See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/337705605

A fully automated vision-based system for real-time personal protective

detection and monitoring

Conference Paper · November 2019

CITATIONS READS

9 5,939

3 authors, including:

Thi Lan Le Hong Hoang Si

Hanoi University of Science and Technology Hanoi University of Science and Technology
177 PUBLICATIONS 1,559 CITATIONS 51 PUBLICATIONS 1,047 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Thi Lan Le on 18 August 2020.

The user has requested enhancement of the downloaded file.

A fully automated vision-based system for real-time
personal protective detection and monitoring
Quang-Huy Tran Thi-Lan Le Si-Hong Hoang
School of Electrical Engineering MICA Institute School of Electrical Engineering
HUST HUST HUST
Hanoi, Vietnam Hanoi, Vietnam Hanoi, Vietnam
[email protected] [email protected]

Abstract—Construction had the most fatal occupational in- using has to be conducted not only on the construction field
juries out of all industries due to the high number of annual but also at site entry.
accidents. There are many solutions to ensure workers’ safety and Besides the PPE detection, the face and identity of the
limit these accidents, one of which is to ensure the appropriate
use of appropriate personal protective equipment (PPE) specified workers have to be defined. However, this is usually per-
in safety regulations. However, the monitoring of PPE use that formed by manually or by others technologies that require
is mainly based on manual inspection is time-consuming and extra hardware plate-form such as RFID (Radio Frequency
ineffective. This paper proposes a fully automated vision-based Identification). In this paper, for the first time, we propose a
system for real-time personal protective detection and monitor- fully automated vision-based PPE detection and monitoring.
ing. The proposed system consists of two main components: PPE
detection and face detection and recognition. Several experiments The proposed system consists of two main components: PPE
have been conducted. The obtained detection accuracy for 6 main detection and face detection and recognition. The main aim of
PPEs is up to 98% while that for face detection and recognition PPE detection is to determine the presence of required PPE
is 96%. The obtained results have demonstrated the ability to while face detection and recognition aims to determine the
detect PPE and to recognize a face with high precision and recall identity of the workers.
in real-time.
Index Terms—PPE detection, deep learning, object detection, II. R ELATED WORKS
automatic monitoring.
The enhancement of onsite construction safety has been
I. I NTRODUCTION increasingly received the attention of researchers and industrial
practitioners. The proposed solutions can be divided into two
Construction has been identified as one of the most dan- groups: a noncomputer vision-based and computer vision-
gerous job sectors. In Vietnam, according to the Report of based technique.
Ministry of labour-invalids, and social affairs in 2017, there The work of Kelm et al. [5] falls into the first group.
were 8956 occupational accidents nationwide, causing 9173 The authors designed an RFID-based portal to check whether
victims, of which 5.4% of cases involved not wearing personal the workers’ personal protective equipment (PPE) complied
protection equipment [1]. with the corresponding specifications. Dong et al. [6] use
Several onsite safety regulations have been established to real-time location system (RTLS) and virtual construction are
ensure construction workers’ safety. In the safety regulations, developed for worker’s location tracking to decide whether
the appropriate use of appropriate personal protective equip- the worker should wear a helmet and give a warning, while
ment (PPE) is specified and the contractors must ensure that the silicone single-point sensor is designed to show whether
the regulations are enforced through the monitoring process. the PPE is used properly for a further behavior assessment.
The monitoring of the use of PPE is normally conducted in However, these methods are limited in their respective ways.
two areas: the site entry and the onsite construction field. For example, the worker’s identification card only indicates
Nowadays, most of the construction fields conduct the that the distance between the worker and PPE is close and the
monitoring of PPE using manually by inspectors. This work loss of sensors may be a consideration when applying.
is tedious, time-consuming and ineffective due to the high Concerning computer vision-based techniques, taking into
number of workers to monitor in the field. account the important role of hardhat, several works have been
Recently, several technologies have been proposed to en- done for hardhat detection. Rubaiyat, et al. [7] incorporate a
hance the construction safety. Among the proposed solutions, Histogram of Oriented Gradient (HOG) with Circle Hough
computer vision has been widely used [2], [3], [4]. However, Transform (CHT) to obtain the features of workers and hard-
most of recent works focus on detecting the use of hardhat on hats. Du et al. [8] combine face detection and hardhat detection
the onsite construction field. Besides hardhat, others equip- based on color information.
ment such as glove, shoes need to be detected in order to In recent years, deep learning developed extremely fast
ensure the worker safety. Moreover, the monitoring of PPE based on a huge amount of training data and improved com-
puting capabilities of computers. The results for the problem of B. PPE and face detection
classification or detection of objects are increasingly improved.
The object detection methods can be divided into two main In our work, we employ the YOLO network for PPE and
approaches: one-stage (e.g., Faster Region-based Convolu- face detection. YOLO network has been introduced by Joseph
tional Neural Networks (Faster R-CNN) [9]) and two-stage Remon’s team [10] for object detection.
detection (e.g., You Only Look Once (YOLO) [10], Single The first version of YOLO is named YOLO v1. Different
Shot Multibox Detector (SSD) [11]). In general, one-stage from two-stage methods, the core idea behind this fast detector
detectors is faster and simpler. Inspired by the performance is a single convolutional network consisting of convolutional
of deep learning-based object detection method, in [12], [13], layers followed by 2 fully connected layers. A single neural
Fang et al. applied the Faster R-CNN algorithm to detect the network predicts bounding boxes and class probabilities di-
absence of hardhats and discovering non-certified work. The rectly from full images in one evaluation. For this, YOLO
proposed method has been evaluated in different situations and network divides the image into an S × S grids. For each
shown its high performance. grid cell, it predicts B bounding boxes, their confidence
However, most of the current works focus on detecting the score and C class probabilities as illustrated in Fig 2. The
use of hardhat on the construction site. Besides hardhat, other design of YOLO enables end-to-end training and real-time
equipment have to be detected. Moreover, to make sure that speeds while maintaining high average precision but still
the workers use proper PPE, the present of PPE should be has limitations when detecting the small objects appeared in
checked at the entry point of the construction field. groups. Therefore, YOLO v2 is introduced [14] and after that
is YOLO v3 [15]. YOLO v3 has significant improvements in
both accuracy and speed, improved the capable of detecting
III. P ROPOSED FRAMEWORK
small objects.
A. Overview
Currently, monitoring the use of personal protective equip-
ment is still done manually. It is costly and inaccurate because
a large number of workers need to be checked over a period
of time. In response to these restrictions, the overall objective
of this paper is to propose a novel solution to address the un-
resolved problem of reliably identifying workers who comply
with safety regulations at the entry point of the construction
field. The proposed system illustrated in Fig. 1 consists of two
main steps: PPE detection and face detection and recognition.
The main aim of PPE detection is to determine the presence
of required PPEs while face detection and recognition aim to
determine the identity of the workers. Inspired from impressive
results of deep convolutional neural network for different
computer vision tasks, in this paper, PPEs and face are detected
Fig. 2: Illustration of PPE or face detection based on YOLO
thanks to YOLO (You only look once) network while FaceNet
network.
is fine-tuned for face recognition.

The YOLOv3 uses a variant of Darknet, which originally

a 53 layer network trained on Imagenet called Darknet-53.
Similar to the VGG models it uses mostly 3 × 3 filters
and double the number of channels after every pooling step.
In addition, it uses Network in Network (NIN) and batches
normalization to stabilize the training process, speed up con-
vergence, and regularize the model. Its structure is detailed on
the layer type, input, output, filter number, filter size (including
stride), as shown in Table I. We finetune YOLO from the
pre-trained model on Imagenet for PPE and face detection
with the weight decay of 0.0005, the momentum of 0.9, the
learning rate of 10−3 . We also employ data augmentation
Fig. 1: The proposed system for real time PPE detection and including random crops, rotations, and hue, saturation, and
monitoring at entry point of the construction field. exposure shifts. Besides, we use Multi-Scale Training for each
of 10 batches with image dimension size to improve network
accuracy with a variety of input dimensions [14].
 TABLE I: Network architecture YOLOv3 accuracy. In our work, we propose to use YOLO for detecting
Type Filters Size Output
Convolutional 32 3x3 256x256
Convolutional 64 3x3/2 128x128
Convolutional 32 1x1
1x Convolutional 64 3x3
Residual 128x128
Convolutional 128 3x3/2 64x64
Convolutional 64 1x1
2x Convolutional 128 3x3
Residual 64x64
Convolutional 256 3x3/2 32x32
Convolutional 128 1x1
8x Convolutional 256 3x3 Fig. 3: Structure of one output cell in YOLO
Residual 32x32
Convolutional 512 3x3/2 16x16
Convolutional 256 1x1
8x Convolutional 512 3x3 6 main PPEs at the entry point of the construction field
Residual 16x16 and workers face (see Fig. 4). The use of YOLO has two
Convolutional 1024 3x3/2 8x8 main advantages in comparison with state of the art methods
Convolutional 512 1x1
4x Convolutional 1024 3x3 proposed for PPE detecting. Firstly, YOLO is very fast. It can
Residual 8x8 obtain the frame rate of 20fps at 416 × 416 resolution on a
Avgpool Global GPU-supported work station.
Connected 1000
Softmax Secondly, YOLO sees the entire image during training
and test time. This is different from region proposal-based
methods which only consider features within the bounding
The newer architecture boasts of residual skip connections boxes. For that reason, YOLO has less than half the number
and upsampling. The most salient feature of v3 is that it of background errors than Fast R-CNN. [16].
makes detections at three different scales. YOLO is a fully
convolutional network and its eventual output is generated
by applying a 1 x 1 kernel on a feature map. In YOLO v3,
the detection is done by applying 1 x 1 detection kernels on
feature maps of three different sizes at three different places
in the network. The detection filters are calculated according
to the number of classes and the attributes of the bounding
box required as Equ. 1
F ilters = bboxes ∗ (5 + classes) (1)
bboxs is the number of bounding box per cell, classes are the
number of classes and “5” is for the 4 bounding box attributes
and one object confidence as shown in Fig. 3. In this work,
bboxes are 3 and classes are 7. In YOLO v3 trained on this Fig. 4: Faces and six types of PPEs: hardhat, shirt, gloves,
dataset, 3 bounding boxes and 7 classes, so the kernel size belt, pant, shoes.
is 1 x 1 x 36. YOLO v3 makes a prediction at three scales,
which are precisely given by downsampling the dimensions
of the input image by 32, 16 and 8 respectively. And at each It is important to note that the PPE and face detection can be
scale gives us the corresponding output features map : 13 x done by using any image captured when the worker comes into
13 x 36, 26 x 26 x 36 and 52 x 52 x36. The 13 x 13 layer the monitoring zone. However, to increase the reliability of
is responsible for detecting large objects, whereas the 52 x the system, we determine the detection results from an image
52 layer detects the smaller objects, with the 26 x 26 layer sequence based on the detection result of each image through
detecting medium objects. Here is a comparative analysis of voting technique. For this, a series of images during 3s are
different objects picked in the same object by different layers. collected and processed. The number of frames is from 20
The structure of an output cell as illustrated in Fig 3. From to 40 depending on the speed of the person. For each class,
the output of the last layer, several bounding boxes can be the ratio between the number of frames where the interested
generated for the same object. To filter out these bounding class is detected and the total of frames. If the ratio is equal
boxes, we use Non-maximum suppression and threshold. After or less than 0.5, the system will confirm the detection of the
that, we get the coordinates of bounding boxes with their class. An example is illustrated in Fig 5. The green frames
highest probability class without duplicating the bounding corresponds to the images that are detected correctly while
boxes of the same object. These bounding boxes will be the red frames are images with incorrect detection (e.g. miss
compared with those of ground-truth to compute the model’s detection of gloves and shoes).
 Thanks to the voting technique, correct detection decisions The FaceNet model is based on the Inception which is the
can be made in some cases even with false alarm and miss winning approach for ImageNet includes to extract features
detection at some images in the sequence. from images then uses an SVM for classification. The special
in FaceNet is that it uses the Triplet loss (refer to Fig. 6) to
minimize distance between similar faces and maximize the
distance to faces that are not similar, so the FaceNet can
distinguish very accurate person to person. FaceNet consists
of a batch input layer and a deep CNN followed by L2
normalization, which results in the face embedding.
A training triplet contains three exemplars: anchor (an image
of a person), positive (another image of the same person)
and negative (an image of a different person). The triplet loss
measures two Euclidean distances: one distance between the
anchor and the positive image named A and another between
the anchor and the negative image named B. The training
Fig. 5: Illustration of the voting technique used for PPE and process aims to reduce A while maximise B, such that similar
face detection. The green frame corresponds to the images in images lie close to each other and distinct images lie far away
which the PPE and face are detected correctly while the red in the embedding space as illustrated in Fig 7.
frames are images with incorrect detection (miss detection of
gloves).

C. Face recognition
FaceNet is introduced in 2015 by Google researchers, which Fig. 7: The triplet loss minimizes the distance between an
is a start-of-art face recognition, verification, and clustering anchor and a positive, both of which have the same identity,
neural network. It has a 22-layers deep neural network that and maximizes the distance between the anchor and a negative
directly trains its output to be a 128-dimensional embedding. of a different identity.
Once the FaceNet model having been trained with triplet loss
for different classes of faces to capture the similarities and
differences between them, the 128-dimensional embedding The embedding is represented by f(x) ∈ Rd . It embeds an
returned by the FaceNet model can be used to clusters faces image x into a d-dimensional Euclidean space. The triplet loss
effectively. Once such a vector space (embedding) is created, learns to ensure that an image xai (anchor) of a specific person
tasks such as face recognition, verification and clustering can is closer to all other images xpi (positive) of same person than
be easily implemented using standard techniques with FaceNet it is to any image xni (negative) of any other person. This is
embeddings as feature vectors. In a way, the distance would visualized in Fig 7. The loss fusion L that is being minimized
be closer for similar faces and further away for non-similar is as follows:
faces. X n
kf (xai ) − f (xpi )k22 − kf (xai ) − f (xni )k22 + α
 
(2)
i

where α is a margin that is enforced between positive and

negative pairs.

IV. E XPERIMENTAL RESULTS

A. Dataset
There is no off-the-shelf PPE dataset available and because
the standards of personal protective equipment for construction
site workers are clearly defined according to the nature of the
works, we selected 6 types of typical protective equipment and
face of workers in the construction site to create a data set for
training and evaluation as illustrated in Fig. 4 . The dataset
is collected outdoors by IP camera and collected over several
days at different times so it has different lighting condition.
The image acquisition camera is placed 2,5 meters high, the
Fig. 6: Illustration of face recognition based on FaceNet [17] length of the road segment is from 3 to 4 meters. The images
are captured at a rate of 10 FPS (frames per second). After
collecting full images of models, we label each object on TABLE IV: Obtained results for PPE and face detection with
images from our dataset with the annotation tool Yolo-mark different numbers of iterations.
[18]. This work included marking bounding boxes of objects Iteration number
and generating annotation files that contain coordinates and 1,000 2,000 3,000 4,000 5,000
Precision 0.85 0.92 0.94 0.98 0.95
box size parameters along with the label type. The dataset is Recall 0.86 0.92 0.99 0.96 0.99
collected with 10 subjects. The dataset is divided into 2 main F1 Score 0.85 0.92 0.96 0.97 0.97
parts: training set and testing set. The training set consists
of 12,000 images of 6 types of PPE and 4,500 images of
10 different subjects. The number of testing images for PPE Concerning face recognition, once a face detection module
detection is 1200 images while those for face detection andhas detected the faces in images, the face recognition method
recognition is 500. Detail information is given in Table II and
based on FaceNet is executed. The face detection results are
Table III. shown in the right-most column in Table V while the obtained
face recognition accuracy is 96% for 500 testing images. The
TABLE II: Number of images for PPE detetion training and
accuracy of face recognition is quite high. However, in some
testing
cases, the face is not correctly detected due to its small size
Hardhat Shirt Glove Belt Pant Shoe Total in images (see Fig. 8) or the presence of emotions (see Fig.
Training 3,000 2,000 2,000 2,000 1,000 2,000 12,000
Testing 350 250 200 150 100 150 1,200
9).

TABLE III: Number of images used for face detection and

recognition training and testing
#subjects #face images for each subject Total number
Training 10 450 4,500
Testing 10 50 500

B. Evaluation measures
The metrics that we use to evaluate the model’s accuracy are Fig. 8: Face is not correctly detected due to the small size.
Precision, Recall and F1-score for object and face detection
and accuracy metric for face classification. Firstly, we have to
define the meaning of TP (true positive), FP (false positive),
and FN (false negative). True Positive results when an object
is correctly identified with IOU (Intersection over Union)
between ground truth and bounding box predicted to be greater
than a threshold (in our case, this threshold is set by 0.5).
False Positive is the result of the wrong identity, which means
that the wrong class can be identified, or IOU < 0.5. False
Negative is the result of miss identification, meaning that the
object appears but is not recognized.
Precision, Recall, and F1-score are defined as follows: Fig. 9: Face is detected but incorrectly recognized due to the
2 ∗ P recision ∗ Recall presence of emotions.
F1 score = (3)
P recision + Recall
TP TP 2) Evaluation of whole system: To evaluate the whole
where P recision = T P +F P , Recall = T P +F N
system, we prepare 4 scenarios. In all scenarios, the subjects
C. Experimental results do not wear all required PPE. In this evaluation, the chosen
1) Evaluation of PPE detection and face detection and network resolution is 416 × 416. For each scenario, we
recognition at image level: To determine the number of iter- perform 5 tests and evaluate the accuracy that is the ratio
ation for the network, we evaluate the network with different between the correct recognized test and the total number of
iterations. The obtained results are shown in Table IV with tests. One test is considered as a correct recognized test if all
network resolution being 416 × 416 and threshold equal to interested PPE are correctly detected and the identity of the
0.25. As is shown, training with a large number of iterations person is correctly classified. As shown in Tab. VI, thanks to
gives better results and allows us to avoid over-fitting. Based the voting technique and the robustness of PPE detection and
on these results, we use the weights after 4,000 iterations. face recognition, the system produced the correct results even
The performance of the method on the testing dataset with some PPEs are omitted or incorrectly detected and the identity
three network resolutions is shown in Tab. V. of the person is wrong classified in some frames.
 TABLE V: Obtained results for PPE and face detection with different network resolutions.
Network resolution Frame rate Evaluation metric Hardhat Shirt Glove Belt Pant Shoes Face
320 × 320 10Hz FN 8 5 0 5 5 0 0
FP 10 10 14 12 13 15 0
TP 332 245 210 155 105 150 500
Precision 0.97 0.96 0.94 0.93 0.89 0.91 1
Recall 0.98 0.98 1 0.97 0.95 1 1
F1 Score 0.97 0.97 0.97 0.95 0.92 0.95 1
416 × 416 8Hz FN 6 5 2 6 6 0 0
FP 5 7 14 12 8 13 0
TP 334 245 208 154 104 150 500
Precision 0.99 0.97 0.94 0.93 0.93 0.92 1
Recall 0.98 0.98 0.99 0.96 0.94 1 1
F1 Score 0.98 0.97 0.96 0.94 0.93 0.96 1
608 × 608 6Hz FN 10 5 2 6 6 0 0
FP 6 7 15 11 14 15 0
TP 330 245 208 154 104 150 500
Precision 0.98 0.97 0.93 0.93 0.88 0.91 1
Recall 0.97 0.98 0.99 0.96 0.94 0.99 1
F1 Score 0.97 0.97 0.96 0.94 0.91 0.95 1

TABLE VI: Detection and recognition results of the proposed[5] A. Kelm, L. Laußat, A. Meins-Becker, D. Platz, M. J. Khazaee,
system with different scenarios. A. M. Costin, M. Helmus, and J. Teizer, “Mobile passive radio
frequency identification (rfid) portal for automated and rapid control of
Types of equipment #tests #false alarms
Accuracy (%) personal protective equipment (ppe) on construction sites,” Automation
Hardhat, shirt, belt, pant, shoe 5 0 100 in Construction, vol. 36, pp. 38 – 52, 2013. [Online]. Available:
Hardhat, shirt, pant, shoe 5 0 100 http://www.sciencedirect.com/science/article/pii/S0926580513001234
Shirt, glove,belt, pant, shoe 5 0 100 [6] S. Dong, Q. He, H. Li, and Q. Yin, Automated PPE Misuse Identification
Hardhat, pant, shoe 5 0 100 and Assessment for Safety Performance Enhancement. [Online]. Available:
https://ascelibrary.org/doi/abs/10.1061/9780784479377.024
[7] A. H. M. Rubaiyat, T. T. Toma, M. Kalantari-Khandani, S. A. Rahman,
L. Chen, Y. Ye, and C. S. Pan, “Automatic detection of helmet uses for
V. C ONCLUSIONS AND F UTURE WORKS construction safety,” in 2016 IEEE/WIC/ACM International Conference on
Web Intelligence Workshops (WIW), Oct 2016, pp. 135–142.
Construction activities are dangerous and very risky. Work[8] S. Du, M. Shehata, and W. Badawy, “Hard hat detection in video sequences
safety management in construction despite being concerned, based on face features, motion and color information,” in 2011 3rd Interna-
still leaves many occupational accidents. From the practical tional Conference on Computer Research and Development, vol. 4, March
2011, pp. 25–29.
needs and the urgent requirements laid down at the construc-[9] S. Ren, K. He, R. B. Girshick, and J. Sun, “Faster R-CNN: towards real-time
tion sites, we have developed a PPE detection and identifica- object detection with region proposal networks,” CoRR, vol. abs/1506.01497,
tion system to replace manual inspection and monitoring, and 2015. [Online]. Available: http://arxiv.org/abs/1506.01497
[10] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi, “You only look
allow management and retrieval of data. Along with that is to once: Unified, real-time object detection,” CoRR, vol. abs/1506.02640, 2015.
build a database of images of 6 types of common protective [Online]. Available: http://arxiv.org/abs/1506.02640
equipment at construction sites. This database along with [11] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, and A. C.
labeling information can be used for the research community. Berg, “SSD: single shot multibox detector,” CoRR, vol. abs/1512.02325,
2015. [Online]. Available: http://arxiv.org/abs/1512.02325
The results of detection and identification on the test dataset [12] Q. Fang, H. Li, X. Luo, L. Ding, H. Luo, T. Rose, and W. An, “Detecting
show that the model gives good results in both accuracy and non-hardhat-use by a deep learning method from far-field surveillance videos,”
Automation in Construction, vol. 85, pp. 1–9, 01 2018.
miss rate. In the future, we will expand the data set with [13] Q. Fang, H. Li, X. Luo, L. Ding, T. M. Rose, W. An,
other equipment with different conditions. Furthermore, the and Y. Yu, “A deep learning-based method for detecting
code will be optimized to increase the speed of the system. non-certified work on construction sites,” Advanced Engineering
Informatics, vol. 35, pp. 56 – 68, 2018. [Online]. Available:
R EFERENCES http://www.sciencedirect.com/science/article/pii/S1474034617303166
[14] J. Redmon and A. Farhadi, “YOLO9000: better, faster, stronger,” CoRR, vol.
[1] M. of labour invalids and social affairs. (2018) Notification abs/1612.08242, 2016. [Online]. Available: http://arxiv.org/abs/1612.08242
notice of the situation of labor accidents in 2017. [Online]. [15] ——, “Yolov3: An incremental improvement,” CoRR, vol. abs/1804.02767,
Available: http://vnniosh.vn/chitietN CKH/id/7559/T hong − bao − 2018. [Online]. Available: http://arxiv.org/abs/1804.02767
tinh − hinh − tai − nan − lao − dong − nam − 2017 [16] R. B. Girshick, “Fast R-CNN,” CoRR, vol. abs/1504.08083, 2015. [Online].
[2] B. E. Mneymneh, M. Abbas, and H. Khoury, “Automated hardhat Available: http://arxiv.org/abs/1504.08083
[17] F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding
detection for construction safety applications,” Procedia Engineering, vol.
196, pp. 895 – 902, 2017, creative Construction Conference 2017, for face recognition and clustering,” CoRR, vol. abs/1503.03832, 2015.
CCC 2017, 19-22 June 2017, Primosten, Croatia. [Online]. Available: [Online]. Available: http://arxiv.org/abs/1503.03832
http://www.sciencedirect.com/science/article/pii/S1877705817331430 [18] AlexeyAB. (2016) GitHub,Yolo-mark gui for marking bounded
[3] ——, “Vision-based framework for intelligent monitoring of hardhat wearing boxes of objects in images for training yolo. [Online]. Available:
on construction sites,” Journal of Computing in Civil Engineering, vol. 32, https://github.com/AlexeyAB/Yolom ark
2019.
[4] Z. Zhenhua, P. Man-Woo, and E. Nehad, “Automated monitoring of hard-
hats wearing for onsite safety enhancement,” in International Construction
Specialty Conference of the Canadian Society for Civil Engineering (ICSC),
2015.

View publication stats