Content uploaded by Xue Yang
Author content
All content in this area was uploaded by Xue Yang on May 24, 2021
Content may be subject to copyright.
Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=tgis20
International Journal of Geographical Information
Science
ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tgis20
Attributing pedestrian networks with semantic
information based on multi-source spatial data
Xue Yang, Kathleen Stewart, Mengyuan Fang & Luliang Tang
To cite this article: Xue Yang, Kathleen Stewart, Mengyuan Fang & Luliang Tang (2021):
Attributing pedestrian networks with semantic information based on multi-source spatial data,
International Journal of Geographical Information Science, DOI: 10.1080/13658816.2021.1902530
To link to this article: https://doi.org/10.1080/13658816.2021.1902530
View supplementary material
Published online: 30 Mar 2021.
Submit your article to this journal
Article views: 160
View related articles
View Crossmark data
RESEARCH ARTICLE
Attributing pedestrian networks with semantic information
based on multi-source spatial data
Xue Yang
a
, Kathleen Stewart
b
, Mengyuan Fang
c
and Luliang Tang
c
a
School of Geography and Information Engineering, China University of Geosciences, Wuhan, China;
b
Department of Geographical Sciences, University of Maryland, College Park, MD, USA;
c
State Key Laboratory
for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, China
ABSTRACT
The lack of associating pedestrian networks, i.e. the paths and roads
used for non-vehicular travel, with information about semantic attri-
bution is a major weakness for many applications, especially those
supporting accurate pedestrian routing. Researchers have developed
various algorithms to generate pedestrian walkways based on data-
sets, including high-resolution images, existing map databases, and
GPS data; however, the semantic attribution of pedestrian walkways
is often ignored. The objective of our study is to automatically extract
semantic information including incline values and the dierent cate-
gories of pedestrian paths from multi-source spatial data, such as
crowdsourced GPS tracking data, land use data, and motor vehicle
road (MVR) networks. Incline values for each pedestrian path were
derived from tracking data through elevation ltering using wavelet
theory and a similarity-based map-matching method. To automati-
cally categorize pedestrian paths into ve classes including sidewalk,
crosswalk, entrance walkway, indoor path, and greenway, we devel-
oped a hierarchical strategy of spatial analysis using land use data
and MVR networks. The eectiveness of our proposed method is
demonstrated using real datasets including GPS tracking data col-
lected by volunteers, land use data acquired from OpenStreetMap,
and MVR network data downloaded from Gaode Map.
ARTICLE HISTORY
Received 21 September 2020
Accepted 9 March 2021
KEYWORDS
Pedestrian networks;
semantic attribution; incline
values; pedestrian path
categorization; multi-source
spatial data
1. Introduction
Pedestrians, including individuals who travel on foot or tiny wheels, e.g. wheelchairs,
scooters, skateboards, etc., are usually recognized as a group of road users who are
vulnerable to dierent aspects of road use, such as limited access, injury, and crime,
especially in an outdoor environment (Zhang et al. 2012, Yang et al. 2020). To ensure
the safety and ecient travel of pedestrians, navigation applications and optimiza-
tion schemes should consider not only information about the geometry and con-
nectivity of pedestrian networks but also the semantic attribution of pedestrian
paths, such as incline values and the dierent categories of paths (John et al. 2017,
Sun et al. 2019). For example, a wheelchair user may not be able to climb a path
with over 10% incline; and a cyclist may choose a mountainous route for training or
CONTACT Luliang Tang tll@whu.edu.cn
Supplemental data for this article can be accessed here.
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE
https://doi.org/10.1080/13658816.2021.1902530
© 2021 Informa UK Limited, trading as Taylor & Francis Group
a at route for commuting (John et al. 2017). Urban planners need to know what
kinds of paths connect, for example, a shopping mall with a metro system, and
further estimate the walkability based on the category of pedestrian paths (Elias
2007, Sun et al. 2019). Semantic attribution of pedestrian paths including incline
value and path category is an essential foundation for routing tools, planners, and
geospatial analysts to better understand and support pedestrian movements.
Over the past decade, studies on pedestrian network information extraction have been
conducted using various databases, such as existing map data, social media data, and GPS
tracking data (Yang et al. 2020). Current approaches for generating these networks have
been summarized from three dierent perspectives: buering methods, image proces-
sing, and collaborative mapping (Karimi and Kasemsuppakorn 2013). Buering methods
have been used to extract the structure of sidewalks and crosswalks from existing MVR
networks (Kim et al. 2009, Ballester et al. 2011, Tal and Handy 2012, Guo et al. 2017). Image
processing methods have been another choice for extracting pedestrian networks with
dierent kinds of path types. Since the rst two methods require signicant post-
processing work such as eliminating route segments where people cannot walk and
lling gaps caused by missing segments that were shielded by trees, buildings, etc.,
a collaborative mapping approach was developed (Kasemsuppakorn and Karimi 2009,
2013b, Martin and Rob 2013). The crowdsourced geographic data used in collaborative
mapping are derived from a combination of local knowledge, eld notes, and ‘Armchair
mapping
1
’ as well. Using crowdsourced tracking data to extract pedestrian networks is
one type of collaborative data collection, however, most research has only focused on
topology and geometry detection, and has ignored semantic information extraction
associated with pedestrian networks, such as incline values and the type or category of
paths (Xie and Ou 2018, Yang et al. 2020). The incline values used for current routing tools
(e.g. OpenTripPlanner
2
) have been mainly extracted from publicly available elevation
datasets, such as Shuttle Radar Topology Mission (SRTM) and Advanced Spaceborne
Thermal Emission and Reection Radiometer (ASTER) data. However, John et al. (2017)
found that these elevation data sources were associated with three issues: a high cost of
data acquisition, data being available only for a limited set of locations, and insucient
horizontal resolution or vertical accuracy. Meanwhile, path category information used for
walkability analyses is still mostly dependent on manual identication (Elias 2007,
Kasemsuppakorn and Karimi 2013, Sun et al. 2019).
In this study, we present an approach for automatically extracting information on
the category and ne-granular incline value of pedestrian paths from multi-source
spatial data including crowdsourced GPS tracking data, land use data, and motor
vehicle road (MVR) network data. The steps include computing pedestrian road
incline values using three-dimensional (3D) GPS tracking data; and pedestrian path
categorization through combining land use and MVR network data, with existing
pedestrian networks. To improve the reliability for computing incline values, two
steps involving the preprocessing of crowdsourced GPS tracking data were needed.
The rst step was to lter the GPS trajectory elevation data using an approach based
on wavelet theory. Then, we matched the GPS data to the existing pedestrian
networks based on a similarity measurement algorithm. The incline values were
calculated based on existing gradient calculation formulas. The types or categories
of pedestrian paths, such as the sidewalk, crosswalk, entrance walkway, indoor path,
2X. YANG ET AL.
and greenway, were detected based on the hierarchical strategy of spatial analysis
(HSSA) method that used land use data and MVR network data. It should be
emphasized that pedestrian path categories including sidewalk and crosswalk are
dened based on the spatial relation between pedestrian paths and other physical
infrastructures such as MVR networks, buildings, and green land; the detailed deni-
tions are shown in Section 3.3. The main contributions of this paper include: (1) how
spatial datasets can be used to extract semantic information relating to pedestrian
networks including path incline values and path categories that lls a gap in studies
involving detailed pedestrian network mapping; (2) how including paths with poten-
tial risks for pedestrians identied based on road incline values benets the approach
by highlighting risks for users of these paths, especially for mobility-restricted indi-
viduals; and (3) the identication of categories of pedestrian paths based on the
proposed HSSA method with an average precision and recall of 87.22% and 91.63%
respectively.
2. Related work
Studies on road information extraction from spatial datasets, e.g. GPS tracking data,
images, videos, etc., have been attracting more attention in recent years. Many of these
works have been conducted using detailed MVR information mining, such as multi-level
MVR information extraction (Uduwaragoda et al. 2013, Ding et al. 2014, Tang et al. 2016,
Yang et al. 2018a, 2018b, Zhang et al. 2020); and MVR change detection (Rade et al. 2018,
Tang et al. 2019a). The details of multi-level MVR information extraction included road
shapes, connectivity, and semantic attribution acquisition, such as road boundary, lane
marking, turns, road type, road width, etc. In contrast, studies on pedestrian network
generation have been relatively few and most of them only explore automatic acquisition
techniques from the perspective of geometry and connectivity (Ballester et al. 2011,
Kasemsuppakorn and Karimi 2013, 2013b, Xie and Ou 2018, Yang et al. 2020). For many
pedestrian-related applications, semantic attribution of pedestrian networks including
incline values and category of the path is also important, e.g. for accurate pedestrian
routing especially for wheelchair users, cyclists, and elderly persons, etc. As few studies
have been conducted to detect incline values and categories of pedestrian networks from
dierent spatial datasets, the problems of information detection with insucient accuracy
and low automation remain (John et al. 2017).
With the evolution of the Web and the generalization of positioning techniques,
massive amounts of spatial data have been produced, e.g. GPS tracking data, social
media data with location pins, and crowdsourced road maps shared via OSM
(OpenStreetMap); and used for extracting various location-related information (Chin
et al. 2008, Ben et al. 2016, Zhang and Ye 2017, Gao et al. 2020). In contrast to the most
common sources of elevation data, e.g. data collected from satellite missions such as
TanDem-X, airborne LiDAR, and terrestrial surveying (John et al. 2017), crowdsourced
tracking data with elevation information are collected by soliciting contributions from
volunteers and are a low-cost and ecient way to extract and create semantic attributions
of pedestrian networks. While individuals are guaranteed to follow pedestrian paths,
a DEM might return elevation values that are a few meters away from the path and that
could already dier by several meters if the terrain is hilly. John et al. (2017) proposed
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 3
extracting road incline values from crowdsourced tracking data based on a segmentation
strategy. Specically, they split streets into segments at the intersection points with other
streets and then extracted the average incline value for a street segment using the
matched tracking points. However, this coarse-grained approach does not provide
detailed information on the slopes of roads, especially for long roads with greater
inclination uctuations. Thus, using an average incline value to represent the changing
characteristics of pedestrian road surfaces is insucient for pedestrian-based geospatial
applications.
3. Methodology for semantic attribution extraction of pedestrian paths
3.1. Scope and overall idea
Semantic information for pedestrian paths such as incline values and path categories can
be extracted from multi-source spatial data, as shown in Figure 1.
In this study, two kinds of spatial data were used to extract the semantic information
relating to pedestrian networks. The rst was crowdsourced GPS tracking data collected
by volunteers using mobile positioning devices (e.g. mobile phones, hand-held GPS
Figure 1. The architecture of semantic information extraction for the pedestrian network using multi-
source spatial data.
4X. YANG ET AL.
devices) that were used to extract incline values for the pedestrian paths. Specically,
a crowdsourced trajectory is comprised of a set of corresponding tracking points, denoted
as T= (p
1
, . . ., p
n
), where n is the number of tracking points belonging to the trajectory.
Each tracking point is represented by p (x, y, z, t), where x, y, z, and t are the longitude,
latitude, elevation, and time stamps, respectively, for a tracking point. The second kind of
spatial data used in this study was land use data and MVR network data, which were
acquired from OSM and Gaode Map respectively. The land use data used in this study
included boundary information for green spaces and buildings within the study area. The
MVR network data included information on roads that captured the plane position of road
segments and nodes, connectivity of each segment, the number of lanes, driving direc-
tion constraints (e.g. one-way, or two-way), and road length. The land use data and MVR
network data together were applied to automatically categorize pedestrian networks.
3.2. Pedestrian path incline values extraction using crowdsourced tracking data
Existing positioning systems such as GPS (global positioning system) do not work per-
fectly in all locations, especially in urban areas. Some outliers caused by tall buildings,
shadowing, and multi-path issues, are mixed in with the raw positioning results. For
crowdsourced tracking data with 3D positioning information, there are positional errors
for both the spatial locations (x, y) and the elevations (z). To reduce the high-frequency
noise that may be present in the elevation data (John et al. 2017), we applied a wavelet
denoising method because of its eectiveness for removing high-frequency noise (Sardy
et al. 2001). Based on the denoising theory of the wavelet method, high-frequency noises
were mainly present in the high-frequency signal component. Thus, after signal decom-
position is applied, the low frequency component was used to reconstruct a ltered
elevation dataset. The specic value of the resolution levels during wavelet denoising
will be described in detail in the case study section of this paper. The second preproces-
sing step for incline values computation was map-matching. This was used to remove the
outliers present in the spatial locations (x, y); and assisted in computing the incline values
of the pedestrian paths. That is, we needed to get the elevation of the pedestrian path rst
based on the matched tracking points and then compute the incline values.
The base map for the pedestrian networks used during map-matching including
geometry and connectivity information was derived from crowdsourced tracking data
using a method proposed in a previous study (Yang et al. 2020). The geometry and
connectivity for pedestrian networks are usually similar but simpler than those for MVR
networks (Yang et al. 2020). Based on this consideration, we applied a similarity-based
map-matching algorithm after reviewing the existing map-matching methods proposed
for MVR networks (Yang et al. 2018b). Compared with existing map-matching methods
(e.g. probabilistic modeling), a similarity-based method was less complex and oered
more exibility concerning similarity modeling. For this study, the similarity between
tracking points and pedestrian paths was calculated based on two criteria: (1) the vertical
distance between the tracking point and the pedestrian path; (2) the angle dierence
between the tracking vector and the pedestrian path. The calculation of the similarity
between the GPS tracking data and pedestrian paths followed the method proposed by
Yang et al. (2018a). The specic values of weights of ω
1
and ω
2
, constant D, and similarity
threshold Ts, for similarity computation during map-matching are discussed further in the
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 5
case study section. The detailed steps for the similarity-based map-matching method are
shown in the Appendix. After map-matching, all tracking points were categorized into
two types: (1) successfully matched points and (2) unsuccessfully matched points.
Tracking points successfully matched to pedestrian paths were used to compute the
incline value of the corresponding pedestrian path based on its elevation, and unsuccess-
fully matched points were regarded as planar drifting points and removed.
To keep the details of incline values for pedestrian paths, we partitioned the pedestrian
path segment ps
j
into a series of sub-segments (denoted as: sps
m
) based on a partition size
(denoted as: α) during the incline values computation. The value of partition size α for any
pedestrian road segment ps
j
was determined based on the planar positioning accuracy of
tracking points.
For a sub-segment sps
m
, there may be some matched tracking points p
1
, p
2
, . . . p
i
, as
shown in Figure 2(a). We computed the average elevation value of these matched
tracking points and this value was used as the elevation information for sub-segment
sps
m
. The elevation data used in this paper was based on the WGS-84 Ellipsoid height (see
Figure 2(b)). Based on existing methods for computing inclines (John et al. 2017), the
incline value between pedestrian path sub-segment sps
m
and sps
m+1
was dened as:
im¼100%�Δem
dis cm;cmþ1
ð Þ (1)
where Δe
m
was the elevation dierence between pedestrian path sub-segment sps
m
and
sps
m+1
; dis(c
m
, c
m+1
) was the Euclidean distance between the sub-segment center point c
m
and c
m+1
, as shown in Figure 2(b). Also, note that the coverage of tracking data for
a pedestrian path mainly depended on the ow of pedestrians who were traveling on
it. This would cause a high coverage of trajectories for some pedestrian paths, while
others had only a few or even no matching points. Therefore, the incline values for some
sub-segments without matching points were inferred based on their adjacent sub-
segments.
3.3. Automatic identication for pedestrian path categories based on HSSA
method
Based on previous studies (Kasemsuppakorn and Karimi 2013, Zhou et al. 2015), the
categories of pedestrian paths in this study were dened as the sidewalk, indoor path,
entrance walkway, greenway, and crosswalk. The denition of indoor paths was essen-
tially the same as used in previous research (Kasemsuppakorn and Karimi 2013, Zhou et al.
2015). Sidewalks were dened as paths that were next to the MVR network. An entrance
walkway was dened as the path between the intersection of pedestrian segments and
the entrance of a building (Figure 3(a)). A greenway was dened as a path that is in
a green space and where the path does not belong to any other type. A crosswalk was
a path that pedestrians use to cross a road from one side to the other. In this study, we
used a circular buer to visualize a crosswalk because it is very hard to get the specic
width of a crosswalk due to the mobility of pedestrians (see Figure 3(b)). The radius
(denoted as r) of the circular buer for a crosswalk was decided based on the road width
(denoted as wml) of MVR networks and pedestrian path width (denoted as ε). Typically,
6X. YANG ET AL.
sidewalks are set on both sides of most MVR networks, and the radius r can be computed
based on Equation (2).
r¼wml þ2ε
2(2)
To detect these categories, we designed an HSSA method that used three steps. As the
rst step entrance walkways and indoor paths were detected based on spatial overlay
analysis of polygon data for buildings and the pedestrian networks. Then, crosswalks and
sidewalks were extracted from the rest of the pedestrian paths by using MVR networks. As
the third step, we detected greenways from the spatial overlay results of polygon data for
Figure 2. Inclination computation based on the matched tracking points and the different colors of
points and lines in the above two panels present the different trace segments and pedestrian path
segments, respectively; (a) the average value computation of elevation data; (b) inclination
calculation.
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 7
green spaces with the remaining pedestrian paths. The detailed operation for the HSSA
method is shown in the Appendix. By applying these steps, the raw pedestrian networks
were enriched with the semantics of the dierent path categories, i.e. indoor path,
crosswalk, entrance walkway, sidewalk, and greenway. Note that the overlay computation
of pedestrian paths with polygon data (e.g. buildings and green spaces) was conducted
by using the built-in function ‘inpolygon’ of the MATLAB 2018a platform. The details of
parameter value setting are discussed in more detail in the case study section.
4. Case study: semantic information extraction for pedestrian networks
The proposed approach for semantic enhancement of pedestrian networks was tested
using real-world multi-source spatial data. These spatial data included: (1) 3D tracking
data generated by pedestrians in the City of Wuhan in China (see Figure 4); (2) land use
data obtained from the OSM platform (see Figure 5(a)); and (3) MVR networks down-
loaded from Gaode platform through a public API, as shown in Figure 5(b). The tracking
data used in this study were collected for two weeks in 2016 by about 83 participants,
using built-in positioning devices in mobile phones. There were about 138,863 tracking
points with approximately 10–15 m planar positioning accuracy and 1–10s sampling
intervals. The study site was in the north of Hongshan district of the City of Wuhan, an
area of approximately 5 square kilometers, which had an undulating terrain and con-
tained many dierent kinds of pedestrian paths, such as sidewalk, crosswalk, entrance
walkway, indoor path, and greenway. The land use information for buildings and green
Figure 3. Categories of pedestrian paths, (a) the entrance walkway of pedestrian networks; (b) the
crosswalk of pedestrian networks.
8X. YANG ET AL.
spaces was represented by polygons, as shown in Figure 5(a). MVR networks were stored
based on an arc-node model, including road attributes and data on the number of lanes,
driving direction constraints (e.g. one-way or two-way), and road length (Figure 5(b)).
4.1. Data preprocessing for crowdsourced tracking data
The method for road incline values computation was tested with preprocessed data
through the steps of elevation data ltering and map-matching. For the methods
described earlier, we used the wavelet denoising tools provided by MATLAB 2018a to
improve the certainty of elevation data. The Symlet wavelet (denoted as ‘Sym5ʹ at
MATLAB platform) was selected as the wavelet basis following an earlier study
(Soleymani et al. 2017). To estimate the quality of data ltering under dierent decom-
position and reconstruction levels, two kinds of approaches were conducted. The rst one
was to compute the elevation dierences between the ltered tracking data and the
ground truth data. Because the high-resolution elevation data was not available for public
use, we used ASTER GDEM (Global Digital Elevation Model) data with a 30 m spatial
resolution to verify the eectiveness of elevation data ltering (Figure 6(a)). To facilitate
the computation of elevation dierences, the tracking data was converted to a raster
format and had the same spatial resolution as the ASTER DEM data (see Figure 6(b)).
The second evaluation step was to compare the elevation of tracking points with or
without wavelet denoising using PSNR (Peak signal to noise ratio) and SSIM (Structural
similarity) indicators. PSNR was used to quantize the distortion of ltered elevation data
under dierent decomposition and reconstruction levels (Sheikh et al. 2006). The higher
the PSNR value, the smaller the dierence between the raw data and processed data. We
adopted SSIM to estimate the structural similarity of the raw and processed data (Wang
et al. 2004). Generally, the value of SSIM ranges from 0 to 1; and the higher the value, the
better the quality of processed data. The values of PSNR and SSIM for ltered data at the
specic level of denoising are computed based on the Equations shown in Appendix.
Table 1 shows the elevation results between the ltered elevations of tracking points and
the DEM data at dierent levels of decomposition and reconstruction. The mean and
Figure 4. Crowdsourced tracking data collected by volunteers in the City of Wuhan, (a) 3D perspective
of the tracking points; (b) 2D perspective of the tracking points.
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 9
Figure 5. Land use data and MVR networks overlaid with pedestrian networks, (a) land use data; (b)
MVR networks.
10 X. YANG ET AL.
standard deviation of elevation dierences between DEM data and the corresponding
raw tracking data were about 12.8911 (m) and 51.4372 (m), respectively. The experimental
results showed that the mean and standard deviation of the ltered elevation data were
improved after wavelet denoising. The processed elevation data did not dier greatly for
dierent decomposition and reconstruction levels.
Figure 6. ASTER DEM and ASTER-derived elevation data on pedestrian trajectories in the experimental
region; (a) DEM data with 30 m spatial resolution; (b) elevation data of tracking data converted to
raster format.
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 11
In Table 2, we nd that the average running time of tracking data ltering increased
with an increasing decomposition level. For a specic decomposition level, the values of
PSNR and SSIM gradually got larger and smaller respectively with an increasing recon-
struction level. The experimental results showed that the value of PSNR and SSIM of
ltered data was associated with the reconstruction level. The value of PSNR and SSIM for
ltered data at the second reconstruction level was the same for dierent decomposition
levels. That means the quality of ltered tracking point elevation data depended mainly
on the reconstruction level. Combined with the results of Table 1 and running time, the
optimized decomposition level and reconstruction level in this study were each set as 2,
as a tradeo.
Map-matching was the second step for preprocessing the tracking data. The unsuc-
cessfully matched points were regarded as planar outliers and removed. The rest of the
matched points were used to extract the incline values of the pedestrian paths. To fulll
the task of map-matching, we rst calculated the similarity between tracking points and
pedestrian paths. As pedestrians can walk in any direction and can freely change their
direction, therefore, using research results from a previous study (Yang et al. 2018c), the
weights for distance and angle of similarity in the evaluation model were set to 0.91 and
0.09; respectively. Since the width of a pedestrian path in an urban area should be
between 2.5 m and 3.0 m (Yang et al. 2020) and the planar positioning accuracy of
tracking data was about 10–15 m, the constant D was set to 13 m. The similarity threshold
(denoted as Ts) was used to decide whether a tracking point could be matched to the
current pedestrian road segment. Its value was 0.8058 when the distance and angle were
3 m and 0 degrees or 180 degrees, respectively. To obtain an optimal value of Ts, we
randomly selected 20 trajectories from the tracking dataset and computed the values of
Table 1. Elevation differences between filtered data and DEM data at different levels of decomposition
and reconstruction.
Wavelet basis Decomposition level Reconstruction level Mean (m) STD (m)
Sym 5 (Symlet wavelet) 2 1 10.6537 8.7003
2 10.6558 8.6095
3 1 10.6680 8.6552
2 10.6558 8.6095
3 10.6836 8.6387
4 1 10.6680 8.6552
2 10.6558 8.6095
3 10.6836 8.6387
4 10.6517 8.5473
Table 2. Evaluation results of RSNR and SSIM for filtered data at different levels of decomposition and
reconstruction.
Wavelet basis Decomposition level Reconstruction level Running Time (s) PSNR SSIM
Sym 5 (Symlet wavelet) 2 1 0.6875 15.3306 0.8741
2 0.7344 26.4584 0.7328
3 1 1.1250 15.3306 0.8741
2 0.9531 26.4584 0.7328
3 0.9063 34.4572 0.6020
4 1 1.3750 15.3306 0.8741
2 1.5000 26.4584 0.7328
3 1.0313 34.4572 0.6020
4 0.8438 40.6113 0.5004
12 X. YANG ET AL.
indicators λ
1
and λ
2
of matching results by manual inspection. The values of λ
1
and λ
2
were
obtained based on Equations (9) and (10) shown in the Appendix.
Table 3 shows the results for Ts through repeated experiments using 1, 685 tracking
points collected in the City of Wuhan. Based on manual inspection, there were about 932
tracking points that should have been matched to the pedestrian paths. The rest of the
tracking points were regarded as outliers because of signal drifting. In Table 3, the relation
exhibits a parabolic trend between Ts and λ
1
. The relation between Ts and λ
2
displayed
a complete reversal trend to that with λ
1
. The value of λ
1
started to fall when the value of
Ts was decreased, even as the value of λ
2
grew. The values of λ
1
reached a peak when Ts
was set to 0.8–0.9. To reduce the uncertainty of road inclination computation and ensure
the integrity of the matched tracking points, the value of Ts was set as 0.8058, as
a tradeo.
4.2. Inclination computation and analysis for pedestrian paths
The semantic information about pedestrian networks was enhanced by adding the
attributes of pedestrian paths to the original database, including path incline values as
well as the semantic categories of the paths. The incline values of paths were calculated
based on the proposed partitioning strategy. Specially, the partition size α for a pedestrian
path was set to 10 m based on the planar positioning accuracy of the tracking data. The
objective of the inclination computation was to identify steep inclines for those path users
who need this information. Based on the design standards of road grades in China, the
longitudinal slope of the main road for pedestrians in a residential area should be less
than 8%. In a hilly area, the longitudinal slope of main roads for pedestrians should be less
than 12%; otherwise, anti-skid treatment should be done. The main roads for pedestrians
should not have stairs. When stairs were necessary, the longitudinal slope of stairs should
be less than 36%. For each branch of the main road, the longitudinal slope is recom-
mended to be less than 18%. Steps should have anti-skid treatments if the slope of
pedestrian paths exceeds 58%. The experimental data used in this study were collected
in the City of Wuhan where the main terrain is mostly plains, hills, and small to medium
relief mountains (Figure 6(a)). To facilitate the identication of pedestrian paths with
potential risks, we classied the inclination of paths into six levels by combining the
design standards for pedestrian roads with the inclination values (Figure 7(a)). In addition,
the design standards of pedestrian road inclination showed that the optimum slope for
setting steps in the hilly area ranged from 23% to 38%. Pedestrian paths with over 38%
inclination are displayed in Figure 7(b).
Table 3. Evaluation of map-matching results for different thresholds.
Ts N NG N
1
N
2
λ
1
λ
2
0.5 1,685 932 932 0 55.31% 100%
0.55 1,644 932 932 0 56.69% 100%
0.6 1,581 932 932 0 58.95% 100%
0.65 1,520 932 919 13 60.46% 100%
0.7 1,393 932 907 25 65.11% 98.61%
0.75 1,152 932 898 34 77.95% 96.35%
0.8 870 932 870 62 100% 93.35%
0.85 805 932 805 127 100% 86.34%
0.9 691 932 691 241 100% 74.12%
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 13
Figure 7. Visualization of inclination of pedestrian paths, (a) six inclination classes; (b) pedestrian paths
with over 38% inclination.
14 X. YANG ET AL.
The total length of pedestrian paths with incline values was about 39,424 m. In Table 4,
we can see that the highest proportion of pedestrian paths had less than 8% inclination
(57.05%). That means most pedestrian paths in the experimental area were at and
suitable for walking. Approximately 8% to 12% of pedestrian paths had incline values
greater than 8.88% because the experimental region was in a hilly area, which was a little
steeper than pedestrian paths with less than 8% inclination. Also, about 25.77% pedes-
trian paths were shown to need steps, and 13.13% of paths should have the anti-skid
treatment. As shown in Figure 7(b), pedestrian paths with over 38% inclination were
marked by red lines. Pedestrians should take care when using these roads, especially for
individuals who are mobility-restricted, such as wheelchair users or people with walking
aids. Based on these incline values, routing tools can customize walking routes for
pedestrians based on their own needs.
We also randomly selected 30 path segments with over 38% inclination and checked if
these paths have anti-skid treatments for steps by manual visual inspection with street
view images of Baidu Map. The results show that about 60% of pedestrian paths with over
38% inclination were steps and had anti-skid treatments such as handrails guardrail. The
rest of them were relatively at and not the steps. That means the inclination of these
pedestrian paths is overestimated, because of GPS drift and elevation uctuations.
Therefore, the improvement of elevation data using other types of sensors such as built-
in barometers in mobile phones could still benet from further study in the future.
4.3. Categories information extraction for pedestrian paths
In this study, ve types of pedestrian paths were automatically detected using the
proposed HSSA method, including indoor paths, entrance walkways, sidewalks, cross-
walks, and greenways (Figure 8(a)). Indoor paths and entrance walkways were identied
as part of the rst step based on overlay results between pedestrian networks and
buildings (Figure 8(b)).
For sidewalk and crosswalk identication, the values of road width wml
t
from MVR
networks, and GPS trajectories positioning error ε were required. The MVR networks used
in this study recorded the number of lanes and driving direction constraints (e.g. one-way,
or two-way) for road segments, however, not all road segments had this information,
especially for residential roads. For main urban roads, road width was computed by
multiplying the values for lane numbers and lane width. For some residential roads, we
needed to infer their road widths based on road construction standards and driving
direction constraints. Road construction standards in China indicate that the widths for
one-way and two-way roads range from 3.5 m to 5 m and 8 m to 12 m, respectively. The
Table 4. Statistics of pedestrian paths at different levels of
inclination.
Incline range Total length (m) Proportion
Less than 8% 22491.39 57.05%
8%-12% 3500.85 8.88%
12%-18% 3272.19 8.30%
18%-38% 4983.19 12.64%
38%-58% 2132.84 5.41%
More than 58% 3043.53 7.72%
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 15
width of each lane in an urban area is about 3.5 m. Based on these standards, the widths
for one-way and two-way roads in this research were set to 5 m and 8 m, respectively. The
value of ε was set to 2.5 m. Figure 8(c) shows the result of sidewalk and crosswalk
identication based on the values of these parameters. We can see from Figure 8(c) that
most of the sidewalks were identied using the proposed HSSA method. Besides, the
extraction of greenways was set in the last step since some paths could be other types
even though they were in an area of green space, as shown in Figure 8(d).
Figure 8. The study results: (a) five classes of pedestrian path were extracted; (b) indoor path and
entrance walkway detection results; (c) sidewalk and crosswalk detection results; (d) greenway
identification results.
Table 5. Statistical results for each detected category of pedestrian paths.
Category of pedestrian paths Total length (m) Proportion (%)
type1: indoor path 2373.63 1.93
type2: entrance walkway 761.65 6.02
type3: sidewalk 16115.62 40.88
type4: crosswalk 821.00 2.08
type5: greenway 5438.93 13.80
other 13913.33 35.29
16 X. YANG ET AL.
The statistics for the dierent path categories within the pedestrian network are
displayed in Table 5. The evaluation indicators included the length of paths in each
category and its proportion to the total length of the entire pedestrian networks. As we
can see from Table 5, sidewalks occupied the highest proportion of paths compared to
other types of pedestrian paths. It should be noted that pedestrian paths identied as
other were either located in separate areas or were missed during category identication.
According to the experimental results, the proportion of other paths was about 35.3%,
and lower than that of sidewalks. We found that the proportion of pedestrian paths of
type1 (indoor paths) was the lowest of all. That is partly because the GPS signal is lost when
pedestrians walk into buildings. How to extract a complete indoor pedestrian map is
another open research challenge that would benet from further study.
To further verify the eectiveness of category identication, we evaluated the detected
results by comparing our results with ground truth data. As shown in Table 6, two
evaluation indicators, i.e. Precision and Recall were calculated using the method of Yang
et al. (2018a). The parameters True positive, False positive, and False negative referred to the
length of pedestrian paths correctly detected, wrongly detected, and missed by the
methods proposed in this paper, respectively. Specially, the lengths of pedestrian paths
correctly detected or missed were respectively measured by using the measurement tools
in QGIS 2.18. The ground truth for pedestrian networks in the study area in the City of
Figure 9. Misclassification: (a) incorrect indoor path identification; (b) incorrect sidewalk identification.
Table 6. Evaluation of category identification of pedestrian paths.
Category of pedestrian paths True positive (m) False positive (m)
False negative
(m) Precision (%) Recall (%)
type1: indoor path 2048.73 324.90 318.00 86.30 86.56
type2: entrance walkway 611.40 150.26 129.72 80.27 82.50
type3: sidewalk 15739.79 375.84 1204.72 94.66 92.89
type4: crosswalk 695.20 125.80 55.30 84.68 92.63
type5: greenway 4975.38 463.58 250.33 91.48 95.21
other 11955.27 1958.06 0.00 85.93 100.00
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 17
Wuhan was obtained by manual visual inspection of online areal images found on Google
Maps.
The evaluation results in Table 6 illustrate that the methods proposed in this paper
were eective at identifying categories of pedestrian paths at an average precision and
recall of 87.22% and 91.63%. However, based on the results in Table 6, there was about
a 13% chance of incorrectly identifying a pedestrian path category, and about 9% of the
paths were not categorized. In urban areas, it can be a challenging task to correctly
recognize the category of all pedestrian paths because of the accuracy of multi-source
spatial data and the randomness of pedestrians with respect to where they walk, as shown
in Figure 9.
Figure 9 shows two typical examples of incorrect path category identication. In Figure 9(a),
some paths were wrongly identied as indoor paths because of positioning errors of traces
even though they belonged to sidewalks. Moreover, some pedestrians walked close to
buildings, compounding this error. Since the entrance walkway was identied at the same
layer as indoor paths, it was also easy to get an incorrect path classication due to an error in
the detection of indoor paths (see Figure 9(a)). The additional processing needed for improv-
ing the detection accuracy of pedestrian paths is a topic for future research. Beyond that, the
spatial data applied in this study were collected from multi-platforms (e.g. volunteer crowds,
OSM platform, and Gaode Map), which caused datasets to have their own reference systems
which made it dicult to integrate all datasets into a common reference space without some
risk of distortion. Meanwhile, some spatial datasets could be transformed to protect the privacy
of users’ positions that led to local deformation. Figure 9(b) shows a transformation failure for
MVR networks, where a part of MVR roads was wrongly overlaid with buildings. This partial
position error with the MVR networks resulted in sidewalks being wrongly identied as other,
and decreased its recall score. Further analyses and improvements are therefore needed.
Overall, the statistical results shown in the above tables veried that the proposed method
in this paper could be applied for enriching the attribution of pedestrian networks. Enhanced
Figure 10. Incline analysis for pedestrian bridges and pedestrian tunnels.
18 X. YANG ET AL.
pedestrian networks with road inclination and semantic categories could be used to better
assess the walkability of a region, recommending a personalized route for pedestrians, and
assisting in decision making for pedestrian path construction.
4.4. Discussion
In this study, we explored how to automatically extract semantic information includ-
ing incline value and path category from multi-source spatial data. These two kinds
of semantic attribution are fundamental for pedestrian-related applications but are
rarely discussed (for an exception see John et al. (2017)). Building on the earlier work
of John et al. (2017), we rened the computation task by partitioning path segments
into a series of sub-segments that made the detection results for path incline more
granular, which results in more accurate pedestrian routing. We developed an auto-
matic categorization method for acquiring the type of pedestrian paths using land
use data and MVR networks, which signicantly enhanced the eciency of pedes-
trian paths’ category identication when comparing with categories manually identi-
ed. Sun et al. (2019) also investigated factors that aect the walkability of
pedestrian networks using manual digitization results to derive pedestrian paths’
categories from existing topographic maps. Our work goes further though as our
approach also oers a low cost and ecient solution to extract semantic information
relevant for pedestrian paths from public data sources that can expand the set of
data acquisition sources for pedestrian-related applications.
In this study, ve types of pedestrian paths were automatically identied by the
proposed HSSA method. These ve types extend the work of Kasemsuppakorn and
Karimi (2009) who also investigated path types and who dened pedestrian bridges
and tunnels associated with pedestrian paths. Our research did not include pedes-
trian bridges and tunnels as these features usually cross the MVR from above and
below, as shown in Figure 10. The variation of slope both for pedestrian bridges and
tunnels follows the principle of steep rst, then at, and steep again. Since the
positional accuracy of crowdsourced GPS tracking data is limited, it is very dicult
to detect this subtle change in road incline. It was challenging to accurately identify
pedestrian tunnels due to GPS signal loss. This was also an issue for complex MVR
intersections detection such as overpass and cloverleaf intersections (Yang et al.
2018a). Using other sensor data such as Street View (e.g. Google Street View) or
high-denition images, and spatial information for existing physical infrastructures
such as trac lights and subway stations to address this issue could be explored in
future work.
As part of this research, we extracted semantic information about pedestrian paths
from open data that were characterized using low-cost and highly accessible crowd-
sourced data. However, these data also have quality issues that resulted in some
uncertainty with information mining, such as data completeness. Crowdsourced GPS
tracking data are collected by volunteers, and its coverage is mainly dependent on
the number of participants and their movement area. In this study, the crowdsourced
GPS tracking data covered about 80.1% of paths of the experimental area, and the
rest of the paths were neglected. This issue was also discussed by Karimi and
Kasemsuppakorn (2013) who indicated that the quality of pedestrian network
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 19
information extraction was heavily dependent on the coverage and accuracy of GPS
tracking data. Similarly, the land use data downloaded from OSM were also acquired
in a collaborative way. The quality of this data from the perspective of completeness
and diversity patterns has been discussed in many studies (see, for example,
Arsanjani 2015, Wang et al. 2020). It is still an open question how to balance the
conict between data acquisition costs and the quality of data sources. Related to
this, the question of what percentage of routes would be accessible is an interesting
topic for the eld of path planning (Cui 2016, Zimmermann et al. 2017), but not
addressed in this study.
5. Conclusion and future work
Semantic attribution of pedestrian networks is essential for a variety of applications,
especially for pedestrian navigation systems and walkability assessments. In the
absence of approaches and techniques for semantic attribution extraction asso-
ciated with pedestrian networks, this study focused on automatically extracting
incline values and categories of paths using multi-source spatial datasets. To
acquire incline values, 3D crowdsourced tracking data was applied. The categories
of pedestrian paths including sidewalk, crosswalk, entrance walkway, indoor path,
and greenway were identied based on a proposed HSSA method, using land use
data, MVR networks, and a pedestrian network base map. Case studies were con-
ducted using three kinds of spatial datasets including GPS tracking data collected
by volunteers in the City of Wuhan, China, land use data acquired from
OpenStreetMap, and MVR networks downloaded from Gaode Map. Based on the
experimental results of road inclination computation, we mapped pedestrian paths
based on the incline values. These pedestrian paths attributed with incline informa-
tion can be used as foundational data for routing tools or walkability analyses. For
path category identication, the evaluation results indicated that the proposed
HSSA method was eective, with an average precision and recall of 87.22% and
91.63% respectively.
In the real world, however, the environment for pedestrian paths and walkways
is complex and their design is widely varying. For instance, some pedestrian paths
are at but with stairs or curbs. In this situation, it can be dicult for some
pedestrians to travel along these paths, e.g. wheelchair users or people with
physical disabilities. Although we extracted the incline values of pedestrian paths,
stairs or curb identication was challenging and could be a topic for future study. It
was also dicult to identify jaywalking paths and marked crosswalks from all
crosswalks; detect whether there is a physical sidewalk adjacent to a road, and
recognize pedestrian bridges and tunnels. For many pedestrian-related applications,
such as assessing the overall public safety of pedestrian networks, accurate pedes-
trian routing especially for accessibility of vulnerable populations (e.g. the elderly,
people with strollers, physical disabilities, etc.), these unidentiable features are
very important and further research is still needed. Future work could address other
limitations including: (1) improving the accuracy of collected elevation data using
built-in sensors such as barometers in mobile devices to assist positioning; (2)
20 X. YANG ET AL.
extending spatial analysis algorithms with other sensor data (e.g. Street View, or
high-denition images), to increase the accuracy of path identication results as
well as the number of categories of pedestrian paths; and (3) further analysis of the
walking environment (e.g. safety, cleanliness, and greenness), and connectivity with
other trac networks (e.g. MVR networks and bicycle networks) that are also
relevant for pedestrian travel.
Notes
1. https://wiki.openstreetmap.org/wiki/OpenTripPlanner
2. https://wiki.openstreetmap.org/wiki/Armchair_mapping
Acknowledgments
The authors would like to sincerely thank the anonymous reviewers for their constructive comments
and valuable suggestions to improve the quality of this article.
Data and codes availability statement
The data and code that support the ndings of this study are available in [gshare.com] with the
identier(s) at the link (https://doi.org/10.6084/m9.gshare.12660467.v2).
Disclosure statement
No potential conict of interest was reported by the author(s).
Funding
This work was funded by the National Natural Science Foundation of China [No. 41901394,
41971405]; Open research fund program of LIESMARS, Wuhan University [No. 19S01].
Notes on contributors
Xue Yang received the Ph.D. degree from Wuhan University, Wuhan, China, in 2018. She is currently
an associate Professor with China University of Geosciences, Wuhan. Her research interests include
intelligent transportation system, spatiotemporal data analysis, and information mining.
Homepage: http://grzy.cug.edu.cn/yangxue1/zh_CN/index.htm
Email: yangxue@cug.edu.cn
Kathleen Stewart is currently a Professor in the Department of Geographical Sciences and Director
of the Center for Geospatial Information Science. She works in the area of geographic information
science with a particular focus on geospatial dynamics. She is interested in mobility and spatial
access, often in a big geospatial data context and using approaches that lie in the expanding eld of
spatial data science. Homepage: https://geog.umd.edu/facultyprole/stewart/kathleen
Email: stewartk@umd.edu
Mengyuan Fang received Bsc degree from Wuhan University, Wuhan, China, 2014. He is currently
a Ph.D candidate at the State Key Laboratory of Information Engineering in Surveying, Mapping and
Remote Sensing, Wuhan University. His research addresses the issue of trac congestion detection
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 21
and prediction using big trace data.
Email: myfang@whu.edu.cn
Luliang Tang received the Ph.D. degree from Wuhan University, Wuhan, China, in 2007. He is currently
a Professor with Wuhan University. His research interests include space–time GIS, GIS for transporta-
tion, and change detection. Homepage: http://www.lmars.whu.edu.cn/index.php/js/298.html
Email: tll@whu.edu.cn
References
Arsanjani, J.J., 2015. Quality assessment of the contributed land use information from
OpenStreetMap versus authoritative datasets. In: J.J. Arsanjani, et al., eds. OpenStreetMap in
GIScience. Cham: Springer, 37–58.
Ballester, M., Pérez, M., and Stuiver, H., 2011. Automatic pedestrian network generation. In:
Proceedings of the 14th AGILE international conference on geographic information science, 18–22
April. Utrecht, Netherlands, 1–13.
Ben, J., Trisalyn, N., and Meghan, W., 2016. Mapping ridership using crowdsourced cycling data.
Journal of Transport Geography, 52, 90–97. doi:10.1016/j.jtrangeo.2016.03.006
Chin, G.K.W., et al., 2008. Accessibility and connectivity in physical activity studies: the impact of
missing pedestrian data. Preventive Medicine, 46 (1), 41–45. doi:10.1016/j.ypmed.2007.08.004.
Cui, J.X., 2016. Detecting urban road network accessibility problems using taxi GPS data. Journal of
Transport Geography, 51, 147–157. doi:10.1016/j.jtrangeo.2015.12.007
Ding, Y., et al., 2014. Inferring road type in crowdsourced map services. Pattern Recognition, 8422 (8),
392–406.
Elias, B., 2007. Pedestrian navigation - creating a tailored geodatabase for routing. In: Workshop on
Positioning, Navigation & Communication, 22–22 March. Hannover, Germany, 41–47.
Gao, J., et al., 2020. Understanding urban hospital bypass behaviour based on big trace data. Cities,
103, 1–12. doi:10.1016/j.cities.2020.102739
Guo, Q., et al., 2017. The eect of road network patterns on pedestrian safety: a zone-based Bayesian
spatial modeling approach. Accident Analysis and Prevention, 99, 114–124. doi:10.1016/j.
aap.2016.11.002.
John, S., et al., 2017. Deriving incline values for street networks from voluntarily collected GPS traces.
Cartography and Geographic Information Science, 44 (2), 152–169. doi:10.1080/15230406.2016.1190300.
Karimi, H. and Kasemsuppakorn, P., 2013. Pedestrian network map generation approaches and
recommendation. International Journal of Geographical Information Science, 27 (5), 947–962.
doi:10.1080/13658816.2012.730148.
Kasemsuppakorn, P. and Karimi, H., 2009. Pedestrian network data collection through
location-based social networks. In: International conference on collaborative computing: network-
ing, applications and worksharing, 11–14 November. Washington, DC: IEEE, 1–9.
Kasemsuppakorn, P. and Karimi, H., 2013b. A pedestrian network construction algorithm based on
multiple GPS traces. Transportation Research Part C: Emerging Technologies, 26 (1), 285–300.
doi:10.1016/j.trc.2012.09.007.
Kasemsuppakorn, P. and Karimi, H.A., 2013. Pedestrian network extraction from fused aerial imagery
(Orthoimages) and laser imagery (Lidar). Photogrammetric Engineering and Remote Sensing, 79 (4),
369–379. doi:10.14358/PERS.79.4.369.
Kim, J., Bang, Y., and Yu, K., 2009. Automatic derivation of a pedestrian network based on existing
spatial data sets. In: ASPRS/MAPPS 2009 Fall Conference, 16–19 November. San Antonio, TX, 1–7.
Martin, D. and Rob, K., 2013. Crowdsourced cartography: mapping experience and knowledge.
Environment and Planning, 45 (1), 19–36. doi:10.1068/a44484.
Rade, S., et al., 2018. Road network fusion for incremental map updates. In: 14th International
Conference on Location Based Services, 15–17 January. Zurich, Switzerland, 91–109.
Sardy, S., Tseng, P., and Bruce, A., 2001. Robust wavelet denoising. IEEE Transactions on Signal
Processing, 49 (6), 1146–1152. doi:10.1109/78.923297.
22 X. YANG ET AL.
Sheikh, H.R., Sabir, M.F., and Bovik, A.C., 2006. A statistical evaluation of recent full reference image
quality assessment algorithms. IEEE Transactions on Image Processing, 15 (11), 3440–3451.
doi:10.1109/TIP.2006.881959.
Soleymani, A., et al., 2017. Characterizing change points and continuous transitions in movement
behaviours using wavelet decomposition. Methods in Ecology and Evolution, 8, 1113–1123.
doi:10.1111/2041-210X.12755
Sun, G., Webster, C., and Zhang, X., 2019. Connecting the city: a three-dimensional pedestrian network of
Hong Kong. Urban Analytics and City Science, 48, 60–75. doi:10.1177/2399808319847204
Tal, G. and Handy, S., 2012. Measuring nonmotorized accessibility and connectivity in a robust
pedestrian network. Transportation Research Record: Journal of the Transportation Research Board,
2299, 48–56. doi:10.3141/2299-06
Tang, J., et al., 2019a. An automatic method for detection and update of additive changes in road
network with GPS trajectory data. ISPRS International Journal of Geo-Information, 9 (8), 1–20.
Tang, L., et al., 2016. CLRIC: collecting lane-based road information via crowdsourcing. IEEE Transactions
on Intelligent Transportation Systems, 17 (9), 2552–2562. doi:10.1109/TITS.2016.2521482.
Tang, L., et al., 2019b. Detecting and evaluating urban clusters with spatiotemporal big data. Sensors,
19 (3), 1–15. doi:10.3390/s19030461.
Uduwaragoda, E.R.I.A.C., Perera, A.S., and Dias, S.A.D., 2013. Generating lane level road data from vehicle
trajectories using kernel density estimation. In: Proceedings of the 16th International IEEE conference on
Intelligent Transportation Systems. Hague, Netherlands, 384–391. doi:10.1109/ITSC.2013.6728262.
Wang, S., Zhou, Q., and Tian, Y., 2020. Understanding completeness and diversity patterns of
OSM-based land-use and land-cover dataset in China. ISPRS International Journal of Geo-
Information, 9 (9), 531. doi:10.3390/ijgi9090531.
Wang, Z., et al., 2004. Image quality assessment: from error visibility to structural similarity. IEEE
Transactions on Image Processing, 13 (4), 600–612. doi:10.1109/TIP.2003.819861.
Xie, X. and Ou, G., 2018. Pedestrian network information extraction based on VGI. Geomatica, 72 (3),
85–99. doi:10.1139/geomat-2018-0006.
Yang, X., et al., 2018a. Generating lane-based intersection maps from crowdsourcing big trace
data. Transportation Research Part C: Emerging Technologies, 89, 168–187. doi:10.1016/j.
trc.2018.02.007.
Yang, X., et al., 2018b. Automatic change detection in lane-level road networks using GPS
trajectories. International Journal of Geographical Information Science, 32 (3), 601–621.
doi:10.1080/13658816.2017.1402913.
Yang, X., et al., 2018c. A data cleaning method for big trace data using movement consistency.
Sensors, 18 (3), 824. doi:10.3390/s18030824.
Yang, X., et al., 2020. Pedestrian network generation based on crowdsourced tracking data.
International Journal of Geographical Information Science, 34 (5), 1051–1074. doi:10.1080/
13658816.2019.1702197.
Zhang, H. and Ye, C., 2017. RGB-D camera based walking pattern recognition by support vector
machines for a smart rollator. International Journal of Intelligent Robotics & Applications, 1 (1), 32.
doi:10.1007/s41315-016-0002-6.
Zhang, Y., et al., 2012. Associations between road network connectivity and pedestrian-bicyclist accidents.
In: Transportation Research Board Annual Meeting, 13–17 January. Washington, DC, 1–17.
Zhang, Y., et al., 2020. A hybrid method to incrementally extract road networks using
spatio-temporal trajectory data. ISPRS International Journal of Geo-Informaiton, 9, 186.
doi:10.3390/ijgi9040186.
Zhou, B., et al., 2015. ALIMC: activity landmark-based indoor mapping via crowdsourcing. IEEE
Transactions on Intelligent Transportation System, 16 (5), 1–11. doi:10.1109/TITS.2015.2423326.
Zimmermann, M., Tien, M., and Emma, F., 2017. Bike route choice modeling using GPS data without
choice sets of paths. Transportation Research Part C: Emerging Technologies, 75, 183–196.
doi:10.1016/j.trc.2016.12.009
INTERNATIONAL JOURNAL OF GEOGRAPHICAL INFORMATION SCIENCE 23