ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
Available online at: http://ajet.
Andalas Journal of Electrical and Electronic Engineering Technology ISSN 2777-0079 Design and Implementation of a Deep Learning-Based Safety Helmet Compliance Detection System Using the Faster R-CNN Method Palman.
Sandy Azizi.
Muhammad Ilhamdi Rusydi*.
Rahmadi Kurnia Department of Electrical Engineering.
Faculty of Engineering.
Universitas Andalas.
Padang, 25163.
Indonesia
ARTICLE INFORMATION
A B S T R A C T
Received: April 2, 2026 Revised: May 2, 2026 Accepted: May 5, 2026 Available online: June 1, 2026 Workplace accidents remain one of the major issues in industrial environments and are often caused by low compliance with the use of Personal Protective Equipment (PPE), particularly safety helmets.
Manual supervision of PPE usage tends to be inefficient and prone to human This study aims to develop an intelligent computer-vision-based system capable of automatically and real-time monitoring helmet compliance.
The proposed system employs the Faster Region-Convolutional Neural Network (Faster R-CNN) algorithm to detect and classify workers who are wearing and not wearing helmets.
The dataset was obtained from CCTV video recordings in industrial areas, which were converted into image frames for training and testing The experimental results show that the system achieved an accuracy of 90% for helmet-wearing workers and 87% for non-helmet-wearing workers during daytime conditions, and 97% and 91% respectively at night.
With an average computation time of 0.
1 seconds per frame, the system is capable of real-time detection at up to 10 frames per second.
These results indicate that the Faster R-CNN method is effective in detecting PPE compliance and has the potential to be implemented as an automated safety-support system in industrial environments.
KEYWORDS
Faster R-CNN.
Computer Vision.
Personal Protective Equipment
CORRESPONDENCE
E-mail: rusydi@eng.
INTRODUCTION
According to global data released by the International Labour Organization (ILO), the number of occupational accidents (OA) and occupational diseases (OD) worldwide has reached 430 million cases per year, consisting of 270 million .
8%) OA
cases and 160 million .
2%) OD cases, resulting in 78 million worker fatalities annually.
Indonesia, the number of workers experiencing OA/OD increased each year from 2019 to 2021 .
Throughout 2022.
PT Semen Padang recorded seven occupational accident cases, comprising four first-aid cases and three severe cases.
Therefore, efforts to prevent workplace accidents must continue to be strengthened and enhanced.
As part of PT Semen PadangAos efforts to prevent workplace accidents and minimize associated risks, the company has implemented an Occupational Safety and Health Management System (SMK.
integrated into the Semen Padang Management System (SMSP) .
One of the key measures for controlling workplace accidents is ensuring the use of personal protective equipment (PPE) in operational areas.
However, compliance with PPE requirementsAiparticularly the use of safety helmetsAiis often neglected by some workers, thereby increasing the risk of workplace accidents .
, .
, .
, .
Previous studies have explored the use of computer visionAebased detection technologies to identify PPE usage .
, .
, .
, .
Convolutional Neural Networks (CNN.
are among the most https://doi.
org/10.
25077/ajet.
commonly applied methods for detecting objects in digital images .
, .
, .
, .
The Faster R-CNN model has demonstrated strong performance in object detection, particularly in identifying region-based features in images, such as worker faces or safety helmets .
, .
, .
, .
Previous research has primarily focused on demonstrating algorithm effectiveness in general or controlled environments, with limited application in complex industrial contexts.
Few studies have systematically designed and implemented real-time systems for detecting personal protective equipment (PPE) usage, while simultaneously addressing the accuracy and processing speed requirements critical for industrial applications.
This study proposes a novel approach to monitoring safety helmet compliance in industrial work environments by leveraging an optimized Faster R-CNN algorithm to detect helmet usage among workers at PT Semen Padang Packing Plant (PP) Bengkulu.
The system is expected to achieve high detection accuracy and provide automated notifications to management when PPE violations occur, thereby contributing to a reduction in workplace METHODS The steps carried out in this research follow the flowchart shown in Figure 1.
In the problem identification stage, information is collected from the field regarding obstacles in monitoring compliance with the use of personal protective equipment (PPE) PALMAN /ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
in the work area.
A literature review is also conducted using journal articles and previously published studies.
In the system design stage, a helmet detection system is developed for field work areas, capable of classifying workers who are wearing helmets and those who are not, as well as providing notifications when non-compliance is detected.
2 meters.
The cameras used were high-resolution devices with a frame size of 1920 y 1080 pixels and an average frame rate of 24 frames per second.
Figure 3.
Front-view camera installation position.
Figure 1.
Flowchart of research steps The next stage is system testing.
The designed system is evaluated to assess its performance.
The testing process involves observing the model during execution and measuring its performance based on the applied method.
All data obtained during the testing phase are recorded.
Finally, the results and analysis stage is conducted to examine the overall system performance.
Improvements are then implemented to enhance detection accuracy and speed.
Model refinement is achieved by adjusting the input size, increasing the amount of training data, and modifying the Faster R-CNN architecture.
System Design The system to be developed will utilize CCTV cameras as input for the machine learning/deep learning model .
, .
, .
The overall system design to be implemented is illustrated in Figure 2.
Figure 2.
System block diagram Objects such as helmets and heads without helmets are detected using a deep learning model.
Data Collection and Annotation Data collection was carried out by recording employees in several rooms and field areas at PT Semen Padang PP Bengkulu.
The recordings were captured as videos at 24 frames per second.
Each recorded video was sampled every 10 seconds .
, .
Video collection was conducted over several days, covering different locations and varying numbers of employees.
The CCTV camera installation layout is shown in Figure 3 and 4.
The CCTV cameras were installed with a spacing of 4.
5 meters between each unit.
The distance between each camera and the outermost pole was https://doi.
org/10.
25077/ajet.
Figure 4.
Side-view camera installation position Data Preprocessing The preprocessing step applied in this study was the resizing process .
The original input image size of 1920y1080 pixels was converted to 640y640 pixels.
Using the original 1920y1080 resolution requires substantial computational resources, making the training process difficult to perform efficiently .
The 640y640 resolution was selected to facilitate the training process in terms of both speed and required computing resources .
, .
, .
The collected dataset was divided into training, validation, and testing subsets .
The training data were used to train the Faster R-CNN model for object detection.
The validation data were employed to evaluate the performance of the trained model and to fine-tune its parameters.
The testing data were used to assess the modelAos performance on real-time data, enabling the derivation of conclusive insights regarding the trained model.
The dataset was split into 80% for training, 10% for validation, and 10% for testing .
, .
, .
, .
Before the data are used to train the artificial intelligence model, a labeling process is required.
The assigned labels consist of head and helmet classes.
The labeling was performed using LabelImg.
Each object was annotated by drawing a bounding box around it and assigning the appropriate classification label for the detected The recording results are presented in Figure 5, while the labeling process using LabelImg is shown in Figure 6.
PALMAN /ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
Table 1.
Comparison of mAP & the processing speed of each Model mAP (%) Processing Speed .
R-CNN Fast R-CNN Faster R-CNN 73.
Figure 5.
Recording using a camera Detection was performed using the Faster R-CNN algorithm because it achieves a processing speed of 0.
2 seconds on the VOC 2007 dataset, enabling Faster R-CNN to reach equal to or more than 5 frames per second .
, .
The Faster R-CNN algorithm learns each label corresponding to the classification of the objects to be detected .
The target objects in this study are heads with helmets and heads without helmets.
The flowchart for training the Faster R-CNN algorithm is presented in Figure 7.
System Training Figure 6.
Labeling and classification using LabelImg Model Tarining Helmet usage detection in this study is carried out using the Faster R-CNN algorithm.
Faster R-CNN generates outputs consisting of object classification results and the bounding box coordinates of each detected object .
, .
This algorithm is chosen because, among various Convolutional Neural NetworkAebased detection methods.
Faster R-CNN achieves inference times of less than one second .
, .
which is essential for real-time implementation.
Furthermore.
Faster R-CNN offers a relatively high mean average precision .
AP) compared to other models .
A comparison of mAP and processing speed for several algorithms is presented in Table 1.
The performance of the proposed system was evaluated using standard classification metrics, including accuracy, precision, and recall, to assess its capability in distinguishing between helmet and no-helmet usage.
Accuracy represents the proportion of correctly classified instances, precision reflects the reliability of positive predictions, and recall measures the modelAos ability to detect relevant objects.
The evaluation was conducted under two distinct conditions: daytime and nighttime.
A confusion matrix was employed to systematically compare the predicted labelsAi helmet, no helmet, and no humanAiwith the corresponding ground truth annotations.
Subsequently, the evaluation metrics were computed based on the confusion matrix using the following formulations .
ycNycE ycNycA yaycaycaycycycaycayc = ycNycE ycNycA yaycE yaycA ycNycE ycEycyceycaycnycycnycuycu = ycNycE yaycE ycNycE ycIyceycaycaycoyco = ycNycE yaycA Figure 7.
Flowchart of the faster R-CNN algorithm for training the model https://doi.
org/10.
25077/ajet.
PALMAN /ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
The calculation formulas for FN.
FP.
TN, and TP for the helmet data follow the equations below:
ycNycE = ycayceycoyco .
caycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoycey.
yaycA = ycayceycoyco .
caycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoycey.
caycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycayc.
yaycE = ycayceycoyco .
caycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoycey.
caycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoycey.
ycNycA = ycayceycoyco .
caycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoycey.
ycaycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycaycu ycaycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoyceyc ycaycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycayc.
available at PT.
Semen Padang.
The model development was implemented using the Python programming language, with TensorFlow serving as the primary framework for computational modeling and deep learning tasks.
Results This study employs the Faster R-CNN model to classify heads as either wearing helmets or not wearing helmets.
After undergoing a training process lasting approximately two hours, with a total of 4,100 iterations, the total loss and classification loss curves of the model were obtained, as illustrated in Figure 8 and Figure 9.
Formula for calculating FN.
FP.
TN, and TP for no-helmet data:
ycNycE = ycaycoyceyco .
caycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoycey.
yaycA = ycayceycoyco .
caycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoycey.
caycaycycycayco ycuycu Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycayc.
yaycE = ycayceycoyco .
caycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoycey.
caycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eayceycoycoycey.
ycNycA = ycayceycoyco .
caycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoycey.
ycaycaycycycayco Eayceycoycoyceyc ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycaycu ycaycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc Eayceycoycoyceyc ycaycaycycycayco ycuycu Eaycycoycaycu ycaycuycc ycyycyceyccycnycaycyceycc ycuycu Eaycycoycayc.
In addition to the previously described metrics, this research also evaluates performance using Average Precision (AP) and mean Average Precision .
AP).
Due to the absence of confidence scores required to construct a complete precisionAerecall curve.
AP is approximated using a single-point estimation based on precision and recall, where AP is computed as the product of precision and recall.
This approach serves as a simplified representation of the area under the precisionAerecall curve.
Furthermore, mAP is calculated as the mean of the AP values across all evaluated classes.
The formulations used to compute AP and mAP are presented as follows.
yaycE OO ycEycyceycaycnycycnycuycu ycu ycIyceycaycaycoyco ycoyaycE = ycA OcycA ycn=1 yaycEycn Figure 8.
Graph of total loss during training .
The evaluation was conducted by testing the model using heterogeneous datasets, including daytime and nighttime conditions, as well as scenarios involving one, two, and three Another critical parameter assessed was the detection An effective system is expected to perform detections promptly, thereby minimizing the number of skipped frames in the video sequences.
Figure 9.
Graph of classification loss during training At the final iteration, the training process indicates that the model achieved a total loss of 0.
013% and a bounding box classification loss of 0.
These loss values are considerably low compared to those observed in earlier iterations, suggesting that the training process can be terminated.
Furthermore, the total loss curve demonstrates a consistent downward trend, indicating that the learning process is stable and efficient.
Daytime and Nighttime Testing RESULTS AND DISCUSSIONS System Prototype The prototype system was developed through a series of stages, including data collection, data annotation, model training, integration of CCTV with the trained model, and the generation of alerts upon detection of violations.
The equipment utilized in this study comprises CCTV systems, computers, and cloud-based resources, all of which are readily https://doi.
org/10.
25077/ajet.
An example of the detection test results under daytime conditions is illustrated in Figure 10.
The overall performance of the detection tests is summarized in the confusion matrix for daytime testing, as presented in Table 2.
PALMAN /ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
Table 3.
Confusion matrix for nighttime testing Predicted Helmet Helmet Human Helmet Actual Helmet Human The performance of the system was evaluated under both daytime and nighttime conditions using the confusion matrix method.
During daytime testing, the helmet class achieved an accuracy of 9%, whereas the non-helmet class reached an accuracy of In contrast, nighttime testing yielded higher accuracies, with the helmet class achieving 98.
3% and the non-helmet class The detailed accuracy, precision and recall results for each class are summarized in Table 4.
Figure 10.
Results of the daytime testing Table 2.
Confusion matrix for daytime testing Predicted Helmet Helmet Human Helmet Actual Helmet Human Subsequently, the results of the tests conducted under nighttime conditions are presented in Figure 11.
The corresponding detection performance is summarized in the confusion matrix for nighttime testing, as shown in Table 3.
Table 4.
Accuracy.
Precision and Recall per Class Class Accuracy Precision Recall Daytime - Helmet - No Helmet Nighttime - Helmet - No Helmet The accuracy of the helmet class is consistently higher than that of the no-helmet class under both daytime and nighttime This can be attributed to the more distinctive color and shape characteristics of helmets compared to bare heads.
The relatively stable geometric structure and stronger visual features of helmets enable the model to learn discriminative features more effectively than those of non-helmet instances.
Furthermore, the evaluation results indicate that the accuracy obtained under daytime conditions is lower than that achieved during nighttime testing.
A similar trend is observed in the mAP Based on the estimation approach, the mAP is calculated to be 75.
8% for daytime testing and 87.
4% for nighttime testing, as presented in Table 5.
Table 5.
AP and mAP .
Class Daytime - Helmet - No Helmet Nighttime - Helmet - No Helmet Figure 11.
Results of the nighttime testing https://doi.
org/10.
25077/ajet.
The higher accuracy and mAP values observed during nighttime testing indicate that the model performs better under nighttime conditions compared to daytime scenarios.
This difference can be attributed to environmental factors affecting image quality.
During daytime, the observed objects are occasionally affected by excessive sunlight exposure, resulting in glare within the CCTV cameraAos field of view.
In contrast, nighttime conditions benefit from more controlled illumination provided by artificial lighting.
PALMAN /ANDALAS JOURNAL OF ELECTRICAL AND ELECTRONIC ENGINEERING TECHNOLOGY - VOL.
6 NO.
MAY-2026
which is sufficiently stable to support clearer object observation by the CCTV system.
between accuracy and detection speed, making it more suitable for real-time industrial applications.
The precision values for both the helmet and no-helmet classes reach 100% under both daytime and nighttime testing conditions, indicating that the model consistently performs correct inter-class classification without producing false positive predictions.
However, the relatively lower recall values suggest that a number of objects remain undetected.
This indicates that the model still requires improvement, particularly in terms of object detection In addition, recent studies on PPE detection using deep learning, such as those based on YOLO architectures .
have demonstrated faster inference times due to their one-stage detection framework.
However, these methods often exhibit lower localization accuracy compared to two-stage detectors like Faster R-CNN.
The findings of this study are consistent with prior work .
which highlights that Faster R-CNN provides higher precision in object localization, particularly for structured objects such as safety helmets.
This explains the consistently high precision values .
%) observed in this study, indicating the modelAos robustness in minimizing false positive detections.
Detection Speed Testing The detection speed evaluation was carried out by measuring the time required by the system to detect an image.
Using computational resources equipped with an Nvidia T4 provided by Google ColaboratoryAifeaturing 12 GB of RAM, 15 GB of GPU RAM, and 112 GB of disk storageAithe system achieved a detection speed of 0.
10 seconds.
This detection time is sufficiently fast and meets the requirements for real-time A comparison of the detection performance from this study with previous related studies is presented in Table 6.
Table 6.
Comparison of Detection Results Accuracy Detection Model Time .
Helmet No Helmet Xu Li et al.
Laily et al.
Our model
DISCUSSIONS