Matrix : Jurnal Manajemen Teknologi dan Informatika

Development of an AI IoT-based smart system to support improved beach visitor safety

Fri, 31 Jul 2026 00:00:00 +0000

The high rate of sea accidents at Indonesian coastal tourist destinations, such as those occurring at Parangtritis and Pangandaran beaches, underscores the limitations of conventional safety systems that rely on manual surveillance. The objective of this research is to develop the Smart Beach Visitor Safety System (Smart BEVISAT), an integrated system based on artificial intelligence (AI) and the Internet of Things (IoT) to enhance beachgoer safety through preventive and reactive functions. The system development method employs the waterfall model, which includes stages of requirements analysis, design, implementation, and testing. The system is composed of three primary components: an AI surveillance camera, an IoT safety buoy, and a ground station. The AI camera utilizes advanced algorithms such as YOLOv11 for object detection, ByteTrack for tracking, and DeepLabv3+ for waterline segmentation, enabling it to monitor visitors and detect safety zone violations in real-time. In the event of an incident, the system automatically sends the victim's coordinates to the IoT Safety Buoy, an Unmanned Surface Vehicle (USV) designed to autonomously navigate to the victim's location for rapid evacuation. The implementation results show a functional prototype with measurable technical specifications, where the buoy achieves a total gross buoyancy of 205.8 kg, a net carrying capacity of approximately 190 kg, and a measured operational speed of 1.20–1.53 m/s under varying load conditions, demonstrating effective rescue capability. The development of Smart BEVISAT is expected to provide a technological solution to minimize accident risks and strengthen the maritime tourism sector in Indonesia

Video-based human activity recognition analysis using YOLOv26

Santi Rahayu, Imam Riadi, Andri Pranolo — Fri, 31 Jul 2026 00:00:00 +0000

Human Activity Recognition (HAR) is a rapidly growing research field in computer vision and has various applications, such as intelligent surveillance systems, sports analysis, healthcare, and human-computer interaction. This study aims to analyze the performance of the YOLOv26 Classification model in recognizing video-based human activities using the UCF101 dataset. This study uses six activity classes, namely JumpingJack, Punch, PushUps, Typing, WalkingWithDog, and WritingOnBoard. The research method is carried out through video frame extraction using a uniform sampling technique, the formation of an image classification dataset, YOLOv26 model training, and evaluation at the frame and video levels. Experiments were conducted using 36 configuration combinations consisting of four YOLOv26 model variants (YOLOv26-n, YOLOv26-s, YOLOv26-m, and YOLOv26-l), three variations in the number of frames (8, 16, and 24 frames), and three variations in the number of epochs (50, 100, and 150 epochs). Video evaluation was conducted using a majority voting approach with accuracy, precision, recall, and F1-score metrics. The results showed that all configurations produced video accuracy above 93%. The best configuration was obtained with the YOLOv26-m model with 16 frames and 50 epochs, which achieved video accuracy of 98.26%, precision of 98.15%, recall of 98.37%, and F1-score of 98.23%. Confusion matrix analysis showed that most predictions were on the main diagonal, indicating the model's ability to distinguish human activities with a low error rate. These results prove that YOLOv26 Classification has excellent performance for video-based human activity recognition and has the potential to be applied to various applications based on automatic human activity analysis.

Daily USD-IDR exchange rate prediction using Long Short-Term Memory (LSTM) with macroeconomic indicators and recursive multi-step prediction

Ni Luh Gede Ina Ari Richardi, Putu Indah Ciptayani, I Putu Bagus Arya Pradnyana — Fri, 31 Jul 2026 00:00:00 +0000

The foreign exchange market’s high volatility and non-linear dynamics pose significant challenges for accurate currency price forecasting. This study develops a predictive model for the daily USD-IDR exchange rate using Long Short-Term Memory (LSTM) method integrated with macroeconomic indicators. Historical price data, global oil prices, interest rates, and the US Dollar Index (DXY) were collected by fetching from Yahoo Finance and FRED. A comprehensive experimental design comprising 432 configurations was evaluated across varying data periods, train-validation-test splits, feature combinations, and hyperparameters. Results indicate that a 10-year historical dataset yields the most stable and accurate performance, achieving a test MAPE of 0.52% and R² of 0.904. The integration of macroeconomic variables improved predictive capability, though the impact of DXY was found to be conditional on data split ratios and hyperparameter settings. The optimal model was deployed as an interactive web application using Streamlit, enabling users to forecast USD-IDR rates up to seven days ahead via recursive multi-step forecasting. The proposed system provides a practical, accessible tool for retail traders and financial analysts to support informed decision-making in volatile forex markets

Image classification of rerajahan ulap-ulap Bali using MobileNetV2 architecture

Fri, 31 Jul 2026 00:00:00 +0000

Rerajahan Ulap-Ulap is a visual expression in Balinese Hindu culture that functions as a sacred element in the mlaspas ritual of buildings, containing sacred scripts and specific ornaments. The visual complexity of these motifs presents a challenge in accurately recognizing their types, necessitating a technological approach to support the documentation of this cultural heritage. This study aims to develop an image classification model using the MobileNetV2 architecture with a transfer learning method. The research utilized a primary dataset of 810 original images across nine motif classes, expanded through augmentation to 3,888 images (432 per class) to address data limitations and prevent overfitting. Model performance was evaluated using a 5-Fold Cross Validation method across three optimization scenarios: AdamW, Nadam, and Adagrad. Experimental results demonstrate that Scenario 1 (AdamW) achieved superior performance with an average accuracy of 98.64%, followed by Nadam (98.52%) and Adagrad (71.23%). These findings indicate that the combination of MobileNetV2 and AdamW optimization provides a reliable solution for automated classification of rerajahan, serving as a digital instrument to support Balinese cultural preservation efforts.

Multi-class skin disease classification using transfer learning architectures

Muhammad Akhdaan, Majid Rahardi — Fri, 31 Jul 2026 00:00:00 +0000

Dermatological conditions are among the most common health problems worldwide, where delayed identification may increase disease severity and complicate treatment procedures. However, restricted access to dermatological expertise and insufficient public awareness often contribute to delayed diagnosis. This study proposes a multi-class skin disease classification approach using deep learning and transfer learning architectures based on digital skin images. The dataset, obtained from the babaruzair/kaggle-skin-disease repository, consists of 1,157 images categorized into eight skin disease classes. Image preprocessing techniques, including resizing, normalization, and data augmentation, were applied to improve data quality and model generalization. Three models were evaluated in this study, namely a baseline Convolutional Neural Network (CNN), MobileNetV2, and EfficientNetB0. Both transfer learning models utilized ImageNet pre-trained weights, followed by customized classification layers and fine-tuning of selected upper layers. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. Experimental results show that the baseline CNN achieved an accuracy of 52.36%, while MobileNetV2 and EfficientNetB0 achieved accuracies of 88.84% and 95.28%, respectively. The findings demonstrate that transfer learning architectures significantly outperform conventional CNN models for limited medical image datasets, with EfficientNetB0 providing the best classification performance. These results indicate the potential of deep learning-based approaches to support early skin disease diagnosis and assist clinical decision-making