Personalized Walkability Assessment for Pedestrian Paths:
An As-built BIM Approach Using Ubiquitous Augmented
Reality (AR) Smartphone and Deep Transfer Learning
Xue, F. 1*, Chiaradia, A.J.F. 2, Webster, C.J. 3, Liu, D. 4, Xu, J. 5 and Lu, W.S. 6
Abstract: The walkability of pedestrian paths serves as an urban semantic infrastructure for the
global smart city development. However, traditional methods cannot fulfill the task of
personalized walkability assessment on (i) automation of assessment, (ii) the level of details, and
(iii) meeting the personalized demands of pedestrians (e.g., with a wheelchair). This paper
presents a novel as-built building information model (BIM) approach based on ubiquitous
augmented reality (AR) smartphone and deep transfer learning. First, an ‘as-is’ 3D point cloud is
scanned by AR smartphone. Then, an as-built BIM containing semantic objects such as paved
footway, guardrail, and obstacles can be created with pre-trained 3D point classification models.
Finally, calculation of walking characteristics of the semantic as-built BIM leads to personalized
walkability assessment. A pilot study involving five walking characteristics was conducted on a
real case near the University of Hong Kong for the personalized requirements of five types of
pedestrians: (1) people in wheelchair, (2) leisure traveler with a baby stroller, (3) business traveler
with a rolling luggage, (4) senior people, and (5) exerciser (joggers). The results of the pilot study
showed that the personalized walkability was correctly assessed using the proposed as-built BIM
approach. With advanced ubiquitous computing devices, the proposed approach is expected to
apply to everyday AR smartphones to enable crowd-sourcing and real-time assessment of the
personalized walkability for pedestrian paths. The semantic as-built BIM, as an intermediate
solution in this approach, can also open new avenues for many innovative smart city applications
to foster a smart, sustainable, and resilient future.
Keywords: Personalized walkability; As-built BIM (building information model); Augmented
reality (AR) smartphone; 3D point cloud; Deep transfer learning; Urban semantics; Ubiquitous
computing; Data-driven modeling; Smart city.

1* Xue, F.
Corresponding author, Department of Real Estate and Construction, The University of Hong Kong, Pok Fu Lam,
Hong Kong SAR
E-mail: xuef@hku.hk
ORCiD: http://orcid.org/0000-0003-2217-3693
Tel: (852) 2219 4174
2 Chiaradia, A.J.F.
Department of Urban Planning and Design, The University of Hong Kong, Pok Fu Lam, Hong Kong SAR
3 Webster, C.J.
Faculty of Architecture, The University of Hong Kong, Pok Fu Lam, Hong Kong SAR
4 Liu, D.
Department of Real Estate and Construction, The University of Hong Kong, Pok Fu Lam, Hong Kong SAR
5 Xu, J.
Department of Real Estate and Construction, The University of Hong Kong, Pok Fu Lam, Hong Kong SAR
6 Lu, W.S.
Department of Real Estate and Construction, The University of Hong Kong, Pok Fu Lam, Hong Kong SAR

1 Introduction
This paper focus on personalized walkability assessment (PWA) for pedestrian paths (or
sidewalks). Walkability is conceptualized as the extent to which the characteristics of the built
environment may or may not be conducive to residents in the area walking for leisure (e.g., with a
baby stroller), traveling (e.g., in a wheelchair), or exercise (e.g., jogging)[1]. Beyond walking,
walkability impacts many aspects of urban living[2] such as health[3], residential housing value, and
neighborhood crime[4]. In view of the fact that 65% world’s population will be urban residents by
2050[5], walkability is not only an essential concept for semantically-rich information retrieval but
also an important analytical concept in infrastructure for human-centered smart, sustainable, and
resilient city development [6].
Researchers have developed many indices and approaches for assessing walkability at the
macroscale (city)[7], neighborhood scale[4], and microscale[2,8-9]. However, it is costly and
time-consuming to apply the traditional methods (e.g., audit tools) to the citywide pedestrian paths
due to the vast amount, heterogeneous conditions, and frequent changes. The task becomes even
more challenging for personalized assessment, e.g., walkability for people in a power wheelchair
means something different to that for a jogging person. Hardware that enables PWA is readily
available in the market, e.g., ubiquitous smartphones are evolving to augmented reality (AR)
devices that can scan the as-is conditions with a centimeter accuracy. Computer vision software
libraries such as deep learning models have also been developed to understand the results of AR
scanning. So far, these advanced technologies have not been integrated into an automatic and
inexpensive approach to PWA, as far as we are aware.
This paper presents a novel as-built building information model (BIM) approach to PWA for
pedestrian paths using the up-to-date ubiquitous AR smartphone, deep transfer learning, and
as-built BIM creation and analytics. The remainder of this paper is organized into the following
sections: related work, a framework of the approach, a pilot case study, discussion, and
conclusion.

2 Related Work
Many conventional walkability assessment methods rely on observational audit tools. For example,
Sun et al. employed 67 characteristics to measure the walking environments of access routes to
urban metro stations in China in an urban planning viewpoint[2]; Griew et al. employed Google
street view for facilitating their audit tool[9]. Although audit conducted by surveyors can, with care
and attention to validation, be reliable and accurate, it is too laborious and costly to apply to PWA
for the complex network of dynamic pedestrian paths. Using geographic information systems (GIS)
technologies, such as GPS records[10] and imagery analysis of aerial photos[11] to collect real-world
walking data, is also investigated in the literature, to supplement audits. Yet, the methods based on
GIS suffer from inaccuracy (i.e., the unsatisfactory error level of GPS signal or path details in
aerial photos for the built environment) and incomplete data (e.g., the paths are mostly a centerline
in most GIS platforms)[12]. In summary, for PWA, especially in the contemporary settings for
smart city applications, traditional assessment methods fail in at least one of three aspects: (i)
automation of assessment, (ii) the level of details, and (iii) meeting the personalized demands of
pedestrians (e.g., different physical impairments)[13].
A building information model (BIM) involves physical as well as functional characteristics of
a facility and serves as an information hub to facilitate construction management along the whole
life-cycle[14]. An as-built BIM focuses more on the as-is conditions[15] such as actual building
geometries, current functions, and real topology to enable various smart city applications such as
construction automation, energy consumption, and facility management [15]. Approaches to as-built
BIM creation can be broadly categorized as data-driven (i.e., based on the input data’s features,
shapes, materials, and statistics)[15-16] and model-driven (i.e., based on the BIM components and
knowledge fitted into the input)[17-18]. Due to the variety of objects on uncontrolled real-world
pedestrian paths, data-driven methods such as those exploiting architectural regularity and random
sample consensus (RANSAC)[16] are more competitive in creating as-built BIMs for PWA[18]. It
should be noted that the generated as-built BIMs can also be exported to GIS platforms to enable
advanced spatial-temporal analyses as well as many existing GIS-based walkability assessment

methods, once the as-built BIMs are geo-referenced[19].
These challenges motivate the as-built BIM approach to PWA presented in this paper. The
as-built BIM is the information hub about various walking characteristics that facilitate the PWA
for different pedestrians. In contrast to conventional assessment methods, the proposed approach
is automatic and ubiquitous, rich in details and urban semantics, and compatible with multiple
types of pedestrians.

3 An As-built BIM Approach to Personalized Walkability Assessment
The approach presented in this paper, as shown in Figure 1, is a straightforward technological
pipeline that integrates the state-of-the-art data standards and processing methods in AR scanning
from remote sensing and robotics, deep transfer learning from artificial intelligent and computer
vision, and as-built BIM creation from civil engineering and construction management. In the
proposed approach, the as-built BIM is the semantic information hub about walkability. The input
to the approach, as shown in Figure 1, is the actual 3D geometry of the built environment, i.e.,
pedestrian paths. In the pipeline, there are three main steps of automatic data processing: (1) 3D
scanning, (2) as-built BIM creation, and (3) assessment. The intermediates in the approach include
the as-is 3D point cloud of the geometry of a path scene and an as-built BIM. The outputs of the
proposed approach include PWA and recommendations from simulations. The proposed approach
is thus classified as a data-driven as-built BIM approach.

Figure 1. The technological pipeline of the proposed as-built BIM approach

Given a target pedestrian path, the first step is 3D AR scanning of the actual geometry to
form an as-is 3D point cloud. Accuracy and density of the point cloud are crucial for the whole
approach. Thus, advanced AR mobile devices which incorporate depth sensing 3D cameras,
motion tracking, and area learning, such as Google Project Tango, are preferred over other AR
devices such as Google ARCore and Apple ARKit. Because of the limited computational capacity
for running deep learning models on current AR smartphones, the sensed as-is 3D point cloud can
be uploaded to a GPU (graphics processing unit) powered workstation for post- processing.
The second step is as-built BIM creation from the as-is 3D point cloud, by taking advantage
of deep transfer learning, shape extraction and object recognition. In computer science, transfer
learning aims at reducing the recollection of training data by transferring knowledge from related
domains (models), and can thus save a significant amount of labeling effort[20]. As a subset, deep
transfer learning focuses on reusing pre-trained deep learning models such as convolutional
neural networks (CNNs), usually accelerated by modern GPUs. For example, Semantic-8[21] is an
open dataset about street scenes, scanned in Germany, and annotated with 8 semantic labels (e.g.,
“man-made terrain,” “natural terrain,” and “buildings”). Deep transfer learning can reuse a
PointNet[22] (or other CNN) model pre-trained on Semantic-8 to segment any 3D point cloud of
street scene captured from anywhere. Based on the segmented labels (e.g., “man-made terrain”
and “buildings”), shape extraction methods and object recognition methods such as a priori rules
can recognize and rectify[16] the semantic components to create a data-driven as-built BIM. This
step can be automatic or semi-automatic, depending on the as-built BIM creation methods.
The final step of the proposed approach is PWA by calculation and simulations. Quantitative
walking characteristics (e.g., slope grade and the number of steps) of the as-built BIM can be
calculated and compared to the pre-defined requirements of a specific type of pedestrians (e.g.,
people in a hand-powered wheelchair or joggers). The walkability of the input pedestrian path can
thereafter be assessed as “OK,” “Limited,” or “Failed” though a floor function (i.e., the worst) on
all the characteristics individually for each type of pedestrians. For example, an “OK” in slope
grade and a “Failed” in the number of steps result in a “Failed” in PWA. In addition, obstacles and
walkability recommendations can also be inferred from a series of ‘what-if’ simulations on the
as-built BIM. For example, a major obstacle implies that its removal improves the walkability

from “Failed” to “OK,” while a minor obstacle means removal refines the walkability from
“Failed” to “Limited” (or from “Limited” to “OK”).

4 A Pilot Study
4.1 As-is 3D scanning using AR smartphone
To validate the proposed approach, we selected a scene of a pedestrian path on Bonham Road next
to the University of Hong Kong, as shown in Figure 2 (a). The scanned area was 56.2 m². We
collected a dense cloud of 569,344 colorful points (over 10,000 points per m²) using an in-house
developed Android app on a Google Tango AR smartphone (model: Lenovo PB2-690Y). The as-is
3D point cloud, though somehow noisy, was adequately registered, stitched, and uniformly
sampled in the Stanford polygon (.ply) format. It was observed in Figure 2 (b) that there existed
two obstacles (i.e., a light pole in the red oval and a meter pole in the blue oval) on the pavement,
two drainage pipes (in the yellow oval) on the wall, and a guardrail (in the flat oval) between the
pavement and the roadway. The time spent on the 3D scanning was about four minutes. At the end
of scanning, the app uploaded the as-is point cloud, as shown in Figure 2 (b), to a GPU powered
BIM workstation in a few seconds via a local 4G mobile network.

(a) A scene of Bonham Road, Hong Kong

(b) As-is cloud of 569,344 points through AR scanning

Figure 2. A pilot case of a pedestrian path near the University of Hong Kong

4.2 As-built BIM creation using deep transfer learning
An as-built BIM was then created on the BIM workstation. The workstation has two Intel XEON
E5-2690 v4 CPUs (2.6GHz, 28 cores), 64 GB memory, an Nvidia Quadro P5000 GPU, and
Windows 10 Enterprise 64-bit operating system. Before the as-built BIM creation, a PointNet[22]
model was already trained on the Semantic-8 dataset[21] on Google TensorFlow (version 1.4, with
GPU support) in Python (version 3.6) to enable the deep transfer learning. First, the uploaded
point cloud was converted from the Stanford polygon (.ply) format to a Numpy (version 1.12) data
file. The pre-trained PointNet model spent 4 seconds on the segmentation of the scanned 569,344
points. Figure 3 (a) shows the result of segmenting the 3D points in Hong Kong with training data
from Germany, which was surprisingly satisfactory. For instance, it can be observed in Figure 3 (a)
that the points labeled as “man-made terrain” were very close to the actual geometry of the
pavement. Next, the excessive 3D points above the height of 1.5 m were removed, as shown in
Figure 3 (b), for focusing on the major space for pedestrians.

(a) 3D point classification (e.g., the points labeled as (b) As-built BIM consisting of semantic objects (walls
“manmade terrain” were detached as the pavement)
omitted in this view)
Figure 3. The results of 3D point classification and the as-built 3D model of the pilot case

Planes and cylinders were detected from the segmented and filtered point cloud using an
up-to-date as-built BIM creation algorithm[16] in less than 1 second. The detected geometric
primitives were further screened by the proximity of their normal direction to the z-axis for
focusing on the vertical and horizontal objects. Furthermore, the boundaries of planar objects were
refined by the 2D concave hull algorithm provided by the point cloud library (PCL, version 1.8.1).
The resulting objects were associated with semantic labels to form semantic BIM components
through a priori rules. Examples of rules are “vertical planes along the pavement are guardrails or
walls,” and “vertical cylinders connected to the pavement are possible obstacles.” Finally, an
as-built BIM was formed as shown in Figure 3 (b). The as-built BIM included the reconstructed
guardrail (i.e., the transparent flat polygon) and the pipes and poles (i.e., the four narrow
cylinders). It was interesting that an unexpected object, which was actually a concrete trace of a
former construction work left on the wall, was also recognized as shown as the wide cylinder in
Figure 3 (b). Although the as-built BIM in this pilot case was preliminary (i.e., not volumetric and
in popular BIM formats), the semantic information (i.e., the geometry and labels) was capable of
reasoning and calculating many detailed walking characteristics and problems.

4.3 Personalized walkability assessment based on the as-built BIM
To demonstrate the PWA of the proposed approach, this paper employs five walking
characteristics: (1) the number of steps, (2) slop grade of the footway, (3) tile grade of the footway,
(4) footway width, and (5) pavement surface clearance (i.e., non-existence of gaps and pebbles).
The as-built BIM was geo-referenced first, as shown in Figure 4 (a), for visualization and
compatibility with GIS platforms. The number of steps was identified as zero by analyzing the
variation in altitude of the paved footway. The pave footway in the as-built BIM had two sections,
as shown in Figure 4 (b). the left section had a slope grade of 1:50 (or 1.1°) while the right one
had a grade of 1:58.8 (or 1.0°). The title grades of the two sections of the footway, as shown in
Figure 4 (c), were 1:47.6 (or 1.2°) and 1:66.7 (or 0.9°), respectively. The width of the paved
footway varied in different sections, as shown in Figure 3 (d). The maximum width was 199 cm in
the beginning and minimum width (i.e., the bottleneck) was only 45 cm, where the average width
was around 105 cm for the remaining parts of the paved footway. The green cylinders in Figures 3
(b) and 4 (d) stand for the possible obstacles. The surface clearance of the paved footway was
good, because no considerable gap was found, e.g., the gap shown in Figure 4 (c) was about 2 cm,
neglectable as an error of planarization of a curved surface footway.

(a) Geo-referenced first-person view

(c) The tilt grades of the paved footway

(b) The slope grades (i.e., tan θ) of the paved footway

(d) The actual widths of the pavement with obstacles

Figure 4. The results of point cloud classification and as-built 3D model of the pedestrian path

Based on the calculated values of the five characteristics, the walkability can be accurately
assessed for different pedestrians regarding their individual requirements. In this pilot study, we
defined five types of pedestrians: (1) people in a wheelchair, (2) leisure travelers with a baby
stroller, (3) business travelers with rolling luggage, (4) senior people, and (5) exerciser (joggers).
Table 1 shows the walkability assessment matrix between the five walking characteristics and the
five types of pedestrians. The bottom row indicates the results of PWA as the worst (floor function)
evaluations in each column. The results in Table 1 indicated that the pilot scene was unwalkable
for any people in wheelchair, and only walkable for travelers with a limited walkability for baby
strollers and rolling luggage.
Table 1. Assessing the walkability of the pilot case for five types of pedestrians
Walking
characteristic

Calculated value

Type of pedestrians
Stroller
Wheelchair ♿
Luggage 🛄🛄
Senior 👴👴 Exercise 🏃🏃
No. of steps
0
OK
OK
OK
OK
OK
Slope grade*
1:50.0~58.8
OK
OK
OK
OK
OK
Tilt grade†
1:47.6~66.7
OK
OK
OK
OK
OK
Footway width‡ 45~199 cm
Limited
Limited
OK
OK
Failed
Clearance
Good
OK
OK
OK
OK
OK
Overall walkability (the worst)
Limited
Limited
OK
OK
Failed
*: Reference maximum slope grade: 1:8~12 (wheelchairs);
†: Reference maximum tilt grade of pavement: 1:15 (wheelchairs);
‡: Reference minimum width: 70~90 cm (wheelchairs), 40~70 cm (strollers), and 30~60 cm (baggage).

In addition, obstacles to the walkability can be identified by a series of simulations. In each
simulation, a ‘what-if’ analysis of PWA (i.e., repeating Table 1) was carried out by removing one
possible obstacle. Table 2 lists the identified obstacles in three categories: major obstacles, minor
obstacles, and inoffensive obstacles. It can be found that the light pole in the middle of the paved
footway was identified as the only major obstacle. As a result, alternative recommendations can be
made, for example, moving the light pole to the guardrail, which can be subject to subsequent
cost-benefit analysis as well as public consultation, leading to a schedule of improvement works to
upgrade a PWA.
Table 2. A list of identified obstacles to walkability
Major obstacles
Light pole

Minor obstacles
(None)

Inoffensive obstacles
Meter pole, drainage pipe #1, #2, and concrete trace on the wall

4.4 Discussion
This pilot study preliminarily confirms the technological feasibility of the proposed approach.
Nevertheless, some minor details such as volumetric BIM components and commercial BIM
platforms (e.g., Autodesk Revit) were not tested in the study. In terms of the computational time
cost, the 3D scanning accounted for the significant portion, while other automatic processes like
network transmission, execution of deep transfer learning, object recognition, and as-built BIM
creation only cost a few seconds on the workstation. The flexible settings of the walking
requirements in the proposed as-built BIM approach lead to a personalized walkability assessment.
The high efficiency (fast processing speed) indicates that it is now possible to crowdsource the
PWA to ubiquitous AR smartphones. When ubiquitous AR smartphones become more and more
potent in computational power (e.g., with powerful GPUs) in future, the proposed approach can
also be fully embedded in everyday AR smartphones for real-time PWA and obstacle detection.
In comparison to the conventional walkability assessment methods, the proposed approach
has several advantages. First, the automation level of walkability assessment is elevated, so that
the traditional reliance on human surveyors by audit tools and inaccurate data by GIS-based
methods is now relieved. The proposed approach leverages state-of-the-art technologies in remote
sensing, deep transfer learning, and as-built BIM creation and achieves an accurate, consistent and
inexpensive assessment. Consistency is very important for the use of a walkability rating system.
Furthermore, the details of pedestrian paths are modeled and assessed, thanks to the centimeter
accurate, dense as-is 3D point cloud scanned by the AR smartphone. Last but not least, the
proposed approach measures personalized walkability for each individual pedestrian like a person

in a power wheelchair or a traveler with a large-model (e.g., 70 cm) baby stroller. To sum up, the
proposed approach yielded a novel, automatic, semantically rich PWA, which is expected to be
very competitive with manual surveying in smart city scenarios.
Disadvantages, nevertheless, can also be identified. In the first step of 3D scanning, noise is
problematic because the capacity of the embedded AR sensors and the related software libraries
are not perfect. Another limitation is the requirement of an additional BIM workstation and
network transmission apart from the AR smartphone, due to the reliance on powerful GPUs. The
GPUs and memory in state-of-the-art smartphone models are not powerful enough to execute
large-scale CNN models. Additionally, the as-built BIM, though successful in the pilot case, was
neither volumetric nor in popular commercial BIM formats. Errors, though not significant, can be
found in the results of PoinNet and as-built BIM creation as shown in Figures 3 (a) and (b).
However, with the promised technological advancement in AR smartphone and continuous
research, all these drawbacks are likely to be resolved in the near future.

5 Conclusion
This study presented an as-built BIM approach to PWA for pedestrian paths. The as-built BIM of a
target pedestrian path represents the actual geometry, semantic components (e.g., potential
obstacles), as well as the real topology, and serves as an information hub for assessing walkability.
In our approach, the everyday AR smartphone becomes a potent walkability assessment tool, and
pre-trained CNN models serve as a part of the automatic as-built BIM generator. The PWA created,
is automatic, highly detailed, and capable of assessing walkability for various types of pedestrians.
A pilot case of a real-world pedestrian path scene gave initial validation of the proposed approach.
This study is expected to enrich research of PWA, as well as as-built BIM study, for
pedestrian paths with a novel ubiquitous computing enabled framework for near-instant automated
modelling of a walking environment. Although the findings are preliminary, the technological
feasibility of crowdsourcing walkability assessment has been demonstrated by our prototyping
pilot study. Further research will aim to reduce noise in the 3D scanning stage, embed the
approach to AR smartphones without assistance from a BIM workstation, integrate with popular
BIM/GIS software and standards, and refine the deep learning models.

Acknowledgements
This study was supported by a grant from the Hong Kong Research Grant Council (GRF
HKU17201717) and two grants from University of Hong Kong (201702159013, 201711159016).

References
[1] Leslie, E., Coffee, N., Frank, L., Owen, N., Bauman, A., and Hugo, G. (2007). Walkability of
local communities: Using geographic information systems to objectively assess relevant
environmental attributes. Health & place, 13(1): 111-122.
[2] Sun, G., Webster, C., and Chiaradia, A. (2017). Objective assessment of station approach
routes: Development and reliability of an audit for walking environments around metro
stations in China. Journal of Transport & Health, 4: 191-207.
[3] Althoff, T., Hicks, J. L., King, A. C., Delp, S. L., and Leskovec, J. (2017). Large-scale physical
activity data reveal worldwide activity inequality. Nature, 547(7663), 336.
[4] Gilderbloom, J. I., Riggs, W. W., and Meares, W. L. (2015). Does walkability matter? An
examination of walkability’s impact on housing values, foreclosures and crime. Cities, 42:
13-24.
[5] World Health Organization. (2015). Global Report on Urban Health—Executive Summary.
http://www.who.int/kobe_centre/measuring/urban-global-report/en/ <accessed 29/04/2018>
[6] Birkmann, J., Welle, T., Solecki, W., Lwasa, S., and Garschagen, M. (2016). Boost resilience
of small and mid-sized cities. Nature, 537(7622): 605-608.
[7] Frank, L. D., Sallis, J. F., Saelens, B. E., Leary, L., Cain, K., Conway, T. L., and Hess, P. M.

(2010). The development of a walkability index: application to the Neighborhood Quality of
Life Study. British journal of sports medicine, 44(13): 924-933.
[8] Clifton, K. J., Smith, A. D. L., and Rodriguez, D. (2007). The development and testing of an
audit for the pedestrian environment. Landscape and Urban Planning, 80(1-2): 95-110.
[9] Griew, P., Hillsdon, M., Foster, C., Coombes, E., Jones, A., and Wilkinson, P. (2013).
Developing and testing a street audit tool using Google Street View to measure
environmental supportiveness for physical activity. International Journal of Behavioral
Nutrition and Physical Activity, 10(1): 103.
[10] Carlson, J. A., Saelens, B. E., Kerr, J., Schipperijn, J., Conway, T. L., Frank, L. D., Chapman,
J. E., Glanz, K., Cain, K. L., and Sallis, J. F. (2015). Association between neighborhood
walkability and GPS-measured walking, bicycling and vehicle time in adolescents. Health &
Place, 32: 1-7.
[11] Kasemsuppakorn, P., and Karimi, H. A. (2013). Pedestrian network extraction from fused
aerial imagery (orthoimages) and laser imagery (lidar). Photogrammetric Engineering &
Remote Sensing, 79(4): 369-379.
[12] Brownson, R. C., Hoehner, C. M., Day, K., Forsyth, A., and Sallis, J. F. (2009). Measuring
the built environment for physical activity: state of the science. American journal of
preventive medicine, 36(4): S99-S123.
[13] Aghaabbasi, M., Moeinaddini, M., Shah, M. Z., Shekari, Z. A., and Kermani, M. A. (2018).
Evaluating the capability of walkability audit tools for assessing sidewalks. Sustainable
Cities and Society, 37: 475-484.
[14] National Institute of Building Sciences. (2015). National Building Information Modeling
Standard. Version 3, Retrieved from https://www.nationalbimstandard.org/ <accessed
29/04/2018>
[15] Tang, P., Huber, D., Akinci, B., Lipman, R. and Lytle, A. (2010), Automatic reconstruction of
as-built building information models from laser-scanned point clouds: A review of related
techniques. Automation in Construction, 19(7): 829-843.
[16] Chen, K., Lu, W., Xue, F., Tang, P., and Li, L. H. (2018). Automatic building information
model reconstruction in high-density urban areas: Augmenting multi-source data with
architectural knowledge. Automation in Construction, 93: 22-34.
[17] Xue, F., Chen, K., Liu, D., Niu, Y., and Lu, W. S. (2018). An optimization-based semantic
building model generation method with a pilot case of a demolished construction. In
Proceedings of the 21st International Symposium on Advancement of Construction
Management and Real Estate: 231-241.
[18] Xue, F., Lu, W., and Chen, K. (2018). Automatic generation of semantically rich as-built
building information models using 2D images: A derivative-free optimization approach.
Computer-Aided Civil and Infrastructure Engineering, in press, doi:10.1111/mice.12378.
[19] Lu, W., Peng, Y., Xue, F., Chen, K., Niu, Y., and Chen, X. (2018). The fusion of GIS and
Building Information Modeling for big data analytics in managing development sites, in
Comprehensive Geographic Information Systems: 345-359.
[20] Pan, S. J., and Yang, Q. (2010). A survey on transfer learning. IEEE Transactions on
knowledge and data engineering, 22(10): 1345-1359.
[21] Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2017). Pointnet: Deep learning on point sets for
3D classification and segmentation. Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition: 652-660. doi:10.1109/CVPR.2017.16
[22] Hackel, T., Savinov, N., Ladicky, L., Wegner, J. D., Schindler, K., and Pollefeys, M. (2017).
Semantic3D.Net: A new large-scale point cloud classification benchmark. ISPRS Annals of
the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2017(IV-1-W1):
91-98.