
often have high cost of living, which may excess the affordability of older adults. High number of
older residents and low number of detected older pedestrians (H/L) grids mainly scattered in low-
and middle- density residential areas. Low number of older residents and low number of detected
older pedestrians (L/L) areas are scattered in the suburban areas or low-density residential areas of
Hong Kong Island (e.g., Stanley), Kowloon (e.g., Yau Tong), and the New Territories (e.g., Sai
Kung). Both H/L and L/L areas are often far away from city center, and amenities, services, and
healthcare facilities are scarce (Sun & Lau, 2021). The lack of walking destinations may hinder the
walking activities among older adults. Furthermore, residents living in those areas often have other
transportation options, e.g., private vehicles, which partly reduce the need of walking.
To account for such spatial mismatch, we develop a new walking index for the older people
(OWI) based on the ratio of observed pedestrians to the resident population. It can reflect the quality
of urban design, human activity, and urban vitality for older adults by measuring the extent to which
older adults are willing to walk. Most studies quantify the walkability of a given area based on
features of the built environment that support walking, such as the density of intersections, the
continuity and directness of paths, and the presence of sidewalks and other pedestrian infrastructure
(Chen et al., 2022; Li et al., 2021; Yang et al., 2019). However, the availability and accessibility of
pedestrian-friendly infrastructure may not reflect actual walking behavior in such an area. Our
walking index can help fill this gap by accounting for the collective walking behavior.

6.3 Spatial mismatch between WI and OWI
This study focuses on aging-friendly built environment and the pedestrian demand of older
adults. Due to the unique needs and physical abilities of older pedestrians, locations that attract
general pedestrians may not accommodate the older pedestrians. Therefore, we mapped the spatial
mismatch between WI and OWI to explore the spatial disparity with four conditions (H/H, L/H,
H/L, and L/L), and further identity the built environment features associated with each condition.
High WI and high OWI areas (H/H) are characterized by higher proportion of commercial area,
lower proportion of residential area and closer distance to the city center, compared with other areas.
A high concentration of commercial spaces increases the availability of various services, shopping,
dining, and entertainment options within walking distance. The areas close to the city center tend to
have better pedestrian infrastructure, such as sidewalks, crosswalks, and street lighting, which
enhances the safety and comfort of pedestrians. Low WI and high OWI (L/H) areas have shorter
distance to MTR stations compared with other areas. This indicates that older adults are more
sensitive to the accessibility of public transit compared with general population. Accessibility to
MTR is particularly appealing to older adults, because it is reliable, comfortable, and cost-effective
(Hess, 2009; Kostyniuk & Shope, 2003) .
High WI and low OWI (H/L) areas are further away from MTR stations and the city center,
and have lower proportions of commercial areas, compared with other areas. This further
underscores the notion that older adults are more sensitive to the accessibility of commercial
destinations and public transport. Low WI and low OWI areas (L/L) are featured by higher
population density and proportion of residential area, lower proportion of commercial, and further
away from the city center, compared with others. Most of these areas are high-density housing
estates featuring multiple high-rise residential buildings in New Territories and the fringe of
Kowloon and Hong Kong Island. The lack of commercial facilities and dense living environment
may hinder the older adults’ willingness to walk.
6.4 Limitations and future research
Based on this study, four possible directions for future research emerge. First, this study has
identified associations between older pedestrian count and built environment factors. However,
more rigorous research designs (e.g., natural experiment) are necessary to find out any causal
relationships between pedestrian volume, built-environment factors and older adults' willingness to
walk. Second, pedestrian behavior can be measured using individual-level mobility data extracted
from mobile data to understand individual walking behavior and its geographic context. Third, other
attributes of pedestrians such as gender and disability are also meaningful and should receive more
attention. Fourth, population-weighted exposure and the Gini index of exposure of different groups
of people to the built environment can be further investigated.
7 Conclusion
In this study, we proposed a novel approach to automatically recognize older adults using SVI,
which could be used to estimate pedestrian demand and evaluate walking environment for older
people. Our proposed model achieved high accuracy (87.1%) in detecting older pedestrians from
SVIs. The results of the multivariate regression models illustrated that the ratio of older pedestrians
to older residents has the potential to be a good indicator of walking attractiveness for older adults,
especially in high residential areas. The result also indicated a spatial mismatch between the walking

and resident population for older people. Researchers and urban planners should consider the
distribution and needs of older pedestrians for further urban planning interventions.
References
Adkins, A., Makarewicz, C., Scanze, M., Ingram, M., & Luhr, G. (2017). Contextualizing
walkability: Do relationships between built environments and walking vary by socioeconomic
context? Journal of the American Planning Association, 83(3), 296–314.
Aune, D., Sen, A., Henriksen, T., Saugstad, O. D., & Tonstad, S. (2016). Physical activity and the
risk of gestational diabetes mellitus: A systematic review and dose–response meta-analysis of
epidemiological studies. European Journal of Epidemiology, 31(10), 967–997.
Barnett, D. W., Barnett, A., Nathan, A., Van Cauwenberg, J., & Cerin, E. (2017). Built environmental
correlates of older adults’ total physical activity and walking: A systematic review and meta-analysis.
International Journal of Behavioral Nutrition and Physical Activity, 14(1), 1–24.
Bennie, J. A., Chau, J. Y., van der Ploeg, H. P., Stamatakis, E., Do, A., & Bauman, A. (2013). The
prevalence and correlates of sitting in European adults-a comparison of 32 Eurobarometer-
participating countries. International Journal of Behavioral Nutrition and Physical Activity, 10(1),
1–13.
Boyle, T., Keegel, T., Bull, F., Heyworth, J., & Fritschi, L. (2012). Physical activity and risks of
proximal and distal colon cancers: A systematic review and meta-analysis. Journal of the National
Cancer Institute, 104(20), 1548–1561.
Brownson, R. C., Hoehner, C. M., Day, K., Forsyth, A., & Sallis, J. F. (2009). Measuring the built
environment for physical activity: State of the science. American Journal of Preventive Medicine,
36(4), S99-S123. e12.
Burton, E. (2000). The compact city: Just or just compact? A preliminary analysis. Urban Studies,
37(11), 1969–2006.
Carr, L. J., Dunsiger, S. I., & Marcus, B. H. (2010). Walk scoreTM as a global estimate of
neighborhood walkability. American Journal of Preventive Medicine, 39(5), 460–463.
Census and Statistics Department. (2020). Hong Kong Population Projections 2020-2069. Census
and Statistics Department. https://www.statistics.gov.hk/pub/B1120015082020XXXXB0100.pdf
Census and Statistics Department, C. and S. D. (2013). Thematic Household Survey Report—Report
No. 56.
Cerin, E., Lee, K., Barnett, A., Sit, C. H. P., Cheung, M., & Chan, W. (2013). Objectively-measured
neighborhood environments and leisure-time physical activity in Chinese urban elders. Preventive
Medicine, 56(1), 86–89. https://doi.org/10.1016/j.ypmed.2012.10.024
Cerin, E., Nathan, A., Van Cauwenberg, J., Barnett, D. W., & Barnett, A. (2017). The neighbourhood
physical environment and active travel in older adults: A systematic review and meta-analysis.
International Journal of Behavioral Nutrition and Physical Activity, 14(1), 1–23.
Chen, B., Tu, Y., Wu, S., Song, Y., Jin, Y., Webster, C., Xu, B., & Gong, P. (2022). Beyond green
environments: Multi-scale difference in human exposure to greenspace in China. Environment
International, 166, 107348.
Chen, L., Lu, Y., Sheng, Q., Ye, Y., Wang, R., & Liu, Y. (2020). Estimating pedestrian volume using
Street View images: A large-scale validation test. Computers, Environment and Urban Systems, 81,
101481.
Chen, L., Lu, Y., Ye, Y., Xiao, Y., & Yang, L. (2022). Examining the association between the built

environment and pedestrian volume using street view images. Cities, 103734.
Chen, S., Wang, T., Bao, Z., & Lou, V. (2022). A path analysis of the effect of neighborhood built
environment on public health of older adults: A Hong Kong study. Frontiers in Public Health, 10.
Chodzko-Zajko, W. J., Proctor, D. N., Singh, M. A. F., Minson, C. T., Nigg, C. R., Salem, G. J., &
Skinner, J. S. (2009). Exercise and physical activity for older adults. Medicine & Science in Sports
& Exercise, 41(7), 1510–1530.
Christiansen, L. B., Cerin, E., Badland, H., Kerr, J., Davey, R., Troelsen, J., Van Dyck, D., Mitáš, J.,
Schofield, G., & Sugiyama, T. (2016). International comparisons of the associations between
objective measures of the built environment and transport-related walking and cycling: IPEN adult
study. Journal of Transport & Health, 3(4), 467–478.
Deng, Y., Luo, P., Loy, C. C., & Tang, X. (2014). Pedestrian attribute recognition at far distance.
Proceedings of the 22nd ACM International Conference on Multimedia, 789–792.
Dhanani, A., Tarkhanyan, L., & Vaughan, L. (2017). Estimating pedestrian demand for active
transport evaluation and planning. Transportation Research Part A: Policy and Practice, 103, 54–
69.
Division, S. A. P. (2010). World population prospects: The 2008 revision (Vol. 1). United Nations
Publications.
Dovey, K., & Pafka, E. (2020). What is walkability? The urban DMA. Urban Studies, 57(1), 93–
108.
Du, Y., Yue, J., Ji, Y., & Sun, L. (2017). Exploration of optimal Wi-Fi probes layout and estimation
model of real-time pedestrian volume detection. International Journal of Distributed Sensor
Networks, 13(11), 1550147717741857.
Ewing, R., & Handy, S. (2009). Measuring the unmeasurable: Urban design qualities related to
walkability. Journal of Urban Design, 14(1), 65–84.
Feng, J. (2017). The influence of built environment on travel behavior of the elderly in urban China.
Transportation Research Part D: Transport and Environment, 52, 619–633.
Fenton, M. (2005). Battling America’s epidemic of physical inactivity: Building more walkable,
livable communities. Journal of Nutrition Education and Behavior, 37, S115–S120.
Forsyth, A. (2015). What is a walkable place? The walkability debate in urban design. Urban Design
International, 20(4), 274–292.
Forsyth, A., Oakes, J. M., Lee, B., & Schmitz, K. H. (2009). The built environment, walking, and
physical activity: Is the environment more important to some people than others? Transportation
Research Part D: Transport and Environment, 14(1), 42–49.
Frank, L. D., Kerr, J., Sallis, J. F., Miles, R., & Chapman, J. (2008). A hierarchy of sociodemographic
and environmental correlates of walking and obesity. Preventive Medicine, 47(2), 172–178.
Ghani, F., Rachele, J. N., Loh, V. H., Washington, S., & Turrell, G. (2018). Do differences in built
environments explain age differences in transport walking across neighbourhoods? Journal of
Transport & Health, 9, 83–95.
Habibzadeh, M., Jannesari, M., Rezaei, Z., Baharvand, H., & Totonchi, M. (2018). Automatic white
blood cell classification using pre-trained deep learning models: Resnet and inception. Tenth
International Conference on Machine Vision (ICMV 2017), 10696, 274–281.
Harvey, J. A., Chastin, S. F., & Skelton, D. A. (2015). How sedentary are older people? A systematic
review of the amount of sedentary behavior. Journal of Aging & Physical Activity, 23(3).
Hess, D. B. (2009). Access to public transit and its influence on ridership for older adults in two US

cities. Journal of Transport and Land Use, 2(1), 3–27.
Hidayati, S. C., You, C.-W., Cheng, W.-H., & Hua, K.-L. (2017). Learning and recognition of
clothing genres from full-body images. IEEE Transactions on Cybernetics, 48(5), 1647–1659.
Huang, G., Li, D., Yu, L., Yang, D., & Wang, Y. (2022). Factors affecting sustainability of smart city
services in China: From the perspective of citizens’ sense of gain. Habitat International, 128,
102645.
Hui, S. C., Chan, C. M., Wong, S. H. S., Ha, A. S. C., & Hong, Y. L. (2001). Physical activity levels
of Chinese youths and the association with physical fitness and demographic variables: The Hong
Kong youth fitness study. Research Quarterly for Exercise and Sport, 72(1), A92–A93.
Husni, M., & Suryana, N. (2010). Crowd event detection in computer vision. 2010 2nd International
Conference on Signal Processing Systems, 1, V1-444-V1-447.
Jiang, Y., Chen, L., Grekousis, G., Xiao, Y., Ye, Y., & Lu, Y. (2021). Spatial disparity of individual
and collective walking behaviors: A new theoretical framework. Transportation Research Part D:
Transport and Environment, 101, 103096.
Kamada, M., Kitayuguchi, J., Shiwaku, K., Inoue, S., Okada, S., & Mutoh, Y. (2011). Differences
in association of walking for recreation and for transport with maximum walking speed in an elderly
Japanese community population. Journal of Physical Activity and Health, 8(6), 841–847.
Koonce, B. (2021). ResNet 50. In Convolutional neural networks with swift for tensorflow (pp. 63–
72). Springer.
Kostyniuk, L. P., & Shope, J. T. (2003). Driving and alternatives: Older drivers in Michigan. Journal
of Safety Research, 34(4), 407–414.
Kwan, M.-P. (2012). The uncertain geographic context problem. Annals of the Association of
American Geographers, 102(5), 958–968.
Lee, I.-M., Shiroma, E. J., Lobelo, F., Puska, P., Blair, S. N., Katzmarzyk, P. T., & Group, L. P. A.
S. W. (2012). Effect of physical inactivity on major non-communicable diseases worldwide: An
analysis of burden of disease and life expectancy. The Lancet, 380(9838), 219–229.
Lee, S., Sung, H., & Woo, A. (2017). The spatial variations of relationship between built
environment and pedestrian volume: Focused on the 2009 Seoul Pedestrian flow survey in Korea.
Journal of Asian Architecture and Building Engineering, 16(1), 147–154.
Lee, S., & Talen, E. (2014). Measuring walkability: A note on auditing methods. Journal of Urban
Design, 19(3), 368–388.
Li, D., Zhang, Z., Chen, X., & Huang, K. (2018). A richly annotated pedestrian dataset for person
retrieval in real surveillance scenarios. IEEE Transactions on Image Processing, 28(4), 1575–1590.
LI, S., Ma, S., Tong, D., Jia, Z., Li, P., & Long, Y. (2021). Associations between the quality of street
space and the attributes of the built environment using large volumes of street view pictures.
Environment and Planning B: Urban Analytics and City Science, 23998083211056340.
Liang, R., Zhu, Y., & Wang, H. (2014). Counting crowd flow based on feature points.
Neurocomputing, 133, 377–384.
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L.
(2014). Microsoft coco: Common objects in context. European Conference on Computer Vision,
740–755.
Liu, D., Jiang, Y., Wang, R., & Lu, Y. (2023). Establishing a citywide street tree inventory with street
view images and computer vision techniques. Computers, Environment and Urban Systems, 100,
101924. https://doi.org/10.1016/j.compenvurbsys.2022.101924

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016). Ssd: Single
shot multibox detector. European Conference on Computer Vision, 21–37.
Liu, X., Zhao, H., Tian, M., Sheng, L., Shao, J., Yi, S., Yan, J., & Wang, X. (2017). Hydraplus-net:
Attentive deep features for pedestrian analysis. Proceedings of the IEEE International Conference
on Computer Vision, 350–359.
Long, Y., & Thill, J.-C. (2015). Combining smart card data and household travel survey to analyze
jobs–housing relationships in Beijing. Computers, Environment and Urban Systems, 53, 19–35.
Lovasi, G. S., Quinn, J. W., Neckerman, K. M., Perzanowski, M. S., & Rundle, A. (2008). Children
living in areas with more street trees have lower prevalence of asthma. Journal of Epidemiology &
Community Health, 62(7), 647–649.
Lu, Y. (2019). Using Google Street View to investigate the association between street greenery and
physical activity. Landscape and Urban Planning, 191, 103435.
Luo, S., Xie, J., & Furuya, K. (2022). Using Google Street View Panoramas to evaluate the
environmental aesthetics quality of blue spaces in urban area. IOP Conference Series: Earth and
Environmental Science, 1092(1), 012001.
Manson, J. E., Hu, F. B., Rich-Edwards, J. W., Colditz, G. A., Stampfer, M. J., Willett, W. C., Speizer,
F. E., & Hennekens, C. H. (1999). A prospective study of walking as compared with vigorous
exercise in the prevention of coronary heart disease in women. New England Journal of Medicine,
341(9), 650–658.
Mendes de Leon, C. F., Cagney, K. A., Bienias, J. L., Barnes, L. L., Skarupski, K. A., Scherr, P. A.,
& Evans, D. A. (2009). Neighborhood social cohesion and disorder in relation to walking in
community-dwelling older adults: A multilevel analysis. Journal of Aging and Health, 21(1), 155–
171.
Moniruzzaman, M., Páez, A., Scott, D., & Morency, C. (2015). Trip generation of seniors and the
geography of walking in Montreal. Environment and Planning A: Economy and Space, 47(4), 957–
976.
Ning, H., Ye, X., Chen, Z., Liu, T., & Cao, T. (2022). Sidewalk extraction using aerial and street
view images. Environment and Planning B: Urban Analytics and City Science, 49(1), 7–22.
Ogawa, M., & Aizawa, K. (2019). Identification of buildings in street images using map information.
2019 IEEE International Conference on Image Processing (ICIP), 984–988.
Organization, W. H. (2007). Checklist of essential features of age-friendly cities. World Health
Organization.
Pahor, M., Guralnik, J. M., Ambrosius, W. T., Blair, S., Bonds, D. E., Church, T. S., Espeland, M.
A., Fielding, R. A., Gill, T. M., Groessl, E. J., King, A. C., Kritchevsky, S. B., Manini, T. M.,
McDermott, M. M., Miller, M. E., Newman, A. B., Rejeski, W. J., Sink, K. M., Williamson, J. D.,
& for the LIFE study investigators. (2014). Effect of Structured Physical Activity on Prevention of
Major Mobility Disability in Older Adults: The LIFE Study Randomized Clinical Trial. JAMA,
311(23), 2387–2396. https://doi.org/10.1001/jama.2014.5616
Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time
object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
779–788.
Redmon, J., & Farhadi, A. (2017). YOLO9000: Better, faster, stronger. Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 7263–7271.
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection

with region proposal networks. Advances in Neural Information Processing Systems, 28.
Rezende, L. F. M. de, Rodrigues Lopes, M., Rey-López, J. P., Matsudo, V. K. R., & Luiz, O. do C.
(2014). Sedentary behavior and health outcomes: An overview of systematic reviews. PloS One,
9(8), e105620.
Richards, F. S. (1961). A method of maximum‐likelihood estimation. Journal of the Royal Statistical
Society: Series B (Methodological), 23(2), 469–475.
Rodriguez, M., Laptev, I., Sivic, J., & Audibert, J.-Y. (2011). Density-aware person detection and
tracking in crowds. 2011 International Conference on Computer Vision, 2423–2430.
Saelens, B. E., & Handy, S. L. (2008). Built environment correlates of walking: A review. Medicine
and Science in Sports and Exercise, 40(7 Suppl), S550.
Saxena, S., Brémond, F., Thonnat, M., & Ma, R. (2008). Crowd behavior recognition for video
surveillance. International Conference on Advanced Concepts for Intelligent Vision Systems, 970–
981.
Schwartz, W. (2000). Bicycle and pedestrian data: Sources, needs, and gaps.
Shigematsu, R., Sallis, J. F., Conway, T. L., Saelens, B. E., Frank, L. D., Cain, K. L., Chapman, J.
E., & King, A. C. (2009). Age differences in the relation of perceived neighborhood environment to
walking. Medicine and Science in Sports and Exercise, 41(2), 314.
Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on image data augmentation for deep learning.
Journal of Big Data, 6(1), 1–48.
Smith, T. C., Wingard, D. L., Smith, B., Kritz-Silverstein, D., & Barrett-Connor, E. (2007). Walking
decreased risk of cardiovascular disease mortality in older adults with diabetes. Journal of Clinical
Epidemiology, 60(3), 309–317.
Sun, G., & Lau, C. Y. (2021). Go-along with older people to public transport in high-density cities:
Understanding the concerns and walking barriers through their lens. Journal of Transport & Health,
21, 101072.
Sung, H., Lee, S., & Cheon, S. (2015). Operationalizing jane jacobs’s urban design theory:
Empirical verification from the great city of seoul, korea. Journal of Planning Education and
Research, 35(2), 117–130.
Thom, N., & Hand, E. M. (2020). Facial attribute recognition: A survey. Computer Vision: A
Reference Guide, 1–13.
Thornton, C. M., Kerr, J., Conway, T. L., Saelens, B. E., Sallis, J. F., Ahn, D. K., Frank, L. D., Cain,
K. L., & King, A. C. (2017). Physical activity in older adults: An ecological approach. Annals of
Behavioral Medicine, 51(2), 159–169.
Tsunoda, K., Tsuji, T., Kitano, N., Mitsuishi, Y., Yoon, J.-Y., Yoon, J., & Okura, T. (2012).
Associations of physical activity with neighborhood environments and transportation modes in
older Japanese adults. Preventive Medicine, 55(2), 113–118.
Van Holle, V., Van Cauwenberg, J., Gheysen, F., Van Dyck, D., Deforche, B., Van de Weghe, N., &
De Bourdeaudhuij, I. (2016). The association between Belgian older adults’ physical functioning
and physical activity: What is the moderating role of the physical environment? PLoS One, 11(2),
e0148398.
Wang, R., Yuan, Y., Liu, Y., Zhang, J., Liu, P., Lu, Y., & Yao, Y. (2019). Using street view data and
machine learning to assess how perception of neighborhood safety influences urban residents’
mental health. Health & Place, 59, 102186.
Wang, X., Zheng, S., Yang, R., Zheng, A., Chen, Z., Tang, J., & Luo, B. (2022). Pedestrian attribute

recognition: A survey. Pattern Recognition, 121, 108220.
Weiss, K., Khoshgoftaar, T. M., & Wang, D. (2016). A survey of transfer learning. Journal of Big
Data, 3(1), 1–40.
Wichrowska, O., Maheswaranathan, N., Hoffman, M. W., Colmenarejo, S. G., Denil, M., Freitas,
N., & Sohl-Dickstein, J. (2017). Learned optimizers that scale and generalize. International
Conference on Machine Learning, 3751–3760.
Wu, S., Moore, B. E., & Shah, M. (2010). Chaotic invariants of lagrangian particle trajectories for
anomaly detection in crowded scenes. 2010 IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, 2054–2060.
Wu, Y., Zhang, D., & Kang, S. (2013). Physical activity and risk of breast cancer: A meta-analysis
of prospective studies. Breast Cancer Research and Treatment, 137(3), 869–882.
Xiang, J., Dong, T., Pan, R., & Gao, W. (2020). Clothing attribute recognition based on RCNN
framework using L-Softmax loss. IEEE Access, 8, 48299–48313.
Yang, Y., He, D., Gou, Z., Wang, R., Liu, Y., & Lu, Y. (2019). Association between street greenery
and walking behavior in older adults in Hong Kong. Sustainable Cities and Society, 51, 101747.
Yin, L. (2017). Street level urban design qualities for walkability: Combining 2D and 3D GIS
measures. Computers, Environment and Urban Systems, 64, 288–296.
Yin, L., Cheng, Q., Wang, Z., & Shao, Z. (2015). ‘Big data’for pedestrian volume: Exploring the
use of Google Street View images for pedestrian counts. Applied Geography, 63, 337–345.