This study presents a comparative analysis of two real-time tree localization and mapping approaches in forestry environments: an eccentricity-based LiDAR mapping method (EB-LiDAR) and a depth-assisted image-based mapping method (DAIM). EB-LiDAR extracts tree trunk clusters from LiDAR data using DBSCAN clustering and PCA-based eccentricity filtering whereas the DAIM approach integrates YOLO-based visual trunk detection with depth clustering, estimating tree positions through arc-based geometric fitting. Both methods are deployed on a multi-sensor mobile robot platform where LiDAR and IMU data are fused via an Extended Kalman Filter (EKF) to ensure robust odometry. Tree positions are refined into a global frame using SLAM-based transformations and Kalman filtering. Performance evaluation conducted in real-world experiments demonstrates that EB-LiDAR achieves an average positional error of 7.13%, whereas DAIM improves accuracy with a lower average positional error of 4.24%. Experimental results demonstrate that DAIM outperforms EB-LiDAR in positional accuracy, while EB-LiDAR provides higher robustness in detecting all physical tree instances. These comparative results highlight the strengths and limitations of each method, offering valuable insights for autonomous forestry navigation and mapping under varying environmental conditions.

抄録全体を表示

This study investigated how the depth of Convolutional Neural Networks (CNNs) affects the perception of spiral illusions. Three CNNs with different layer depths were trained to distinguish between spirals and concentric circles and were evaluated using 14 illusion images. The results showed that shallower CNNs were more likely to classify the images as a spiral than deeper ones. These findings suggest that CNN depth influences the perception of visual illusions, providing insights into both artificial visual systems and human visual cognition.

抄録全体を表示

This paper introduces a structured architecture for integrating OpenStreetMap (OSM) data with ontology-based representations to advance the development and validation of automated driving systems. OSM contributes static geospatial information, while the Lanelet2 format offers detailed lane-level and topological road structure data. The fusion of these data sources enables the creation of high-fidelity simulation environments for evaluating vehicle behavior under realistic conditions. The ontology component facilitates semantic representation of roadway features and traffic regulations, thereby enabling the representation of complex and context-rich driving situations. By combining spatial and semantic layers, the proposed framework supports accurate simulation, traffic flow analysis, route planning, and autonomous decision-making processes. This integration enhances the scalability, precision, and robustness of autonomous driving applications, contributing to improved safety and reliability in real-world driving environments.

抄録全体を表示

This study focuses on Google Maps as the platform to investigate the relationship between review counts and review ratings. Specifically, we conducted a questionnaire-based survey in which participants were asked to choose between two options, each presenting a different combination of review count and rating. Based on the questionnaire survey responses, we performed a correlation analysis, a random forest analysis, and an analysis of the impact of differences in review counts to explore the relationship between review quantity and quality. The results revealed that review ratings (qualitative information) tend to have a stronger influence on users’ choices than review counts (quantitative information) when selecting tourist destinations. Furthermore, it was found that as the difference in review ratings increases, the ability of a higher review count to compensate for a lower rating diminishes.

抄録全体を表示

This study presents an in-depth evaluation and comparison of three mainstream object detection architectures — YOLOv5, YOLOv7, and YOLOv8 — in the context of intelligent music score recognition. Utilizing a labeled musical note dataset from Roboflow, all models were trained and tested under a consistent experimental workflow using Google Colab. Experimental results demonstrate that YOLOv8 outperforms the other models in key metrics such as mAP@0.5, precision, recall, and convergence stability, indicating its superior capability for small object detection and generalization. While YOLOv5 exhibited stable training behavior, its performance in precision and recall remained limited. YOLOv7, despite its theoretical advantages in inference speed, failed to deliver consistent detection results in this task. The findings validate YOLOv8’s robustness in musical note detection under complex conditions and offer technical insights for selecting optimal models for intelligent score recognition systems. This work also extends the authors’ previous research on an automatic soprano recorder-playing system, enhancing its perceptual module through improved note recognition accuracy.

抄録全体を表示

In professional tennis, ranking systems such as the ATP rankings often exhibit inconsistencies between player rankings and actual match outcomes. These discrepancies, referred to in this study as violations, are exacerbated by non-uniform tournament participation and the selective inclusion of results in official rankings. This study proposes a novel ranking method that aggregates match outcomes from all tournaments used in each player’s ATP ranking calculation. The method employs a directed graph representation and an optimization model to minimize violations while accounting for asymmetries in data inclusion. An empirical evaluation using the top 10 ATP players as of December 2023 demonstrates that the proposed method improves overall consistency with match outcomes, while causing some shifts in individual rankings. The method offers a fairer framework for evaluating player performance under conditions of selective and asymmetric tournament participation.

抄録全体を表示

This study investigates the effectiveness of sentiment analysis on Twitter-based news content in enhancing stock market prediction. Tweets from the official accounts of NHK News and Nikkei were analyzed using two sentiment analysis approaches: MeCab combined with the PN Table, and Google Natural Language API (GNL). The final sentiment scores were integrated with Nikkei 225 closing prices to train a predictive model based on Long Short-Term Memory (LSTM) networks. Experimental results indicate that incorporating sentimental features significantly improve forecast performance. While the R² value of baseline model relied solely on historical stock prices is 0.451, the R² value of best-performing model incorporating multiple sentiment indicators is 0.705. These findings show that sentimental signals extracted from financial news tweets will play an important role in stock price prediction as valuable inputs.

抄録全体を表示

Because of non-integrable constraints, we cannot regulate the nonholonomic systems towards arbitrary directions in the state space. The stabilization and tracking control laws developed for nonholonomic systems are generally mutual, i.e., the stabilization law is inapplicable for tracking uses and vice versa. The current work investigates the control problem of differentially-driven wheeled mobile robots and demonstrates an initial idea of a novel control design approach, named practical linearization control, for nonholonomic systems. We define an external dynamic oscillator and fuse it with robot states, followed by converting the nonholonomic robot model into a fully-actuated and linearizable one. Such fusion and conversion introduce a new control input without increasing the number of states to be regulated. Finally, we propose a continuous control law that can be used for both tracking and stabilization control tasks. It is, in Lyapunov’s sense, proven that the tracking errors can be driven into an arbitrarily small ball enclosing the origin. Simulation results are carried out to validate the proposed control law.

抄録全体を表示

Mobile robots are expected to play an active role in the logistics industry, and their development is underway. There are many examples of mechanisms with three omni-rollers as omnidirectional mobile mechanisms, and methods for analyzing and verifying their kinematics have been proposed. In this study, we derive a robot trajectory equation for roller drive. In addition, we verified the kinematics in a simulation environment on a PC, with the goal of reducing costs and time in a simple verification experiment with good visibility.

抄録全体を表示

Driver fatigue, feelings of emotional distress and impairment as a result of stressful events can pose significant risks of road safety. This research paper proposes a deep learning-based pose estimation system, which aims to identify unsafe driver states, by quantifying the driver's posture, head orientation, and their gesture movements. The developed model is trained in a diverse range of driving situations and identifies physiological and behavioural markers associated with fatigue driving. Unlike existing methods, it integrates pose estimation in conjunction with emotional and motion cues, allowing it to function reliably even during low-lighting or partially obscured conditions.

抄録全体を表示

Image processing dates to the 1960s when it was first applied to improve image quality. With the increased use of digital products in the form of smartphones and cameras, low-light image improvement has come into sharp focus. Several methods have been adopted such as histogram equalization, illumination map estimation, normalizing flows, neural networks, and dark region-aware enhancement. This study offers a function for RGB to CIELAB color space transformation and a step-by-step improvement process. Transformation into CIELAB color space offers the feature of separating brightness and color details and improving contrast and image quality. The device-independent CIE 1976 (L*, a*, b*) formula that is well adapted to improve images from different sources is employed. An easy-to-use interface has been implemented, enabling users to download low-light photos and restore the improved ones.

抄録全体を表示

Potato is a staple crop cultivated widely across the globe, but its production is often threatened by diseases like early and late blight, which can lead to substantial economic losses. In recent times, deep learning has proven to be an effective approach for automating plant disease identification using image-based analysis. This research explores the application of the locally adapted VGG16 deep learning framework, which was pre-trained with the ImageNet dataset. Beginning layers were frozen to exploit the benefit of transfer learning. A custom field-captured expert-annotated dataset, referred to as the Potato Leaf Dataset (PLD), obtained from Okara, Pakistan, was used to train, validate, and test the developed system. The Synthetic Minority Oversampling Technique (SMOTE), followed by comprehensive preprocessing, was applied to avoid the class imbalance issue and improve learning stability across disease categories. The model’s performance was assessed through various evaluation metrics, including accuracy, precision, recall, and F1-score. Findings suggest that the use of region-specific imagery combined with tailored preprocessing steps improves the model’s dependability in classifying potato leaf diseases. This research highlights the importance of developing context-aware and scalable AI models for agricultural use, particularly in areas with limited technical resources and internet connectivity. By focusing on a practical and locally optimized approach, the study aims to support timely disease diagnosis and contribute to improved crop management practices in rural farming communities.

抄録全体を表示

This study presents a flipper control method for crawler robots based on terrain geometry, without requiring sequential optimization. A filtered 2.5D terrain map was used to determine the target surfaces and automatically switch between the driving and traversing modes according to the terrain, extending the step-climbing sequences to 3D unstructured environments. By predicting the contact between the robot and terrain, all four flippers were independently controlled to achieve collision avoidance and posture stabilization during arbitrary movements, including turning. Simulations of step climbing, stair climbing, and oblique-step traversal demonstrated that the method suppressed excessive pitch and roll, avoided collisions, eliminated manual flipper operation, and expanded the operational range of crawler robots.

抄録全体を表示

Cyberattacks on hydraulic infrastructure pose serious risks by disrupting critical services and causing widespread societal and environmental harm. However, such incidents are difficult to investigate due to limited and fragmented data. To address this, we developed CySEC-vRT, a virtual testbed designed with reference to digital twin technology to simulate and analyze cyberattacks in water-sector ICS. The testbed integrates a Diamond Model-based framework to reconstruct cross-layer kill chains and reproduce real incidents for assessing system impact. These insights assist field operators in enhancing preparedness and developing effective defense strategies.

抄録全体を表示

The collaborative behavior of multiple systems provides valuable functions and services to users. However, it is actualized by large and complex implementations, which frequently include serious failures. In this study, we propose an automated random simulation (ARS) technique for checking the abstract collaborative behavior of multiple systems at the design level. The abstract collaborative behavior is expected to be designed using an extended place/transition net with attributed tokens (EPNAT), and the checking is performed dynamically based on the design called "EPNAT model". The ARS technique consists of (1) an algorithm for model execution using random search with the evaluation of constraints including feasibility, and (2) a stopping criterion for model execution focusing on glue transitions. The ARS technique requires tool support; therefore, we developed a prototype tool. We ran the prototype tool with a trial model and three faulty models, and found its effectiveness and future challenges.

抄録全体を表示