• ETRI Journal
• /
• v.44 no.3
• /
• pp.426-437
• /
• 2022

To understand the multilateral characteristics of human behavior and physiological markers related to physical, emotional, and environmental states, extensive lifelog data collection in a real-world environment is essential. Here, we propose a data collection method using multimodal mobile sensing and present a long-term dataset from 22 subjects and 616 days of experimental sessions. The dataset contains over 10 000 hours of data, including physiological, data such as photoplethysmography, electrodermal activity, and skin temperature in addition to the multivariate behavioral data. Furthermore, it consists of 10 372 user labels with emotional states and 590 days of sleep quality data. To demonstrate feasibility, human activity recognition was applied on the sensor data using a convolutional neural network-based deep learning model with 92.78% recognition accuracy. From the activity recognition result, we extracted the daily behavior pattern and discovered five representative models by applying spectral clustering. This demonstrates that the dataset contributed toward understanding human behavior using multimodal data accumulated throughout daily lives under natural conditions.

• The Journal of Korea Robotics Society
• /
• v.19 no.2
• /
• pp.188-195
• /
• 2024

This paper proposes a photorealistic real-time dense 3D mapping system that utilizes a neural network-based image enhancement method and mesh-based map representation. Due to the characteristics of the underwater environment, where problems such as hazing and low contrast occur, it is hard to apply conventional simultaneous localization and mapping (SLAM) methods. At the same time, the behavior of Autonomous Underwater Vehicle (AUV) is computationally constrained. In this paper, we utilize a neural network-based image enhancement method to improve pose estimation and mapping quality and apply a sliding window-based mesh expansion method to enable lightweight, fast, and photorealistic mapping. To validate our results, we utilize real-world and indoor synthetic datasets. We performed qualitative validation with the real-world dataset and quantitative validation by modeling images from the indoor synthetic dataset as underwater scenes.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2017.04a
• /
• pp.787-790
• /
• 2017

Data quality is the main issue in the classification problems; generally, the presence of noisy instances in the training dataset will not lead to robust classification performance. Such instances may cause the generated decision tree to suffer from over-fitting and its accuracy may decrease. Decision trees are useful, efficient, and commonly used for solving various real world classification problems in data mining. In this paper, we introduce a preprocessing technique to improve the classification accuracy rates of the C4.5 decision tree algorithm. In the proposed preprocessing method, we applied the naive Bayes classifier to remove the noisy instances from the training dataset. We applied our proposed method to a real e-commerce sales dataset to test the performance of the proposed algorithm against the existing C4.5 decision tree classifier. As the experimental results, the proposed method improved the classification accuracy by 8.5% and 14.32% using training dataset and 10-fold crossvalidation, respectively.

• ETRI Journal
• /
• v.45 no.6
• /
• pp.1079-1089
• /
• 2023

Stroke is one of the leading causes of long-term disability worldwide, placing huge burdens on individuals and society. Further, automatic human activity recognition is a challenging task that is vital to the future of healthcare and physical therapy. Using a baseline long short-term memory recurrent neural network, this study provides a novel dataset of stretching, upward stretching, flinging motions, hand-to-mouth movements, swiping gestures, and pouring motions for improved model training and testing of stroke-affected patients. A MATLAB application is used to output textual and audible prediction results. A wearable sensor with a triaxial accelerometer is used to collect preprocessed real-time data. The model is trained with features extracted from the actual patient to recognize new actions, and the recognition accuracy provided by multiple datasets is compared based on the same baseline model. When training and testing using the new dataset, the baseline model shows recognition accuracy that is 11% higher than the Activity Daily Living dataset, 22% higher than the Activity Recognition Single Chest-Mounted Accelerometer dataset, and 10% higher than another real-world dataset.

• ETRI Journal
• /
• v.44 no.3
• /
• pp.476-490
• /
• 2022

With the recent advances in technology, automatic speech recognition (ASR) has been widely used in real-world applications. The efficiency of converting large amounts of speech into text accurately with limited resources has become more vital than ever. In this study, we propose a method to rapidly recognize a large speech database via a transformer-based end-to-end model. Transformers have improved the state-of-the-art performance in many fields. However, they are not easy to use for long sequences. In this study, various techniques to accelerate the recognition of real-world speeches are proposed and tested, including decoding via multiple-utterance-batched beam search, detecting end of speech based on a connectionist temporal classification (CTC), restricting the CTC-prefix score, and splitting long speeches into short segments. Experiments are conducted with the Librispeech dataset and the real-world Korean ASR tasks to verify the proposed methods. From the experiments, the proposed system can convert 8 h of speeches spoken at real-world meetings into text in less than 3 min with a 10.73% character error rate, which is 27.1% relatively lower than that of conventional systems.

• Korean Journal of Remote Sensing
• /
• v.39 no.4
• /
• pp.395-407
• /
• 2023

Super-resolution (SR) has great significance in image processing because it enables downstream vision tasks with high spatial resolution. Recently, SR studies have adopted deep learning networks and achieved remarkable SR performance compared to conventional example-based methods. Deep-learning-based SR models generally require low-resolution (LR) images and the corresponding high-resolution (HR) images as training dataset. Due to the difficulties in obtaining real-world LR-HR datasets, most SR models have used only HR images and generated LR images with predefined degradation such as bicubic downsampling. However, SR models trained on simple image degradation do not reflect the properties of the images and often result in deteriorated SR qualities when applied to real-world images. In this study, we propose an image degradation model for HR satellite images based on the modulation transfer function (MTF) of an imaging sensor. Because the proposed method determines the image degradation based on the sensor properties, it is more suitable for training SR models on remote sensing images. Experimental results on HR satellite image datasets demonstrated the effectiveness of applying MTF-based filters to construct a more realistic LR-HR training dataset.

• Journal of Biomedical Engineering Research
• /
• v.44 no.6
• /
• pp.384-391
• /
• 2023

Among the elderly, fatal injuries and deaths are significantly attributed to falls. Therefore, a pre-impact fall detection system is necessary for injury prevention. In this study, a robust threshold-based algorithm was proposed for pre-impact fall detection, reducing false positives in highly dynamic daily-living movements. The algorithm was validated using public datasets (KFall and FARSEEING) that include the real-world elderly fall. A 6-axis IMU sensor (Movella Dot, Movella, Netherlands) was attached to S2 of 20 healthy adults (aged 22.0±1.9years, height 164.9±5.9cm, weight 61.4±17.1kg) to measure 14 activities of daily living and 11 fall movements at a sampling frequency of 60Hz. A 5Hz low-pass filter was applied to the IMU data to remove high-frequency noise. Sum vector magnitude of acceleration and angular velocity, roll, pitch, and vertical velocity were extracted as feature vector. The proposed algorithm showed an accuracy 98.3%, a sensitivity 100%, a specificity 97.0%, and an average lead-time 311±99ms with our experimental data. When evaluated using the KFall public dataset, an accuracy in adult data improved to 99.5% compared to recent studies, and for the elderly data, a specificity of 100% was achieved. When evaluated using FARSEEING real-world elderly fall data without separate segmentation, it showed a sensitivity of 71.4% (5/7).

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.11a
• /
• pp.124-125
• /
• 2023

Defect detection during liquid crystal display (LCD) manufacturing has always been a critical challenge. This study aims to address this issue by proposing a data augmentation method based on generative adversarial networks (GAN) to improve defect identification accuracy in LCD production. By leveraging synthetically generated image data from GAN, we effectively augment the original dataset to make it more representative and diverse. This data augmentation strategy enhances the model's generalization capability and robustness on real-world data. Compared to traditional data augmentation techniques, the synthetic data from GAN are more realistic, diverse and broadly distributed. Experimental results demonstrate that training models with GAN-generated data combined with the original dataset significantly improves the detection accuracy of critical defects in LCD manufacturing, compared to using the original dataset alone. This study provides an effective data augmentation approach for intelligent quality control in LCD production.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.8
• /
• pp.4021-4037
• /
• 2018

In network intrusion detection research, two characteristics are generally considered vital to building efficient intrusion detection systems (IDSs): an optimal feature selection technique and robust classification schemes. However, the emergence of sophisticated network attacks and the advent of big data concepts in intrusion detection domains require two more significant aspects to be addressed: employing an appropriate big data computing framework and utilizing a contemporary dataset to deal with ongoing advancements. As such, we present a comprehensive approach to building an efficient IDS with the aim of strengthening academic anomaly detection research in real-world operational environments. The proposed system has the following four characteristics: (i) it performs optimal feature selection using information gain and branch-and-bound algorithms; (ii) it employs machine learning techniques for classification, namely, Logistic Regression, Naïve Bayes, and Random Forest; (iii) it introduces bulk synchronous parallel processing to handle the computational requirements of large-scale networks; and (iv) it utilizes a real-time contemporary dataset generated by the Information Security Centre of Excellence at the University of Brunswick (ISCX-UNB) to validate its efficacy. Experimental analysis shows the effectiveness of the proposed framework, which is able to achieve high accuracy, low computational cost, and reduced false alarms.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.18 no.2
• /
• pp.51-58
• /
• 2023

This paper proposes a deep learning method for estimating the heart rate from facial videos. Our proposed method estimates remote photoplethysmography (rPPG) signals to predict the heart rate. Although there have been proposed several methods for estimating rPPG signals, most previous methods can not be utilized in low-power single board computers due to their computational complexity. To address this problem, we construct a lightweight student model and employ a knowledge distillation technique to reduce the performance degradation of a deeper network model. The teacher model consists of 795k parameters, whereas the student model only contains 24k parameters, and therefore, the inference time was reduced with the factor of 10. By distilling the knowledge of the intermediate feature maps of the teacher model, we improved the accuracy of the student model for estimating the heart rate. Experiments were conducted on the UBFC-rPPG dataset to demonstrate the effectiveness of the proposed method. Moreover, we collected our own dataset to verify the accuracy and processing time of the proposed method on a real-world dataset. Experimental results on a NVIDIA Jetson Nano board demonstrate that our proposed method can infer the heart rate in real time with the mean absolute error of 2.5183 bpm.