G Model

AAP-3927; No. of Pages 11 ARTICLE IN PRESS

Accident Analysis and Prevention xxx (2015) xxxxxx

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

A Big-Data-based platform of workers behavior: Observations from

the eld
S.Y. Guo , L.Y. Ding, H.B. Luo, X.Y. Jiang
School of Civil Engineering and Mechanics and Institute of Construction Management, Huazhong University of Science and Technology, Wuhan, China

a r t i c l e i n f o a b s t r a c t

Article history: Behavior-Based Safety (BBS) has been used in construction to observe, analyze and modify workers
Received 28 June 2015 behavior. However, studies have identied that BBS has several limitations, which have hindered its
Received in revised form effective implementation. To mitigate the negative impact of BBS, this paper uses a case study approach
25 September 2015
to develop a Big-Data-based platform to classify, collect and store data about workers unsafe behav-
Accepted 29 September 2015
ior that is derived from a metro construction project. In developing the platform, three processes were
Available online xxx
undertaken: (1) a behavioral risk knowledge base was established; (2) images reecting workers unsafe
behavior were collected from intelligent video surveillance and mobile application; and (3) images with
Keywords:
Big Data
semantic information were stored via a Hadoop Distributed File System (HDFS). The platform was imple-
Behavior-Based Safety mented during the construction of the metro-system and it is demonstrated that it can effectively analyze
Behavior observation semantic information contained in images, automatically extract workers unsafe behavior and quickly
Intelligent video surveillance retrieve on HDFS as well. The research presented in this paper can enable construction organizations
Mobile application with the ability to visualize unsafe acts in real-time and further identify patterns of behavior that can
HDFS jeopardize safety outcomes.
2015 Elsevier Ltd. All rights reserved.

1. Introduction it has been dependent upon peoples observations; it can be a

The frequent occurrence of accidents cannot be ignored in con-
struction; most occur because of workers unsafe behavior (e.g.
Heinrich et al., 1950; Lingard and Rowlinson, 2005). To the reduce
incidents and improve safety management, it is therefore necessary
to focus attention on the unsafe behavior of workers (Choudhry,
2012). Research undertaken by Lingard and Rowlinson (2005), for
example, has suggested that unsafe behavior can be controlled
by implementing a series of safety measures, and by the workers
themselves.
Behavior-Based Safety (BBS) is an approach that can be used to
modify the behavior of people so that they undertake work more
safely (Geller, 2001). As a result, BBS has been widely adopted
by several industries (Chen and Tian, 2012), such as oil and gas
(e.g. Ismail et al., 2012), manufacturing (e.g. Yeow and Goomas,
2014; Nielsen et al., 2015), nuclear (e.g. Cox et al., 2004), and
construction (e.g. Lingard and Rowlinson, 1998; Li et al., 2015).
Yet while it has been effective in some instances (e.g. Krause et al.,
1999; Yeow and Goomas, 2014), it still has widely criticized as
Corresponding author.
E-mail address: gsy.yu@foxmail.com (S.Y. Guo).
time-consuming and an expensive process (Cameron and Duff,
2007). Moreover, insufcient samples, particular the identica-
tion of common unsafe behaviors (Laitinen et al., 1999) and the
passive participation of workers have contributed to its demise
(DePasquale and Geller, 2000; Han et al., 2011).
With this in mind, this paper proposes and implements a Big-
Data-based platform to identify patterns of unsafe behavior so that
organizations can implement strategies and techniques to improve
the safety of their construction sites. A behavioral risk knowledge
base is built to classify and code the observed behavior. Then, intel-
ligent video surveillance and a mobile application are used to collect
image data on site. Thus, large and complicated images from behav-
ior observation can be transferred into specic information, which
can be recognized by machines, so as to guarantee the data can be
nally stored on Hadoop Distributed File System (HDFS).
2. Behavior Based Safety
BBS or Applied Behavior Analysis (ABA) (e.g. Williams and
Geller, 2000), is derived from Skinners behavioral science the-
ory (Skinner, 1953) has been widely applied in construction to
improve safety management on-site and better educated subcon-
tractors (e.g. Lingard and Rowlinson, 1998; Chen and Tian, 2012;

http://dx.doi.org/10.1016/j.aap.2015.09.024
0001-4575/ 2015 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
2 S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx

Table 1
Camera function and its application in construction.

Function Functional explanation Application in construction

Intrusion detection
Stay identication
Reverse detection
Trafc analysis
Congestion detection
Line-crossing detection
Departure identication
Appearance identication
Disappearance
identication
Objects-leaving
identication
Objects-taking away
identication
Scene-changing
identication
Analyzing, positioning and tracking objects entering
prohibited area
Tracking over-stayed objects in protected area
Monitoring objects moving against the rules
Numbering the amount of people entering into a monitoring
area automatically
Identifying the congestion situation in monitoring area
Detecting objects crossing the warning line
Identifying people or objects leaving a given area
Identifying people or objects entering into a given area in the
picture (emphasize its sudden appearance but external
appearance)
Identifying people or objects disappearing from a given area
suddenly
Identifying remnants in given area
An object in the picture is taken away
The angle of camera is turned, leading to change of scene
Workers enter into dangerous lifting area
Workers stay too long at the edge of the foundation pit
Crane-driving and excavator-driving violators
None
Cross operation in the same area when mounting and
dismounting scaffolding and formwork
Workers cross dangerous area such as the edge of foundation
pit and the sides of the gangway, etc.
No conductors guardianship in operating radius of working
excavator
Workers go up and down the foundation pit by establishing
channel in private
Workers dont replace the reghting apparatus after using it
Workers place material on steel support
Workers remove their helmets on worksite
None

Note: As auxiliary functions, trafc analysis and identication of scene changes cannot indentify some kind of unsafe behaviors specically.

Han and Lee, 2013; Choudhry, 2014; Li et al., 2015). Noteworthy,
a detailed review of BBS in construction and its limitations can be
found in studies such as Zhang and Fang (2013) and Li et al. (2015).
Essentially, however, BBS focuses on what workers do, analyzes
why they do it, and then applies a research-supported interven-
tion technique to improve behavioral processes (Geller, 2005). BBS
intervention adds goal setting and/or feedback to alter response
probability (e.g. DePasquale and Geller, 2000; Cameron and Duff,
2007). As workers move toward a goal of safety compliance level,
the positive feedback they receive can contribute to improving
safety performance. The BBS process contains the following steps
(Geller, 2001): (1) unsafe behavior is listed, which requires close
attention on site; (2) workers unsafe behavior is observed and its
frequency is recorded; (3) feedback is given and their unsafe behav-
ior is modied. Despite studies highlighting the limitations of BBS,
it has also has been proven demonstrated as being an effective
approach for improving safety management in construction (e.g.
Geller, 2001; McSween, 2003; Turnbeaugh, 2010). For example,
Krause et al. (1999) study that compared workers casualty gures
from 73 companies over a ve year period, found that workgroups
that had been subjected to an intervention through BBS recorded
fewer accidents. Yeow and Goomas (2014) proposed an outcome-
and-Behavior-Based Safety incentive program (OBBSIP) to reduce
accidents in a uid manufacturing plant, the results demonstrated
the number of accident were signicantly reduced by 75% using
BBS.
A pertinent issue observed by Lingard and Rowlinson (1998),
however, is that BBS is over reliant on workers observations and
validation of unsafe behaviors. However, in many instances they are
unwilling to report their co-workers (Saba and Mohamed, 2008).
Thus, a reliable and objective technique that can readily determine
unsafe behaviors is required to support BBS.
3. Big Data and behavior observation
Big Data is capable of transforming the way in which organiza-
tions can visualize, function, and perform its routines and practices
(Mayer-Schnberger and Cukier, 2013). According to International
Data Corporation (IDC), Big-Data describes a new generation of
technologies and architectures that are designed to economically
extract value from very large volumes and a wide variety of data,
by enabling high velocity capture, discovery, and/or analysis (Gantz
and Reinsel, 2011).
In construction, a large number of workers behavioral data is
generated on site every day, which is rarely, or not all used to exam-
ine unsafe acts. According to Heinrichs (1950) for every 300 unsafe
acts there are 29 minor injuries and 1 major injury: since this obser-
vation there has been limited empirical evidence that has been able
to observe and quantify repeated patterns of unsface behavior. If
unsafe acts through the observation process can be controlled, then
the potential for accidents can be signicantly reduced; behavior
observation in this instance therefore becomes a valuable tool. With
this in mind, Big Data-based technologies can be used to observe,
to collect and store image data reecting workers unsafe behavior.
The development of image recognition provides support for
collecting unstructured data. Camera-based behavior analysis tech-
nology, for example, a branch of image recognition research, can
assist with the discovery of abnormal behavior and extract their
feature images from video sequences. Several studies have exam-
ined how video cameras can be used in construction (e.g. Chi and
Caldas, 2011). For example, the use of video monitoring system
was been found to contribute to the improvement of productiv-
ity and safety on construction site (e.g. Aguilar and Hewage, 2013;
Teizer et al., 2013). A functional of explanation and the application
of cameras in the construction can be seen in Table 1.
Camera technology has been utilized in transportation (e.g.
Hu et al., 2014), commerce (e.g. Lablack and Djeraba, 2008) and
security (e.g. Ko, 2008). However the application of direct image
recognition has its limitations, such as sematic gap between low
level image features and high level semantics tends to exist (Zhang
et al., 2012). To address limitation, the derived semantic informa-
tion extracted by people can be labelled so as to quickly understand
the expression of images. This is widely used in image retrieval (e.g.
Harmandas et al., 1997; Hu et al., 2003), as it can avoid analyzing
visual elements of images. Using a Java version of Google File Sys-
tem (GFS) (Ghemawat et al., 2003), Hadoop Distributed File System
(HDFS) provides a distributed le system satisfying the demand of
huge amounts of data storage. It can potentially provide high reso-
lution images of unsafe behaviors on-site, which can be produced
efciently.

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx 3

Data Classification and Coding Data Collection Data Storage

Expert
Expert Standard for Monitoring
Monitoring Mobile
Mobile Behavioral
Behavioral Risk
Risk
Accident
Accident Case
Case Experience
Experience Differentiating Work Video
Video Application
Application Visualization
Visualization

Behavior Analysis JHA&VSM

Safety
Safefty Standard
Standard
List
List ofof
Critical Video-capture Hadoop
Hadoop Distributed
Distributed File
File System
System
Critical Metro Construction Video-capture ofof Photographs
Photographs ofof
Unsafe Unsafe (HDFS)
(HDFS)
Unsafef WBS Unsafef Behavior
Behaviorr Unsafe
Unsafef Behavior
Behaviorr
Behavior
Behavior
Operating
Operating
Instruction
Instruction

Data Flow
Data
Data
Work Flow Source
Source

Data
Data Data Behavioral
Behavioral Risk
Risk Classification Structured Unstructured
Unstructured Data
Data Classififcation Behavioral
Behavioral Risk
Risk Recognition
Recognition Structured Data
Data Data
Storage
Storage and
and Coding
Coding
Data
Data
Types Process
Process
Types

Fig. 1. Big-Data-based platform for workers behavior observation.

3.1. A Big-Data-based platform
To overcome the limitations of traditional behavior observation,
a Big-Data-based platform is developed using an array of advanced
technologies. The Big-Data platform is presented in Fig. 1 and com-
prises of three parts.
Data classication: Critical unsafe behavior is listed from safety
standards, operating instructions, accident cases and expert expe-
rience. Combined with a Work Breakdown Structure (WBS)
classication standards, workers unsafe behavior during construc-
tion are classied and encoded.
Data collection: Data derived from images is collected through
intelligent monitoring videos and photographs taken by site man-
agers (or workers). Video capture of unsafe behavior is recognized
and recorded by camera-based behavior analysis technology. This
technology can obtain data automatically by recognizing abnor-
mal behaviors and extracting images from surveillance videos; the
extraction of which marries with the way in Big Data is collected.
Photographs of unsafe behavior are then matched with unsafe acts
in the knowledge database by means of Job Hazard Analysis (JHA)
and Vector Space Model (VSM). JHA is a process of identifying
potential hazards for each step of an activity and proposing safety
rules to prevent potential incidents related to these hazards (Wang
and Boukamp, 2011), so as to determine factors that can inuence
unsafe behavior. In addition, VSM can transfer the analysis of texts
into the calculation of vectors and express the semantic similarity
with spatial similarity (Salton et al., 1975). So keywords reecting
these factors, which are included in observers description, can be
recognized by the mobile application. At this point, it is envisaged
that behavioral risk recognition is realized.
Data storage: Collected multi-source isomerous data (structured
data and unstructured data) is stored by HDFS; it is designed to
store very large data sets reliably, and to stream those data sets
at high bandwidth to user applications (Shvachko et al., 2010).
Recorded images of unsafe behavior are marked by complete
semantic information, so behavioral risk visualization can be real-
ized.
4. Research approach
To test and determine the effectiveness of the developed Big-
Data-based platform, a case study approach (Yin, 2013) was
adopted. The selected case for this study is a underground metro
constuction project that is being undertaken in Wuhan (Hubei
Province) China. Wuhan is the fth largest city in China with
a population of approximately 10.4 million in 2015. As a result
of increasing trafc congestion and population and the need to
improve transportation infrastructure, the city has developed a
strategy to increase the length of its underground metro system.
For this project 15 stations and 8 tunnels from two lines (Line 3
and 6) were selected as data was made readily available from the
company responsible with constructing these metro lines.
4.1. Data classication and coding
The primary task of BBS is to conrm the checklist of unsafe
behavior, which is the basis for its observation. Choudhry (2014)
dened ve different types of unsafe behavior: (1) personal pro-
tective equipment (PPE), (2) housekeeping, (3) access to heights,
(4) plant and equipment, and (5) scaffolding. Similarly, Han and
Lee (2013) identied unsafe behavior using achecklist which
was derived from OSHA accident statistics. In the case of the
metro construction project, workers unsafe behavior were clas-
sied into different types according to safety standards, operating
instructions, accident cases and expert experience. The projects
established WBS, was encoded to form the behavioral risk knowl-
edge base.
Several safety standards and operating instructions in China
were used as references to list partial unsafe behavior, such as
Standard for Construction Safety Assessment of Metro Engineering

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
4 S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx

Table 2
Example of the lifting classication for unsafe behavior.

Name Code Possible injury Affecting factors

Whether the weight of suspended objects 001 Crane accident Action: lifting
exceed machines rated load is not Object-worked-on: heavy object
veried Location: none
Resource: crane
Nearby object: none
Nearby action: none

Transporting workers by hanging object 002 High falling Action: lifting

Object-worked-on: worker
Location: none
Resource: crane
Nearby object: none
Nearby action: none

Placing other oating objects on the 003 Object striking Action: lifting
hanging object Object-worked-on: oating object
Location: none
Resource: crane
Nearby object: none
Nearby action: none

Lifting combustible and explosive goods 004 Explosion Action: lifting

such as oxygen bottles without safety Object-worked-on: explosives
measures Location: none
Resource: crane
Nearby object: none
Nearby action: none

Lifting objects underground without 005 Crane accident Action: lifting

verifying the situation Object-worked-on: objects underground
Location: none
Resource: crane
Nearby object: none
Nearby action: none

Not taking warning during lifting, 006 Object strikes Action: lifting
unwanted workers get into the Object-worked-on: object
dangerous areas Location: none
Resource: crane
Nearby object: worker
Nearby action: none

The main hook and auxiliary hook work 007 Crane accident Action: lifting
simultaneously when operating a crane Object-worked-on: object
Location: none
Resource: main and auxiliary hook
Nearby object: none
Nearby action: none

Walking with objects which are 500 mm 008 Crane accident Action: walking with objects
above the ground Object-worked-on: object
Location: none
Resource: crane
Nearby object: none
Nearby action: none

(GB 50715-2011) and Quality and Safety Check Points of Urban Rail
Transit Engineering (2011) as laws and regulations must be adhered
to. In addition, records of 1731 accidents as as a result of work-
ers unsafe behaviors were gleaned from government records from
2008 to 2014 were examined to identify common recurring unsafe
acts. After analysing accident mechanisms, workers unsafe behav-
ior were extracted and used to develop a checklist. For instance,
a safety accident occured during the construction of the Kunming
(Yunnan province) metro line 2 in August 2011, whereby a worker
inhailed unknown harmful gases; this led to a fatality and two peo-
ple being critically injured. The unsafe behavior exacted in this
instance was in the process of munual insertion of pile founda-
tion, workers do not check whether there are harmful gases before
they begin to work. To complement the data derived from the
government and safety reports derived from the metro project,
common unsafe beahviors from safety management data (daily
safety check documents) on 98 construction sites from Shenyang,
Wuhan, Zhengzhou, and Shenzhen were also collected.
From the datasources presented above, a total of 522 unsafe
behaviors were identied and classied into themes: re ght-
ing (01), scaffolding (02), foundation pit supporting (03), formwork
engineering (04), tunnel construction by shield method (05), tunnel
construction by mining method (06), safety protection (07), tower
crane (08), gantry crane (09), building material elevator (10), lift-
ing (11), construction equipment (12), construction in tunnel track
area (13), manual hole digging pile (14), blasting (15), construc-
tion of electricity (16). An example of the partial unsafe behavior
of lifting classication is presented in Table 2.
In the case of the metro-system construction work is catogerised
in accordance with that advocated by Standard for Differentiating
Unit Work, Division Work and Subdivision Work of Rail Transportation
(QGD-004-2005) as: (1) unit work; (2) subunit work; (3) division
work; (4) subdivision work; (5) subentry work. The partial WBS
for a metro station is presented in Fig. 2. A typical metro station
contains six subunit works. Specically, the main structure con-
struction of open-cut metro station includes three division works:

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx 5

Metro
Metro Station
Station
Construction
Construction
01
01

Main
Main Structure
Structure Outdoor
Outdoor Building
Building
Construction
Construction of
of Open-
Open-
Environment
Environment
Cut
Cut Metro
Metro Station
Station 06
06
01
01

Foundation
Foundation Waterproof
Waterproof Main
Main Structure
Structure
Enclosure and
Enclosure and Engineering
Engineering 03
03
Treatment
Treatment 01
01 02
02

Foundation
Foundation Foundation
Foundation
Earthwork
Earthwork
Enclosure
Enclosure Treatment
Treatment
02
02
01
01 03
03

Precipitation
Precipitation and
and Earth
Earth Excavation
Excavation Support
Support Erection
Erection Earth
Earth Backfill
Backfi
f ll
Drainage
Drainage 002
002 003
003 004
004
001
001

Fig. 2. The partial WBS of unit work of metro station.

foundation enclosure and treatment, waterproof engineering and
main structure, foundation enclosure. Foundation enclosure and
treatment includes three subdivision works: foundation enclo-
sure, earthwork and foundation treatment. Earthwork includes
four subentry works: precipitation and drainage, earth excavation,
support erection and earth backll. By this way, work tasks are
subdivided in metro construction.
Classied unsafe behavior is encoded by the following rule:
XX-XX-XX-XX-XXX-XX-XXX (unit work-subunit work-division work-
subdivision work-subentry work-behavior type-unsafe behavior). For
instance, the unsafe behavior Not taking warning during lifting,
unwanted workers get into the dangerous areas arose during the
earth excavation using an open-cut method is recorded in 01-01-
01-02-002-11-006. Behavioral risk knowledge base is formed by
coding items to conduct behavior observation.
4.2. Data collection
4.2.1. Intelligent video surveillance
The video surveillance process is automated, but its use requires
extensive human intervention for analyzing and monitoring its
result; intelligence is required to automate this task. The core
of intelligent video surveillance is camera-based behavior anal-
ysis technology. This technology usually contains motion/object
detection, object classication, object tracking, behavior and activ-
ity analysis and understanding, person identication, and camera
handoff and data fusion (Brmond et al., 2006). Thus, workers par-
tial unsafe behavior can be recognized automatically from images
taken by cameras. The process of camera-based behavior analysis
involves the following steps: (1) behavior analysis function of the
camera is selected; (2) abnormal cases (unsafe behavior) is dened;
(3) observation areas on the worksite are debugged; (4) workers
unsafe behavior is tracked; (5) multiple-frame images are exacted.
The behavior observation system for metro construction work-
ers has been developed based on the intelligent video surveillance
(Fig. 3). Its hardware environment consists of a Webcam, Power
Over Ethemet (POE) module and Switch, while its software environ-
ment consists of camera-based behavior analysis technology. The
system identies workers unsafe behavior captured by webcams
according to preset rules and then, images with sematic informa-
tion upload promptly to the storage system.
4.2.2. A mobile application taking on-site photographs
Besides intelligent video surveillance, additional methods are
needed to collect data. During daily safety checks, managers iden-
tify potential safety risks by taking photographs. Some of these
photographs may identify workers unsafe acts, but they only
reect individuals observations and may be misinterpreted others.
Therefore, the mechanism to determine workers unsafe behavior
from on-site photographs requires an in-built intelligence function
that is able to ascertain semantics to comrm the inappropriate
behaviors.
Workers unsafe behavior may arise in different phases of con-
struction (Goh and Chua, 2009). Thus in alignment with BBS,
to identify unsafe behaviors the following steps are undertaken:
(1) determining workers action(s); (2) determining object(s) and
resource(s) undertaken during work activities; (3) identifying
location(s) of work. Factors extracted include action(s), object(s)-
worked-on, resource(s), location(s), nearby object(s) and nearby
action(s). Thus, to describe workers unsafe behavior, the following
semantic structure is used (Goh and Chua, 2009): Action(s) exe-
cuted on object(s)-worked-on using resource(s) at location(s) with
nearby object(s) and nearby action(s). Fig. 4 illustrates the unsafe
behavior of workers undertaking the activity of welding next to
an excavator that is in operation. Here the semantic structure is
Digging (action) executed on soil (object-worked-on), using excava-
tor (resource) in the excavation pit (location) with workers (nearby
object) doing welding work (nearby action)., so the potential safety
risk brought by this behavior can be clearly seen.
Training observers (i.e. site managers or workers) are asked to
describe collected photographs in accordance with the semantic
structure identied above. Then affecting factors are extracted from
the text and semantic information is labelled on each photograph.
All of the 522 unsafe behaviors are dened according to the
proposed semantic structure. The steps to undertaken via the VSM
process are: (1) all the keywords contained in the listed 522 unsafe
behaviors can be viewed as a document vector space. The fourth
column of Table 2 identies several keywords extracted from the

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
6 S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx

Fig. 3. The composition of intelligent video surveillance.

input (voice input method considers workers educational level).

The voice-recognition technology is used to transfer voice infor-
mation into written messages. After the observed photograph is
matched with one of the listed unsafe behaviors, its correspond-
ing code is determined according to current construction progress.
Then the photograph is uploaded with semantic information to the
storage system.

4.3. Data storage

To meet the demand for storing large amounts of behavior data,

Fig. 4. An example of photograph analysis.
unsafe acts; (2) the weight of each keyword is calculated by TF-IDF
method. A detailed description of TF-IDF method can be found in
Salton and Buckley (1988); (3) document vectors of listed unsafe
behavior from behavioral risk knowledge base are determined; (4)
query vector of keywords exacted from photograph description are
determined; (5) the photograph is matched with the right unsafe
behavior by calculating the angles between the query vector and
different document vectors by means of the Cosine Law. The doc-
ument vector of listed unsafe behavior k and the query vector of
observers photograph description l in the vector space are shown
as:
V = W1 , W2 , . . ., Wi , Wi+1 , . . ., Wj , Wj+1 , . . ., Wk , Wk+1 , . . ., Wl , Wl+1 , . . ., Wm , Wm+1 , . . ., Wn

      
i j k l

V (dk ) = W1 , W2 , . . ., WN
 time to obtain these images links with the data of the actual
queryVectorl = W1 , W2 , . . ., WN
i n means the component vectors of the six affecting factors in
the vector space, V (dk ) means the corresponding vector of listed
unsafe behavior k, queryVectorl means the corresponding query
vector of observers photograph description l, N means the Nth key-
word of vector space. Eq. (4) is the formula of the Cosine Law. By
calculating, the larger cosine between query and document vector,
the higher similarity between an observers photograph description
and listed unsafe behavior can be determined.
 
queryVectorl , dk
cos(queryVectorl , dk ) =    (4)
 queryVectorl   dk 
 
Therein, queryVectorl , dk means inner product of the
 
queryVectorl and dk ,  queryVectorl  means mold of the
 
queryVectorl ,  dk  means mold of the dk .
Ranking the possible matched unsafe behavior with their cosine
value from large to small, after the given query vector is identi-
ed, and enables each document vector to be calculated. The largest
value that is calculated identies the unsafe behavior. The mobile
application provides two input forms: (1) voice and (2) character
HDFS is used to store data from intelligent video surveillance and
the mobile application, behavioral risk knowledge base, schedule
management information and semantic information of identied
images (Fig. 5).
Intelligent video surveillance and mobile application include
unstructured (video screenshot and site photograph) and struc-
tured data (e.g. time and location). The behavioral risk knowledge
base contains a list of unsafe behaviors and the WBS from the
metro construction. Schedule management information is struc-
tured data and includes the actual project progress. The semantic
information of identied images contains structured data listed as
unsafe behavior, behavior type, possible injury and code, which are
(1)
   
m n
(2)
abstracted from behavioral risk knowledge base. In addition, the
(3)
progress, so the current phase in the WBS can be made explicit.
4.3.1. Process for data storage
HDFS adopts the master-slave structure, with the NameNode
as the master node and the DataNode as the slave node, one
NameNode can manipulate several DataNodes. The role of the
NameNode is to store the meta data information in the HDFS
and manage the block information in the DataNodes. The Meta
data information includes attribute information (the name of image
les, the collection time, location, the collector and the weather)
and le directory derived from the behavioral risk knowledge base.
Due to the default block size of the HDFS, which is designed to store
large les at 64 M, the data of unsafe behavior images may merge to
data blocks in 64 M. Thus the block information needs to be stored
in the NameNode. The DataNode is used to store the data blocks
and offer service to the client.
Every DataNode stores a couple of data blocks. The client
includes the ports of uploading and retrieving. The NameNode will
allocate free DataNodes, calling data blocks to store these struc-
tured and unstructured data, after the gathered unsafe behavior
images with complete semantic information have been entered

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx
Fig. 5. The framework of data sources and storage in HDFS.
into the uploading ports. When a user needs to query the data
stored, the rst step is to obtain the meta data conforming to the
query conditions in the NameNode from the retrieving port, and
then read image data with complete semantic information from
the data blocks in the DataNode. The NameNode maintains the
mapping information between the DataNode and the data block.
The HDFS guarantees the normal use of the stored data by using a
disaster tolerance. The default backup factor of the HDFS is 3; for
every data block, in addition to this DataNode, another DataNode
in this frame, and a DataNode in another frame will retain copies
of the data block. When the DataNode is in outage, tasks can be
distributed to data blocks in other DataNodes.
5. Implementation
Metro construction in Wuhan is experiencing a period of
rapid development; eight lines are being constructed simulta-
neously, workers on site are subjected to a dynamic and changing
work environment that is exacerbated by complex geological and
hydrological conditions. The schedule demands being placed upon
workers juxtaposed with the work environment provides a situ-
ation where unsafe behaviors are likely to occur due to prevailing
demands and constraints. To demonstrate a possible unsafe behav-
ior Not taking warning during lifting, unwanted workers get into
Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
7
the dangerous areas, the following examples are presented using
the intelligent video surveillance and mobile application.
Fig. 6 illustrates the application of intelligent video surveillance
at the Shuangdun Station on metro line 3. Fig. 6(a) denotes the
webcam as the front data acquisition equipment; Fig. 6(b) identi-
ed the spot of POE module and the switches, so as to guarantee
the data transmission and electric power supply. After the camera
is directed at the lifting area, the system backstage is set (Fig. 6c).
The behavior analysis engine of the camera is adjusted to the intru-
sion detection function; the ruled case is dened as unsafe behavior
above and pre-dened dangerous areas (the red circle position)
within the scope of monitor. Fig. 6(d) identies the process there-
after that the system recognized workers entry into dangerous
areas automatically and exacted multi-photographs by observing
workers motion.
In Fig. 7 the collection process of the photographs reecting
the same unsafe behavior via the mobile application is denoted.
Fig. 7(a) identies the observer takes the photograph reecting this
unsafe behavior; Fig. 7(b) presents a description process is inputted
voice or typing characters.
The description of the photograph above was Lifting executed
on object using crane at the foundation side with worker nearby,
with the factors being extracted were lifting, object, foundation
side, crane and worker.
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
8 S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx

Table 3
The computed result about the weight of vectors and Cosine value.

queryVector d1 d2 d3 d4 d5 d6 d7 d8

The weight of the 1st component 0.5713 0.9522 0.9522 0.9522 0.9522 0.9522 0.7141 0.9522 0
The weight of the 2nd component 0.6970 0 0 0 0 0 0.8713 1.1617 1.1617
The weight of the 3rd component 0.5924 0 0 0 0 0 0 0 0
The weight of the 4th component 0.5781 0.9635 0.9635 0.9635 0.9635 0.9635 0.7226 0 0.9635
The weight of the 5th component 0.0206 0 0 0 0 0 0.0257 0 0
The weight of the 6th component 0 0 0 0 0 0 0 0 0
Cosine value 0.6642 0.6642 0.6642 0.6642 0.6642 0.8751 0.7032 0.7399

The value of the best match is given in bold.

Fig. 7. The interface design of the mobile application.

The query vector of this description is dened by calculating

Fig. 6. The application of intelligent video surveillance on site.
the corresponding weights of each keyword by means of TF-IDF.
Document vectors of unsafe behavior in Table 1 were respectively
supposed as d1 d8 , then its components in six dimensions
were calculated. So the similarity of documents is judged by
their cosine value. The computed result presented in Table 3 is
cos(queryVector, d6 ) = 0.8751, which is the maximum. Therefore,
the best match is d6 named Not taking warning during lifting,
unwanted workers get into the dangerous areas. In addition, the
analysis of the other 514 unsafe behaviors via mobile application
provided a similar conclusion.
With complete semantic information, images collected using
intelligent video surveillance and a mobile application the infor-
mation obtained can be stored in the HDFS. To search and retrieve
information from the HDFS is a straightforward process. For exam-
ple, Fig. 8 illustrates the HDFS interface. Here in Fig. 8a the search
area can lter data according to different limitations such as time,
location, weather, collector, image type (video screenshot or site
photograph), behavior type (matched unsafe behavior) and possi-
ble injury. In Fig. 8b data is searched according to the limitations.
For example, under the limitation of time-9th October it can be
seen that there were a total of 23,867 images, but only 2197 of
them were effectively exacted because of unavoidable informa-
tion loss in the process of uploading and storage. In Fig. 8(c) the
limitation of behavior type is added by lifting. Behavior data can
be retrieved with complete semantic information, including iden-
tication (ID), time, image, collector, source, location, description
(in voice or characters only contained in on-site photographs col-
lected by the mobile application), unsafe behavior, behavior type

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx 9

Fig. 8. The interface of HDFS.

and possible injury. As a result, the visualization of behavior risk is
realized.
6. Enabling BBS by a Big-Data-based platform
Observation is an integrated function of BBS, however, as noted
above this activity can lead to subjective and bias decisions being
made. To improve its effectiveness, a Big-Data-based platform is
developed. The contribution of this platform is threefold:
(1) The development of behavioral risk knowledge base is a novel
idea for construction. It can not only be used as a reference
for behavior observation, but also as a tool to proactively man-
age safety site. More specically, when combined with a WBS,
workers unsafe behavior can be dynamically classied and
coded when the construction stage changes.
(2) Methods for effectively collecting data are improved. Dif-
ferent ways of monitoring usage can be realized by using
camera behavior analysis technology to observe unsafe acts

Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
10 S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx
and determine by pre-set rules. In this instance, machines
replace manual observation, to solve time-consuming prob-
lems and labor-intensive activities previously undertaken. The
mobile application was developed to facilitate the capturing of
the photographs during manual observation, to improve data
sparseness of intelligent video surveillance and ensure suf-
cient samples of behavioral data. JHA and VSM are used as
effective methods to determine the unsafe behavior in accord
with photographs.
(3) The behavioral data storage by HDFS improves the operating
efciency of developed platform, and ensures fast reading and
writing of data services so as to guarantee retrieval of behavioral
data for managers to make timely decisions. Thus unsafe acts
can be visualized in real-time.
However, several limitations exist and thus need to be acknowl-
edged. For camera-based behavior analysis technology, a single
camera over a period of time only recognizes a specic kind
of unsafe behavior, so different periods of construction need to
change preset rules frequently. JHA and VSM are used to match
photographs description with the corresponding unsafe behav-
ior. However, since the language usage of individuals can differ,
dening the photograph description may confuse observers cog-
nitive conditioning to a certain behavior. While this platform can
assist with the collection and storage of workers behavioral data
observed from the eld, it ignores internal factors, such as emo-
tions, attitudes and mood states, which can also inuence an
individuals behavior.
Future research should consider how a single camera can iden-
tify a variety of unsafe behaviors and the integration with motion
perception technology that can automatically determine more
types of unsafe acts. Also, a synonyms library should be built to
eliminate data omissions caused by different words with synony-
mous meanings. In contrast to BBS, People-Based Safety (PBS) goes
beyond observable behavior to address the human dynamics of
workplace safety. Thus data collection should not only be derived
from observations on site, but also from questionnaires and inter-
views to obtain a better understanding of unobservable factors such
as peoples emotions, and cognitive disposition.
7. Conclusion
People are the main participants of construction; their behav-
ioral safety needs to be given utmost attention to reduce the
rate of accidents and incidents. BBS is an effective approach to
decrease workers unsafe behavior, through this has been predomi-
nantly undertaken using observation. By combining the traditional
behavior observation with advanced technologies, a Big-Data-
based platform for workers behavior observation was developed
to overcome the limitations associated with traditional BBS. The
behavioral risk knowledge base was formed, which can provide
the basis of behavior observation to eliminate the misunderstand-
ing of the same behavior with different observers. Intelligence
video surveillance and a mobile application are both used to
collect workers unsafe behavior; data acquisition is more compre-
hensive and convenient when these technologies are used. They
provide the possibility to further analyze workers unsafe behavior,
and are more intuitive as they are expressed by visual language.
Moreover, HDFS provides a solution with characteristics of high
fault-tolerance and high throughput to solve data storage issues
that may arise from using Big Data.
The developed platform has been applied during the con-
struction of a metro project. The data collected by intelligent
video surveillance and mobile applications in daily observations
amounted to over 300GB, containing over 150.000 of image data
Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
reecting workers unsafe behavior with complete semantic infor-
mation. The number of images reecting unsafe acts exacted from
intelligent surveillance video is signicantly higher than from the
on-site photographs. Though with the accumulation of such data,
cannot guarantee the complete collection of workers unsafe behav-
ior.
The developed platform has used advanced technologies to
ensure the use of Big Data can be used by other construction
projects in China. With the collection of more data, workers behav-
ioral patterns can be identied so as to lay good foundation for
behavior modication in future research.
Acknowledgements
The presented work has been supported by the National Science
Foundation of China (NSFC) through grant 71471072. The authors
gratefully acknowledge NSFCs support. The authors would also
like to thank Wuhan Metro Group Co., Ltd. and Wuhan Municipal
Construction Group Co., Ltd. for supporting this study. The authors
would like to thank the three anonymous reviewers for their con-
structive and helpful comments.
References
Aguilar, G.E., Hewage, K.N., 2013. IT based system for construction safety
management and monitoring: C-RTICS 2. Autom. Constr. 35, 217228.
Beijing MTR Construction Administration Corporation, 2005. Standard for
Differentiating Unit Work, Division Work and Subdivision Work of Rail
Transportation, Retrieved from: http://wenku.baidu.com/view/
1acc38134431b90d6c85c7dc.html?re=view.
Brmond, F., Thonnat, M., Zniga, M., 2006. Video-understanding framework for
automatic behavior recognition. Behav. Res. Methods 38 (3), 416426.
Cameron, I., Duff, R., 2007. A critical review of safety initiatives using goal setting
and feedback. Constr. Manag. Econ. 25 (5), 495508.
Chen, D., Tian, H., 2012. Behavior based safety for accidents prevention and
positive study in china construction project. Proc. Eng. 43, 528534.
Chi, S., Caldas, C.H., 2011. Automated object identication using optical video
cameras on construction sites. Comput.-Aided Civil Infrastruct. Eng. 26 (5),
368380.
Choudhry, R.M., 2012. Implementation of BBS and the impact of site-level
commitment. J. Prof. Issues Eng. Educ. Pract. 138 (4), 296304.
Choudhry, R.M., 2014. Behavior-based safety on construction sites: a case study.
Accid. Anal. Prevent. 70, 1423.
Cox, S., Jones, B., Rycraft, H., 2004. Behavioural approaches to safety management
within UK reactor plants. Saf. Sci. 42 (9), 825839.
DePasquale, J.P., Geller, E.S., 2000. Critical success factors for behavior-based safety:
a study of twenty industry-wide applications. J. Saf. Res. 30 (4), 237249.
Gantz, J., Reinsel, D., 2011. Extracting value from chaos. IDC iView, 112.
Geller, E.S., 2001. Keys to Behavior-based Safety. Government Institutes.
Geller, E.S., 2005. Behavior-based safety and occupational risk management.
Behav. Modif. 29 (3), 539561.
Ghemawat, S., Gobioff, H., Leung, S.T., 2003. The google le system. SIGOPS Oper.
Syst. Rev. 37 (5), 2943.
Goh, Y.M., Chua, D.K.H., 2009. Case-based reasoning for construction hazard
identication: case representation and retrieval. J. Constr. Eng. Manag. ASCE
135 (11), 11811189.
Han, S., Lee, S., 2013. A vision-based motion capture and recognition framework
for behavior-based safety management. Autom. Constr. 35, 131141.
Han, S., Lee, S., Pena-Mora, F., 2011. Application of dimension reduction techniques
for motion recognition: construction worker behavior monitoring. In:
Proceedings of the 2011 ASCE International Workshop on Computing in Civil
Engineering, pp. 1922.
Harmandas, V., Sanderson, M., Dunlop, M.D., 1997. Image retrieval by hypertext
links. In: Proceedings of the ACM SIGIR Forum, pp. 296303.
Heinrich, H.W., Petersen, D., Roos, N., 1950. Industrial Accident Prevention.
McGraw-Hill, New York.
Hu, B., Dasmahapatra, S., Lewis, P., Shadbolt, N., 2003. Ontology-based medical
image annotation with description logics. In: Proceedings of the Tools with
Articial Intelligence in 15th IEEE International Conference, pp. 7782.
Hu, H.Y., Qu, Z.W., Li, Z.H., 2014. Multi-level trajectory learning for trafc behavior
detection and analysis. J. Chin. Inst. Eng. 37 (8), 9951006.
Ismail, F., Hashim, A.E., Zuriea, W., Ismail, W., Kamarudin, H., Baharom, Z.A., 2012.
Behaviour based approach for quality and safety environment improvement:
Malaysian experience in the oil and gas industry. Proc.-Soc. Behav. Sci. 35,
586594.
Ko, T., 2008. A survey on behavior analysis in video surveillance for homeland
security applications. In: Proceedings of 37th IEEE Applied Imagery Pattern
Recognition Workshop, New York, pp. 8491.
G Model
AAP-3927; No. of Pages 11 ARTICLE IN PRESS
S.Y. Guo et al. / Accident Analysis and Prevention xxx (2015) xxxxxx
Krause, T., Seymour, K., Sloat, K., 1999. Long-term evaluation of a behavior-based
method for improving safety performance: a meta-analysis of 73 interrupted
time-series replications. Saf. Sci. 32 (1), 118.
Lablack, A., Djeraba, C., 2008. Analysis of human behaviour in front of a target
scene. In: Proceedings of 19th International Conference on Pattern
Recognition, New York, pp. 38313834.
Laitinen, H., Marjamki, M., Pivrinta, K., 1999. The validity of the TR safety
observation method on building construction. Accid. Anal. Prevent. 31 (5),
463472.
Li, H., Lu, M., Hsu, S.-C., Gray, M., Huang, T., 2015. Proactive behavior-based safety
management for construction safety improvement. Saf. Sci. 75,
107117.
Lingard, H., Rowlinson, S., 1998. Behavior-based safety management in Hong
Kongs construction industry. J. Saf. Res. 28 (4), 243256.
Lingard, H., Rowlinson, S.M., 2005. Occupational Health and Safety in Construction
Project Management. Taylor & Francis.
Mayer-Schnberger, V., Cukier, K., 2013. Big Data: A Revolution that will Transform
How We Live, Work, and Think. Houghton Mifin Harcourt.
McSween, T.E., 2003. Values-based Safety Process: Improving Your Safety Culture
with Behavior-based Safety. John Wiley and Sons.
Ministry of Housing and Urban-Rural Development of the Peoples Republic of
China, 2011a. Quality and Safety Check Points of Urban Rail Transit
Engineering, Retrieved from: http://www.zgjsjl.org.cn/uploadle/2011 12/
temp 11121215128737.pdf.
Ministry of Housing and Urban-Rural Development of the Peoples Republic of
China, 2011b. Standard for Construction Safety Assessment of Metro
Engineering (GB 50715-2011), Retrieved from: http://www.spsp.gov.cn/page/
CN/2011/GB%2050715-2011.shtml.
Nielsen, K.J., Kines, P., Pedersen, L.M., Andersen, L., Andersen, D.R., 2015. A
multi-case study of the implementation of an integrated approach to safety in
small enterprises. Saf. Sci. 71, 142150.
Please cite this article in press as: Guo, S.Y., et al., A Big-Data-based platform of workers behavior: Observations from the eld. Accid.
Anal. Prev. (2015), http://dx.doi.org/10.1016/j.aap.2015.09.024
11
Saba, F., Mohamed, Y., 2008. Implementing incident prevention (leading
indicators) program in an industrial construction company. In: Proceedings of
the Congres Annuel, pp. 1013.
Salton, G., Buckley, C., 1988. Term-weighting approaches in automatic text
retrieval. Inform. Process. Manag. 24 (5), 513523.
Salton, G., Wong, A., Yang, C.-S., 1975. A vector space model for automatic
indexing. Commun. ACM 18 (11), 613620.
Shvachko, K., Kuang, H., Radia, S., Chansler, R., 2010. The hadoop distributed le
system. In: Proceeding of 26th IEEE Mass Storage Systems and Technologies
(MSST) Symposium, Incline Village, pp. 110.
Skinner, B.F., 1953. Science and Human Behavior. Simon and Schuster.
Teizer, J., Cheng, T., Fang, Y., 2013. Location tracking and data visualization
technology to advance construction ironworkers education and training in
safety and productivity. Autom. Constr. 35, 5368.
Turnbeaugh, T.M., 2010. Improving business outcomes: behavior-based safety
techniques can inuence organizational performance. Prof. Saf. 55 (3), 4149.
Wang, H.H., Boukamp, F., 2011. Ontology-based representation and reasoning
framework for supporting job hazard analysis. J. Comput. Civil Eng. 25 (6),
442456.
Williams, J.H., Geller, E.S., 2000. Behavior-based intervention for occupational
safety: critical impact of social comparison feedback. J. Saf. Res. 31 (3),
135142.
Yeow, P.H.P., Goomas, D.T., 2014. Outcome-and-behavior-based safety incentive
program to reduce accidents: a case study of a uid manufacturing plant. Saf.
Sci. 70, 429437.
Yin, R.K., 2013. Case Study Research: Design and Methods. SAGE Publications.
Zhang, D., Islam, M.M., Lu, G., 2012. A review on automatic image annotation
techniques. Pattern Recognit. 45 (1), 346362.
Zhang, M.Z., Fang, D.P., 2013. A continuous behavior-based safety strategy for
persistent safety improvement in construction industry. Autom. Constr. 34,
101107.