You are on page 1of 10

Construction Research Congress 2014 ©ASCE 2014 1033

Automated Monitoring of Operation-Level Construction Progress Using 4D


BIM and Daily Site Photologs

Kevin K. HAN1 and Mani GOLPARVAR-FARD2


1
PhD Student, Dept. of Civil and Environmental Eng., Univ. of Illinois at Urbana-
Champaign, Urbana, IL 61801; PH (510) 229-9648; email: kookin2@illinois.edu
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

2
NCSA Faculty Fellow and Assistant Professor of Civil Engineering, and of
Computer Science, Univ. of Illinois at Urbana-Champaign, Urbana, IL 61801; PH
(217) 300-5226; email: mgolpar@illinois.edu

ABSTRACT
Recent research efforts on improving construction progress monitoring have
mainly focused on model-based assessment methods. In these methods, the expected
performance is typically modeled with 4D BIM and the actual performance is sensed
through 3D image-based reconstruction method or laser scanning. Previous research
on 4-Dimensional Augmented Reality (D4AR) models– which fuse 4D BIM with
point clouds generated from daily site photologs– and also laser scan-vs.-BIM have
shown that it is possible to conduct occupancy-based assessments and as an indicator
of progress, detect whether or not BIM elements are present in the scene. However, to
detect deviations beyond typical Work Breakdown Structure (WBS) in 4D BIM,
these method also need to capture operation-level details (e.g., current stage of
concrete placement: formwork, rebars, concrete). To overcome current limitations,
this paper presents methods for sampling and recognizing construction material from
image-based point cloud data and using that information in a statistical form to infer
the state of progress. The proposed method is validated using the D4AR model
generated for a building construction site. The preliminary experimental results show
that it is feasible to sample and detect construction materials from the images that are
registered to a point cloud model and use frequency histograms of the detected
materials to infer the actual state of progress for BIM elements.

INTRODUCTION
Early detection of performance deviations in field construction activities is
critical as it provides an opportunity for project management to avoid them or
minimize their impacts. Despite its significance, traditional monitoring practice
includes manual site data collection (e.g., daily construction reports), extensive data
extraction from different construction plan documents, and non-systematic reporting.
To overcome these limitations, research efforts on improving construction
monitoring are mainly focused on model-based assessment methods. In these
methods, the expected performance is typically modeled with 4D Building
Information Models (BIMs) and the actual performance is sensed through 3D laser
scanning (Turkan et al. 2013, Kim et al. 2013, Turkan et al. 2012, Bosché and Haas
2008) or 3D image-based reconstruction methods (Golparvar-Fard et al. 2012, 2011,
2009, Wu et al. 2010). These 3D sensing techniques produce raw point cloud models
and thus their products do not contain additional semantics of object oriented data
(e.g., materials, element interconnectivity information).

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1034

Assessment of progress deviations using raw point cloud data and BIM has
primarily focused on techniques that can register BIM with the point cloud models in
a common 3D environment and then analyzing progress deviations based on
assessing the density of reconstructed points in 3D. To better deal with both static and
dynamic occlusions, Golparvar-Fard et al. (2012) also proposed a metric to measure
the expected visibility per element from the camera convex hull. The outcome of all
these methods –per element– is the assessment of 3D occupancy in a binary fashion,
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

whether the element is within the expected 3D volume as indicated in BIM or not.
Nonetheless, accurate monitoring still requires detecting construction progress
at the operation-level. For example, for construction of a concrete foundation wall,
the operational details may involve several stages of formwork, installing
reinforcement bars, placing concrete, finishing, waterproofing, insulation, and
backfilling of the excavated areas (See Figure 1).

Figure 1. Ex. of operational-details for constructing a concrete foundation wall

As a step toward accomplishing this goal, Turkan et al. (2013) expanded the
Scan-vs.-BIM comparison method of Bosché and Haas (2008) by revising the point
matching metrics to detect secondary and temporary objects (rebar and formwork for
concrete structures) in highly accurate and dense laser scanning point cloud models.
Their work also proposed a method for the analysis of points contained within open
spaces for recognition of concrete shoring. Reducing false positives and false
negatives in these detections even where highly dense point clouds are available– as
the authors have also indicated as open research problems – would require:
(1) Understanding and modeling the appearance of elements; i.e., texture and
color of the 3D elements– Occupancy based methods cannot easily differentiate
between several types of progress: e.g., concrete vs. waterproofing.
(2) Reasoning interconnectivity among elements– by formalizing
construction sequencing and using that to infer progress under partial and full
occlusions.
(3) Assessing confidence in each measurement– estimating expected
visibility per BIM element as a confidence indicator for progress monitoring.
This paper builds on our prior work on automated progress monitoring using
D4AR- 4 dimensional augmented reality- models (Golparvar-Fard et al. 2012) and
proposes two methods for recognizing appearance of elements and inferring a state of
progress. In our prior work (Golparvar-Fard et al. 2012), the assessment of

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1035

confidence using image-based point cloud models as a way of capturing as-built was
addressed. Here we focus on the problem of recognizing an up-to-date state of
progress assuming the confidence based on BIM and the camera convex hull is
already assessed. The proposed technique is validated using actual BIM, construction
schedule, and imagery data obtained during construction of a reinforced concrete
building. In the following, we briefly review the prior work on the D4AR modeling,
and relevant material recognition methods. We then present our method and discuss
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

experimental results.

BACKGROUND

The D4AR – 4D augmented reality – modeling


In prior work, (Golparvar-Fard et al. 2009), we proposed the D4AR modeling
method for progress tracking and as-built model visualization using unordered daily
construction images collections as well as BIM. Figure 2 illustrates the modeling and
monitoring process which has four key steps:

Figure 2. The D4AR modeling and the process of progress monitoring

1) 3D as-built point cloud modeling– Given a set of unordered and un-


calibrated site photographs –taken over a short period for which the site has not had
significant progress– a 3D point cloud model of the site is reconstructed using a
pipeline of Structure-from-Motion (SfM), Multi View Stereo, and voxel coloring
algorithms;
2) Generating 4D as-built point clouds– For photos taken over longer periods
of time (span of weeks or months), several point cloud models are generated and
superimposed using a derivative of the Iterative Closest Point (ICP) algorithm which
considers the scale factor. Here the common visual features detected across these
point cloud models are used for initialization of the scale+ICP algorithm;
3) Superimposing BIM and 3D/4D point clouds– Using user-selected points
from the point cloud and by finding their correspondences with the BIM, the point
cloud is scaled and transformed into the BIM coordinate system. Here, we used a
closed-form solution to solve the least square registration problem of absolute
orientation for 7 DOF (uniform scale, rotation, and translation).
4) Progress monitoring using the D4AR models– In our approach for
detecting progress deviations, the method first traverses and labels the as-built scene
for occupancy. The BIM is subsequently traversed and labelled for expected progress
visibility. A multiple one-vs.-all Support Vector Machine (SVM) classifier – built
upon a Bayesian model to deal with partial visibilities and occlusions– is proposed to
automatically detect physical components in units of voxels. The resulting D4AR

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1036

model enables the plan and as-built models to be jointly explored with an interactive,
image-based 3D viewer where deviations are color-coded over the BIM elements.
Figure 3-a and b show the as-built point cloud which is generated using 160
two mega-pixel images (Video demo can be found at: http://vimeo.com/16987971).
Subfigures c and d further illustrate an upgraded application where we jointly
visualize images with the BIM, which will be discussed later in this paper.
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

Figure 3. (a-b) as-built point cloud model; (c-d) new joint visualization

Material recognition in support of progress monitoring


Identifying construction material for the purpose of image retrieval and
identifying building elements from 2D images was first proposed by Brilakis et al.
(2005). More recently, Dimitrov and Golparvar-Fard (2014) also proposed an image-
based material recognition algorithm. In this method, material appearance was
modeled by a joint probability distribution of responses from a filter bank and
principal Hue-Saturation-Value (HSV) color values in a Bag-of-Words model. The
work utilizes Leung and Malik (2001) filter bank which contains 48 filters– 36
oriented filters at 6 orientations, 3 scales, and 2 filters, in addition to 8 derivative
filters and 4 low-pass Gaussian filters. This distribution is derived by concatenating
frequency histograms of filter response and HSV color clusters. These material
histograms are classified using multiple one-vs.-all SVM classifiers with χ2 kernel.
Classification performance was compared to the state-of-the-art algorithms both in
computer vision and Architecture/Engineering/Construction (AEC) communities. A
new database containing 20 typical construction materials with more than 150 images
per category was also assembled and used for validating the proposed method (see
Figure 4).
Overall, for material classification an average accuracy of 97.1% for 200×200
pixel are reported. To deal with small image patches (e.g., ~30×30 pixels) – which is
expected to happen for progress monitoring purposes where images might be
captured far from construction elements– a reflection/rotation process was also
proposed to synthetically generate additional pixels for analysis purposes, and
average accuracy of 90.8% for 30×30 pixel image patches was reported. In this paper,
we build on our prior work on material recognition and infer progress using small 2D
image patches that are sampled from site-registered images (in point cloud) that
observe the BIM elements.

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1037
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

Figure 4. Material recognition method (Dimitrov and Golparvar-Fard 2014)

APPEARANCE-BASED DETECTION METHOD


Our paper particularly proposes two methods for 1) sampling and recognizing
construction material from image-based point cloud data in the D4AR models; and 2)
using that information in a statistical form to infer the state of progress for the BIM
elements. Figure 5 shows the overall process. Here we assume a D4AR model is
assembled using BIM, construction schedule, and daily site photologs. In the
following, we discuss how we can utilize very small image patches and infer the state
of progress for visible BIM elements using the underlying D4AR model and a
reasonable material recognition model:

BIM
1
Back Projection of Material
BIM on site images Recognition
Model
Schedule D4AR
2 3
Site Sample Image Recognize Construction
Images Patches per element material per patch

4
Infer State of Progress
Update and color-code per BIM element

Figure 5. Overall workflow of our proposed appearance-based monitoring

Step 1: Back-projection of BIM on site images and generating depth map


In the D4AR models, the image-based point cloud is transformed into the site
coordinate system, and thus, we can back-project all expected BIM elements
( ) onto all images that observe them (
) using the transformed camera projection parameters that are derived
from the SfM step and the scene Euclidean transformation matrix.
In the SfM step, we use pinhole camera model and we estimate both intrinsic
and extrinsic parameters of all cameras: , , , , and , representing the focal
length, two radial distortion coefficients, 3×3 matrix for the camera rotation, and a 3-
vector for the translation, respectively. We outline the process of 3D-to-2D projection
for each image in steps 4-6 in Figure 6. First, the coordinates of the vertices ( )
for each face ( ) from the BIM element ( ) are transformed into the camera

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1038

coordinates by applying R and T– the outcome is a 3-vector. Then to deal with


homogeneous coordinates, this 3-vector ( is divided by its third row. The last
step of the back-projection is converting into pixel coordinates using a scaling factor,
r, to undo the radial distortion.

Input: Camera parameters <f, k1, k2, R, T> for & BIM elements ( )
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

Output: Depth map and back-projections of max for all onto each image
1. for each element
2. for each
3. for each vertex
4. transform into camera coordinate &
perspective division
5. undo radial distortion
6. transform onto image plane
7.
face with max 2D back-projected area for element
8. , :center location of camera and :center of in 3D
9. sort(B) for all ( ) by for depth map
10. return depth map and plot only for all ( ) as in sort(B)
Figure 6. Back-projection and depth map of BIM onto an image
Once the back-projections are completed, for each image we identify the
element face that has produced the maximum 2D area and use those pixels that are
associated with only one face of the element for appearance-based assessment
purposes. The intuition here is to guarantee that the sampling for material
classification is conducted over flat surface. This strategy provide us with larger areas
for extracting the required patches, and also minimizes the risk of taking samples
from location where edges and corners of element might be present as artifacts.
Next, the back-projected elements are sorted in order of their distances to each
camera location. This information is then used for generating a depth map per image
by simply plotting the sorted elements, starting from the farthest (blue) and ending
with the closest (red) to the camera. The pseudo code (Figure 6) outlines the entire
process of back-projection and depth map generation. Figure 7 shows the outcome of
this process for one of the registered site imagery.
Due to the presence of static and dynamic occlusions presented in a photo
(static: backfilling in from of an element not modeled in BIM; dynamic: presence of
equipment in front of the BIM element) we expect some of these samplings to be
irrelevant to the surfaces that are under inspected. Nonetheless because we take a
large number of samples from the scene, we expect the highest frequency of sampling
return more image patches from the relevant surface material. Our results as shown
later in the experimental result section further validates this hypothesis.

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1039
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

Figure 7. Back-projection and depth map of back-projected BIM elements

Step 2 and 3: Sampling image patches per element and material classification
Once the depth maps are created for every image, the progress of each
element ) is assessed by 1) selecting several image patches from the relevant
parts of all images that observe the element, 2) classifying material per patches, and
3) inferring progress using statistical distribution of the material classes and the
scores from SVM classifiers. Thus, from the back-projected of each element
, in each image that observes each , sample patches of pixels– which may
have some overlap and all their pixels are within the boundaries of the – are
randomly stored. Each image patch is placed into the material classification (steps 1-3
in Figure 8) adopted from Dimitrov and Golparvar-Fard (2014). Figure 9 shows
examples of these material patches that are extracted from back-projected face of a
BIM element.

Input: Depth map and back-projections of max for all onto all images
Output: The observed material for each element
1. for each
2. for all from all images ( s) that observe
3. randomly extract sample patches of pixels within each
4. for each
5. classify Material and return the class with highest score
6. ; maximum frequency of material
Figure 8. Extraction of sample patches and material recognition

Step 4: Infer state of progress per BIM element


Based on the outcome of material classification for all image patches, a
frequency histogram based on all existing materials in the classification model is
plotted and the class that returns the highest score is chosen to infer the up-to-date
state of progress. For instance, if formworks were presented, it is work-in-progress as
opposed to concrete that indicates finished work.

EXPERIMENTAL RESULTS AND DISCUSSION

GroundTruth and setup


In order to evaluate the performance of the proposed material recognition and
construction monitoring methods, experiments were conducted by taking 160 2-mega

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1040

pixel photos during construction of four-story concrete building project in


Champaign, IL. These images were used to generate a dense 3D point cloud model
and the outcome was superimposed with BIM – containing 141 visible elements– to
form the D4AR model. The material classification method was trained on our
Construction Materials Library (CML) which contains more than 3,000 images for 20
different construction materials collected from various site over a period of about six
months (http://raamac.cee.illinois.edu/materialclassification). The point cloud model
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

and BIM were registered using 15 controlled points with 5.6 cm accuracy. For each
image that observes an element, a total of 10 random 30×30 samples were collected.

Figure 9. Extraction of sample patches and material recognition

Material recognition method


We trained and tested the material recognition method on the CML dataset.
First, by reversing and rotating each image patch of 30×30 pixels, we synthetically
created 90×90 pixel image patches and tested the performance of the algorithm. Our
results indicated that the average accuracy in material recognition was 90.8%. Figure
10 shows the confusion matrix of our classification process. As observed in our
dataset we have several similar categories, and thus for progress inference we
grouped these material together (e.g., concrete-1 and concrete-2 are treated as the
same material). This further minimizes confusion in classification as it is currently
observed in this figure.

Figure 10. Confusion matrix for material classification using CML

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1041

Progress inference per element.


By classifying the material categories for all image patches of all 3D elements
in BIM, a frequency histogram of materials per BIM element were put together.
Figure 11 shows three sampled material frequency histograms. For these three cases,
we expected concrete surfaces. In this figure, class #5 represents concrete material
and, therefore, have the highest observed frequencies for all three graphs, achieving
accurate inferences on the right material associated with the current state of progress.
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

Figure 11. Example results of material frequency histograms

Figure 12 further illustrates the results on material inference for all visible
BIM elements that are made only of concrete. In this Figure, the frequency of
observing “concrete” is plotted in blue. The second highest frequency for the detected
materials is also plotted in “orange” for comparison purposes. Our preliminary results
show that even in presence of static and dynamic occlusions, inaccuracies in
sampling due to BIM-vs.-point cloud registration and presence or edges and corners
as artifacts, the method successfully returns concrete as the correct material class for
all observed elements. In our method, the visibility to elements are maximized by
sampling a large number of image patches from different perspectives. This strategy
minimize the chances of having occlusions or artifacts from particular viewpoints
may affect the outcome of material classification. Thus, it also reduces the risk of
inferring incorrect states for element progress.
1
0.9
0.8
Frequency

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131

Elements #
Material with highest freq Material with second highest freq.

Figure 12. Highest (concrete) and second highest (others) frequency observed

Discussion on Research Challenges and Future Work


While promising results are achieved in our preliminary experiments, several
research challenges remain open: 1) inferences on materials associated with slender
elements (e.g., columns) wherein the number of pixels in one of the image axes is
minimal, (e.g., limits ability to extract patches if an element has less than 30 pixels in

Construction Research Congress 2014


Construction Research Congress 2014 ©ASCE 2014 1042

horizontal or vertical direction); 2) formalization of sequencing for construction


activities so that inference of progress at operation-level and also for highly-occluded
elements could be conducted using embedded semantics in the BIM; 3) devising a
rationale for assigning percentages to physical work-in-progress according to an
observation. These are currently being explored as part of ongoing research and
results will be presented soon.
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 02/26/18. Copyright ASCE. For personal use only; all rights reserved.

CONCLUSION AND FUTURE WORK


To overcome current limitations associated with operation-level assessment of
construction progress, this paper presents two methods for sampling and recognizing
construction material from image-based point cloud data and inferring progress using
a statistical representation from the material classification. The preliminary results
show that it is feasible to sample and detect construction materials from images that
are registered to point clouds and use that to infer the state of progress for BIM
elements.

REFERENCES
Bosché F. and Haas C.T. (2008). “Automated retrieval of 3D CAD model objects in
construction range images.” Automation in Construction, 17 (4), 499-512.
Brilakis, I., Soibelman, L., and Shinagawa, Y. (2005). “Material-based construction
site image retrieval.” J. Comput. Civ. Eng., 19, 341-355.
Brilakis, I. and Soibelman, L. (2006). ”Multimodal Image Retrieval from
Construction Databases and Model-Based Systems.” J. Constr. Eng.
Manage., 132(7), 777–785.
Dimitrov, A., and Golparvar-Fard, M. (2014). “Vision-based material recognition for
automated monitoring of construction progress and generating building
information modeling from unordered site image collections.” Advanced
Engineering Informatics.
Golparvar-Fard, M., Peña-Mora, F., Savarese, S. (2009). “Application of D4AR – A
4-Dimensional augmented reality model for automating construction progress
monitoring data collection, processing and communication.” ITcon, 14, 129-153.
Golparvar-Fard, M., Peña-Mora, F., and Savarese, S. (2011). ”Integrated Sequential
As-Built and As-Planned Representation with D4AR Tools in Support of
Decision-Making Tasks in the AEC/FM Industry.” J. Constr. Eng.
Manage., 137(12), 1099–1116.
Golparvar‐Fard, M., Peña‐Mora, F., and Savarese, S. (2012). “Automated Progress
Monitoring Using Unordered Daily Construction Photographs and IFC‐Based
Building Information Models.” J. Comput. Civ. Eng., 10.1061
Kim C., Son, H., and Kim C. (2013). “Automated construction progress measurement
using a 4D building information model and 3D data”, Auto. Constr., 31, 75-82.
Leung, T., and Jitendra, M. (2001). “Representing and recognizing the visual
appearance of materials using 3D textons.” Int. J. Comp. Vision, 43(1), 29-44.
Turkan Y., Bosché F., Haas C.T., and Haas R. (2012). “Automated progress tracking
using 4D models and 3D sensing technologies.” Auto. Constr., 22 (1), 414-421.
Wu Y., Kim H., Kim C., and Han S.H. (2010). “Object recognition in construction
site images using 3d cad-based filtering.” J. Comput. Civ. Eng., 24 (1), 56-64.

Construction Research Congress 2014

You might also like