Trajectory prediction for humans is essential for autonomously moving agents such as robots. However, prior research has mainly been developed using top-down datasets, whereas real-world drones, mobile robots, and humanoids typically mount cameras facing forward or slightly downward rather than toward the floor. As a result, a viewpoint mismatch makes it difficult to directly apply existing trajectory prediction models. Therefore, aiming at a real-time system operating on embedded hardware while accounting for such camera viewpoints, we designed a ROS2-based integrated system that predicts human trajectories from camera inputs. ROS2 is a robot software framework, and we adopted it because its modular architecture facilitates building real-time pipelines. The proposed system was designed for the NVIDIA Jetson Orin Nano, one of NVIDIA’s embedded computing boards. The system detects people from RGB images captured by an Intel RealSense D455 camera using YOLOv8, maintains consistent object IDs across frames via ByteTrack, and accumulates the observed trajectories over time. These observed trajectories are then preprocessed and used as inputs to a pretrained trajectory prediction model, which predicts the future motion trajectories of humans. We validated the proposed system in real time in a corridor environment and confirmed that the integrated processing stream—from camera input to detection, tracking, and prediction—runs smoothly on the Jetson Orin Nano, and that the fine-tuned model generates future trajectories that align with human movement patterns.

Emotion recognition in conversations is a key task in natural language processing that enhances the quality of human–computer interactions. Although existing deep learning and Transformer-based pretrained language models have shown remarkably enhanced performances, both approaches have inherent limitations. Deep learning models often fail to capture the global semantic context, whereas Transformer-based pretrained language models can overlook subtle, local emotional cues. To overcome these challenges, we developed EmoBERTa–CNN, a hybrid framework that combines EmoBERTa’s ability to capture global semantics with the capability of convolutional neural networks (CNNs) to extract local emotional features. Experiments on the SemEval-2019 Task 3 and Multimodal EmotionLines Dataset (MELD) demonstrated that the proposed EmoBERTa–CNN model achieved F1-scores of 96.0% and 79.45%, respectively, significantly outperforming existing methods and confirming its effectiveness for emotion recognition in conversations.

This poster presents the Smart_Safe system, a modular platform for real-time safety management in indoor industrial environments. The system integrates wearable sensors, Auto-ID technologies (such as RFID and Bluetooth), and AI-based analytics to detect, evaluate, and prevent occupational safety risks. Its core functionality includes real-time tracking of workers, detection of critical events (such as falls or zone violations), and prevention of collisions between people and mobile robots or forklifts. The system is designed to be scalable, interoperable with existing infrastructure, and privacy-respecting through the use of anonymized tracking and local processing. Integration with edge computing and digital twins enables context-aware decision-making and dynamic response to incidents. Smart_Safe supports applications in warehouses, smart factories, and production halls with a focus on high-risk or high-traffic areas. Initial testing demonstrates the feasibility of using hybrid sensor networks and lightweight AI models to ensure workplace safety and optimize movement flows. The poster also outlines the international collaboration between Czech and Korean partners, highlighting the hardware-software co-design process and the future roadmap for deployment.

This poster presents a novel system for inventory automation in large outdoor warehouses using a lightweight drone-mounted UHF RFID reader. The system leverages autonomous aerial platforms equipped with custom-developed RFID readers operating in the 865–868 MHz band, combined with intelligent software modules for tag readability evaluation. The proposed solution addresses major limitations of traditional stationary or handheld readers by enabling efficient scanning from a distance, including hard-to-reach areas. Two prototypes of RFID readers were developed, tested and optimized for use on commercial drones, and validated in real-world environments in both Czech and Korean warehouse facilities. The results confirm the reader’s ability to reliably detect tags at distances up to 28 meters and demonstrate its applicability for regular inventory processes. In addition to the hardware, two supporting methodologies and a software tool were developed for integration into existing information systems. The project was carried out within an international research consortium, where the Korean partners focused on AI-based image processing for drone navigation and risk avoidance. The presented solution contributes to the development of autonomous inventory systems, improves safety, and significantly reduces human effort and inventory time in logistics operations.

This paper addresses the problem that RGB cameras have a relatively low recognition rate for finding people in low-light environments, and aims to use a them1al imaging camera simultaneously to increase the recognition rate for finding people. For this purpose, a bounding box is created based on what is found in the thermal image using the YOL0v8 model, and the boundary line is extracted using the Canny Edge detection algorithm. Then, by calculating the displacement between the border of the high-intensity image and the border of the thermal image, the error of the RGB camera and the them1al imaging camera is shifted upward. Then, adjust the fineness of the boundary line based on the military value of the Canny Edge detection algorithm. As a result, compared to the initial low-light image, the final low-light image with the border selected had a human recognition rate of 22% on average.

Out-of-Domain (OOD) intent classification is an important task for a dialog system, as it allows for appropriate responses to be generated. Previous studies aiming to solve the OOD intent classification task have generally adopted metric learning methods to generate decision boundaries in the embedding space. However, these existing methods struggle to capture the high-dimensional semantic features of data, as they learn decision boundary using scalar distances. They also use generated OOD samples for learning. However, such OOD samples are biased, and they cannot include all real-world OOD intents, thus representing a limitation. In the current paper, we attempt to overcome these challenges by using Advanced Proxy-Anchor loss, which introduces a margin proxy and shared proxy. First, to generate a decision boundary that has the high-dimensional semantic features of training data, we use a margin proxy for learnable embedding vectors. Next, the shared proxy, which is shared by all In-Domain (IND) samples, is introduced to make it possible to learn the discriminative feature between IND intents and OOD intent, ultimately leading to the improved classification of OOD samples. We conduct evaluations of the proposed method using three benchmark datasets. The experimental results demonstrate that our method achieved an improved performance compared to the methods described in previous studies.

The binarization of degraded documents represents a crucial preprocessing task for various document analyses, including optical character recognition and historical document analysis. Various convolutional neural network models and generative models have been used for document binarization. However, these models often struggle to deliver generalized performance on noise types the model has not encountered during training and may have difficulty extracting intricate text strokes. We herein propose a novel approach to address these challenges by introducing the use of the latent diffusion model, a well-known high-quality image-generation model, into the realm of document binarization for the first time. By leveraging an iterative diffusion-denoising process within the latent space, our approach excels at producing high-quality, clean, binarized images and demonstrates excellent generalization using both data distribution and time steps during training. Furthermore, we enhance our model’s ability to preserve text strokes by incorporating a gated U-Net into the backbone network. The gated convolution mechanism allows the model to focus on the text region by combining gating values and features, facilitating the extraction of intricate text strokes. To maximize the effectiveness of our proposed model, we use a combination of the latent diffusion model loss and pixel-level loss, which aligns with the model’s structure. The experimental results on the Handwritten Document Image Binarization Contest and Document Image Binarization Contest benchmark datasets showcase the superior performance of our proposed model compared to existing methods.

With the recent development of drone technology, drones are being used in various fields. Drones have the advantage of being equipped with various devices to move freely and perform various tasks. In the field of inventory management, many studies have been conducted into management automation based on the drone. Drones scan a marker, such as a quick response code (QR code), attached to the shelves to obtain location information of the shelves on which the inventory is loaded. At the same time, drones perform inventory management by scanning the marker attached to the inventory to obtain inventory information. However, unlike indoor warehouses, where grids or shelves are well-defined, a storage yard is not fixed in the location where the inventory is stored. It is difficult to recognize the loading position from the marker for a storage yard without shelves and grids. Furthermore, the loading position of the inventory is not fixed. For the automation of inventory management of warehouses where shelves and grids are undefined, this paper proposes a framework that estimates the inventory 3D position in the video frame based on a deep learning model. The proposed framework uses the image segmentation model to detect and decode the marker in the video frame to estimate the 3D position of a drone and inventory. In addition, the estimated inventory 3D position is corrected using the continuity of the video frame. Experiment results on the video dataset verified that the proposed framework improved the 3D position estimation performance of the inventory. Consequently, efficient inventory management based on drones can be performed through the proposed framework for the 3D position estimation of inventory in all types of warehouses.

Type 2 diabetes (T2D) and non-alcoholic fatty liver disease (NAFLD) are worldwide chronic diseases that have strong relationships with one another and commonly exist together. Type 2 diabetes is considered one of the risk factors for NAFLD, so its occurrence in people with NAFLD is highly likely. As the high and increasing number of T2D and NAFLD, which potentially followed by existing together number, an analysis and assessment of T2D screening scores in people with NAFLD is necessary to be done. To prevent this potential case, an effective early prediction model is also required to be developed, which could help the patients avoid the dangers of both existing diseases. Therefore, in this study, analysis and assessment of T2D screening scores in people with NAFLD and the early prediction model utilizing a forward logistic regression-based feature selection method and multi-layer perceptrons are proposed. Our analysis and assessment results showed that the prevalence of T2D among patients with NAFLD was 8.13% (for prediabetes) and 37.19% (for diabetes) in two population-based NAFLD datasets. The variables related to clinical tests, such as alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), gamma-glutamyl transferase (GGT), and systolic blood pressure (SBP), were found to be statistically significant predictors (p-values < 0.001) that indicate a strong association with T2D among patients with NAFLD in both the prediabetes and diabetes NAFLD datasets. Finally, our proposed model showed the best performance in terms of all performance evaluation metrics compared to existing various machine learning models and also the models using variables recommended by WHO/CDC/ADA, with achieved accuracy as much as 92.11% and 83.05% and its improvement scores after feature selection of 1.35% and 5.35%, for the first and second dataset, respectively.

Psychological well-being is vitally important for the quality of life of the elderly and is only increasing in importance with the rapidly increasing elderly population worldwide. Emerging elderly problems include a deterioration in physical function, loss of friends or spouse, reduced social participation, and reduced economic ability. Hence, the importance of coping with and managing stress in the elderly is also rapidly increasing. This study proposed psychotherapy narration was designed to assist elderly mental well-being by combining person-centered therapy, positive psychology, and cognitive behavioral therapy. Extending from current mainstream psychotherapeutic methods, postmodern psychotherapeutic techniques based on various psychological theories or techniques have begun to be more widely applied. However, almost no previous studies have developed a systematic psychotherapy narration for the elderly. Therefore, this study developed a postmodern psychotherapeutic narration and confirmed its aspects by analyzing elderly satisfaction regarding the corresponding emotion. This satisfaction analysis study found the value of the psychotherapy-narrative model according to the elderly’s stressful situations and emotions. This study can be an initial model of postmodernist-psychotherapy narration for the elderly. Therefore, based on the model of this study, future-oriented development and research on the diversity of the elderly and the effects of each narration are important. The future of this study will give mental self-sustainability to clients who need psychotherapy.

Self-supervised learning is a method that learns general representation from unlabeled data. Masked image modeling (MIM), one of the generative self-supervised learning methods, has drawn attention for showing state-of-the-art performance on various downstream tasks, though it has shown poor linear separability resulting from the token-level approach. In this paper, we propose a contrastive learning-based multi-view masked autoencoder for MIM, thus exploiting an image-level approach by learning common features from two different augmented views. We strengthen the MIM by learning long-range global patterns from contrastive loss. Our framework adopts a simple encoder–decoder architecture, thus learning rich and general representations by following a simple process: (1) Two different views are generated from an input image with random masking and by contrastive loss, we can learn the semantic distance of the representations generated by an encoder. By applying a high mask ratio, of 80%, it works as strong augmentation and alleviates the representation collapse problem. (2) With reconstruction loss, the decoder learns to reconstruct an original image from the masked image. We assessed our framework through several experiments on benchmark datasets of image classification, object detection, and semantic segmentation. We achieved 84.3% in fine-tuning accuracy on ImageNet-1K classification and 76.7% in linear probing, thus exceeding previous studies and showing promising results on other downstream tasks. The experimental results demonstrate that our work can learn rich and general image representation by applying contrastive loss to masked image modeling.

Machine learning algorithms can be used to forecast future blood glucose (BG) levels for diabetes patients, according to recent studies. In this study, dataset from continuous glucose monitoring (CGM) system was used as the sole input for the machine learning models. To forecast blood glucose levels 15, 30, and 45 minutes in the future, we suggested deep neural network (DNN) and tested it on 7 patients with type 1 diabetes (T1D). The suggested prediction model was evaluated against a variety of machine learning models, such as k-nearest neighbor (KNN), support vector regression (SVR), decision tree (DT), adaptive boosting (AdaBoost), random forest (RF), and eXtreme gradient boosting (XGBoost). The experimental findings demonstrated that the proposed DNN model outperformed all other models, with average root mean square errors (RMSEs) of 17.295, 25.940, and 35.146 mg/dL over prediction horizons (PHs) of 15, 30, and 45 minutes, respectively. Additionally, we have included the suggested prediction model in web-based blood glucose level prediction tools. By using this web-based system, patients may readily acquire their future blood glucose levels, allowing for the generation of preventative alarms prior to crucial hypoglycemia or hyperglycemic situations.

Predicting future glycemic events such as hypoglycemia, hyperglycemia, and normal for type 1 diabetes (T1D) remains a significant and challenging issue. In this study, an artificial neural network (ANN)-based model is proposed to predict the future glycemic events of T1D patients. We utilized five T1D patient datasets to build the models and predict future glycemic events with a prediction horizon (PH) of 30 and 60 minutes ahead of time. We applied the data preprocessing method based on the sliding window approach by sliding the blood glucose time-series data from the past 60 minutes (the last 12 data points) as input and using the next 30 and 60 minutes (the next 6 and 12-th data points) as output. All the numeric blood glucose output data are then transformed into a multi-class classification label, such as hypoglycemia, hyperglycemia, and normal. Our proposed model is then used to learn and create the prediction model from the preprocessed blood glucose dataset. Four performance metrics such as accuracy, precision, recall, and f-1 score were utilized to measure the performance of the classification models used in this study, such as Naïve Bayes (NB), Decision Tree (DT), Support Vector Machine (SVM), and K-Nearest Neighbour (KNN). The results showed that our proposed ANN-based model performed better at predicting future glycemic events than other models, with an average accuracy, precision, recall, and f-1 score of 88.649%, 76.661%, 71.731%, 72.609%, and 83.364%, 60.437%, 61.345%, 60.62% for the PH of 30 and 60 minutes, respectively. As a result, knowing this future glycemic event sooner can help patients avoid potentially dangerous conditions and can eventually be used to improve diabetes management.

Developing a prediction model from risk factors can provide an efficient method to recognize breast cancer. Machine learning (ML) algorithms have been applied to increase the efficiency of diagnosis at the early stage. This paper studies a support vector machine (SVM) combined with an extremely randomized trees classifier (extra-trees) to provide a diagnosis of breast cancer at the early stage based on risk factors. The extra-trees classifier was used to remove irrelevant features, while SVM was utilized to diagnose the breast cancer status. A breast cancer dataset consisting of 116 subjects was utilized by machine learning models to predict breast cancer, while the stratified 10-fold cross-validation was employed for the model evaluation. Our proposed combined SVM and extra-trees model reached the highest accuracy up to 80.23%, which was significantly better than the other ML model. The experimental results demonstrated that by applying extra-trees-based feature selection, the average ML prediction accuracy was improved by up to 7.29% as contrasted to ML without the feature selection method. Our proposed model is expected to increase the efficiency of breast cancer diagnosis based on risk factors. In addition, we presented the proposed prediction model that could be employed for web-based breast cancer prediction. The proposed model is expected to improve diagnostic decision-support systems by predicting breast cancer disease accurately.

As a part of the supply chain, inventory management includes, among other things, maintaining the storage of stock, controlling the amount of product for sale and order fulfilment. In business terms, inventory management means the right stock, at the right levels, in the right place, at the right time. In the case of large outdoor warehouses, common identification methods are lengthy and inappropriate. One way to determine inventory easily and quickly is to deploy UAV’s (unmanned aerial vehicle) for product identification purposes. In this case, however, there is a problem in determining where the goods are located. A drone moves at higher altitudes, which can lead to a situation where we will not be able to determine the exact location of the goods. This article deals with a method of determining the correct flight level suitable to distinguish the identified items located at least 2 m apart. The evaluation is performed based on an RSSI (received signal strength indicator) value. The experiment proved that even at maximum reading distance of selected passive UHF RFID tags the two objects can be distinguished.

Predicting future blood glucose (BG) values for diabetic patients, particularly for type 1 diabetes (T1D), remains an important and challenging issue. To overcome it, several well-known machine learning models have been used in recent years. Thus, a personalized model based by using random forest (RF) regression is implemented to forecast the future BG level of T1D patients. To create the models and forecast the BG value, a clinical dataset of T1D patients is employed. In this study, the future BG value, or prediction horizon (PH), is used for the next 15 and 30 minutes. Several performance metrics, including coefficient of determination (R 2 ), mean absolute percentage error (MAPE), and root mean square error (RMSE), are measured during the experiments to calculate the prediction models' performances. The results revealed that the proposed BG prediction model outperformed other models including Multi-Layer Perceptron, Support Vector Regression, Decision Tree, K-Nearest Neighbour, with an average RMSE, MAPE, and R 2 of 15.54 mg/dL, 8.94%, 0.88, and 27.61 mg/dL, 15.52%, and 0.66 for PH of 15 and 30 minutes, respectively. In addition, the RMSE score was reduced by around 1.14 mg/dL and 1.32 mg/dL for the next 15 and 30 minutes, respectively, after adding statistical-extracted data as additional features for the regression models, compared to regression models without statistical data. Ultimately, it is expected that the results of the present study could be used to improve diabetes care.

Emotion Recognition (ER) is an essential research area of ​​natural language processing that can be applied to various fields. Texts in the fields of health care, marketing, and psychological counseling take various forms, and it is very important from a business point of view to find the emotions inherent in these texts. Recently, ER using text embeddings generated through a pre-trained language model with a large corpus was performed. However, since the embeddings are generalized to various domains, there is a limitation to directly using them for ER. In this study, to overcome the limitation, we propose a method that modifies generalized embeddings to emotional embeddings by performing proxy-based metric learning. In the proposed method, we fine-tuned the pre-trained language model by using proxy-anchor loss so that embeddings represent emotion appropriately. Previous studies only added linear classifiers. But, it is possible to capture emotional relationships between data by using proxy-based metric learning. In this study, we conducted ER experiments with benchmark datasets. The experimental result shows that the proposed method achieves better performance than the baseline and creates emotion-specific embeddings.

Named Entity Recognition (NER) is at the core of natural language understanding. The quality and amount of datasets determine the performance of deep-learning-based NER models. As datasets for NER require token-level or word-level labels to be assigned, annotating the datasets is expensive and time consuming. To alleviate efforts of manual anotation, many prior studies utilized weak supervision for NER tasks. However, using weak supervision directly would be an obstacle for training deep networks because the labels automatically annotated contain a a lot of noise. In this study, we propose a framework to better train the deep model for NER tasks using weakly labeled data. The proposed framework stems from the idea that mixup, which was recently considered as a data augmentation strategy, would be an obstacle to deep model training for NER tasks. Inspired by this idea, we used mixup as a perturbation function for consistency regularization, one of the semi-supervised learning strategies. To support our idea, we conducted several experiments for NER benchmarks. Experimental results proved that directly using mixup on NER tasks hinders deep model training while demonstrating that the proposed framework achieves improved performances compared to employing only a few human-annotated data.

This study examined the occurrence of emotion types and the contents and meanings of individual emotion types to improve the quality of life of South Korean senior patients in convalescent hospitals. This research is a sequential mixed study in which we conducted emotion frequency and content analyses with 20 elderly resident patients in a convalescent hospital. In the emotion frequency analysis, we performed emotion occurrence frequency analysis and clustering to create groups of subjects that showed similar distributions of emotions. The study results found that South Korean senior patients displayed six major emotions: joy, sorrow, anger, surprise, fear, and tranquility, including mixed emotional states. In the emotion content analysis, we used NVivo to categorize and analyze the interview contents based on emotion types. The study results show the characteristics of emotions according to patients' treatment and recovery, life within narrow boundaries, relationships with new people and family, and the appearances of themselves that they could not easily but must accept. In addition, these characteristics appeared in health, environment, relationships, and psychological structures. Ultimately, the study results suggest that improving the quality of life of South Korean senior patients requires understanding of their emotions and examining diverse emotions in multiple dimensions.

Heart disease (HD) is number one chronic disease and becomes a major cause of worldwide disability and death. Aside of HD, type 2 diabetes (T2D) is also as the most deathful diseases that causes serious issues if untreated and undetected. HD and T2D predictions are the most effective measures to control the HD and T2D. Thus, early HD and T2D predictions are important to help individuals in preventing the occurrence of the worst cases. This study proposes a chronic disease prediction model for HD and T2D prediction. The proposed study utilized random forest combined with DBSCAN as outlier detection method and SMOTE-ENN as data balancing method. Two HD datasets (Statlog and Cleveland) and one T2D dataset (NHIS Korea) were used for building the model and comparing the results with other existing machine learning (ML) algorithms, including GNB, LR, MLP, DT, and SVM. To measure the performance of the model, k-fold (10) cross-validation and several performance metrics including accuracy, precision, f-measure, and recall are applied in this study. The results show the model that we proposed outperforms other classification models, as well as previous studies, with accuracy rates 97.63%, 97.69%, and 94.85% for Statlog HD dataset, Cleveland HD dataset and NHIS T2D dataset, respectively. By utilizing the proposed model, it could increase the expectation in preventing the occurrence of the worst case and helping individuals in taking fast and precise actions when status of HD and T2D are detected.