Data visualizations and analysis from diagrams, such as bar charts and pie charts, remain largely inaccessible to blind and visually impaired individuals, despite their widespread use in education, science, and public communication. In this work, we present a learning-based framework that automatically detects, extracts, and interprets the contents of bar charts, translating them into multimodal outputs such as text-to-speech narratives and personified representations. Our system employs a YOLOv8 object detection model trained on a custom-generated data set of annotated bar charts, capable of identifying key component of the chart, including bars, slices, angles, axes, titles and labels. Following detection, we integrate Optical Character Recognition (OCR) to decode textual labels and numerals, while a pixel-to-value mapping algorithm interprets bar heights relative to axis scales. The extracted information is dynamically translated into audio descriptions and interactive query-based speech outputs, enabling non-visual comprehension of chart data. To ensure usability, we designed the output flow to follow a hierarchical structure from summary to specific values. Preliminary evaluations show that the system achieves high accuracy in detecting component parts of the chart and generating bar values, with promising feedback from a blind user in pilot tests. This research contributes to inclusive AI, opening new directions for accessible data communication through the integration of computer vision, speech synthesis, and assistive reasoning in an abstract environment.
This study presents a comprehensive investigation into optimizing a SQL Server in a business production environment that was experiencing severe performance problems and frequent system freezing. We implemented a systematic methodology consisting of three key phases: thorough diagnosis, strategic optimization, and longitudinal monitoring. Our findings demonstrate statistically significant improvements across critical performance indicators, including a 26% reduction in CPU utilization, a 4.3% increase in available memory, stabilization of the Page Life Expectancy indicator at optimal values exceeding 300 seconds, and complete elimination of system freezing episodes. This research provides a transferable methodological framework for SQL Server optimization across diverse business environments while contributing valuable insights to the efficient management of relational databases in high-availability contexts.
Web-based supervisory interfaces for distribution networks require a communication channel that delivers timely telemetry to browser clients while preserving ordered and consistent delivery with low overhead. This paper presents the communication core of LineWatch, a campus-scale monitoring platform that adopts a WebSocket-based design to stream feeder measurements and fault notifications over topic-scoped channels such as telemetry/line/feederId and event/fault. A lightweight server broadcasts compact JSON envelopes, and a browser dashboard binds map markers, device tables, and per-device histories directly to the most recent frames. Functional validation from the running system demonstrates four key mechanisms: (1) selective fan-out that keeps the device list and map view synchronized without overloading unsubscribed clients with irrelevant telemetry, (2) in-place popover binding that exposes the latest server timestamp and prevents regressions, (3) an ordered and coherent snapshot across devices in the tabular view with identical render timestamps, and (4) bounded per-device replay retrieved in reverse chronological order at a stable five-second cadence, enabling contextual inspection without interrupting live delivery. The findings indicate that a small, domain-aware contract layered on standard WebSockets can satisfy the timeliness and consistency requirements of real-time distribution dashboards while remaining simple to deploy behind conventional web infrastructure. The paper also discusses security and scalability considerations for long-lived browser connections, including TLS termination, token-based access control, topic-level authorization, and brokered fan-out.
The integration of Virtual Reality (VR) into higher education has gained significant attention for its ability to provide immersive and experiential learning opportunities. This pilot study examines the impact of using VR on Year-2 students' engagement, learning, and professional skills development within the Ethical and Professional Issues module in a UK University, with a specific focus on employability skills. The research utilizes BodySwaps, a VR simulation tool designed to enhance students' soft skills, particularly job interview skills. Through interactive and immersive experiences, students engaged in simulated interviews, receiving real-time feedback on their articulation, body language, eye contact, and use of filler words.To evaluate the effectiveness of VR-based interview training, feedback collected from students is used for sentiment analysis. The study examines deep learning models for sentiment analysis techniques for classification. BERT and DistilBERT are applied to detailed feedback from participants. Each approach is used to classify feedback into either positive or negative categories, with the most effective method determined based on accuracy, precision, recall and F1-Score. DistilBERT model slightly outperform the BERT model with higher speed and lower resources.
Effort estimation is an important asset for software project management, which supports in efficient planning, scheduling and cost estimations. However, it remains a challenge to predict effort precisely as requirements are uncertain, environment is dynamic and software systems evolve with frequent additions of new functionality and removal of old functionality. Traditionally, solo techniques such as expert judgment, regression models, analogy-based estimation, support vector machines and neural networks have been applied to predict software development effort. While these methods have achieved some degree of success, predictability for them is mostly limited by their stability. In attempting to be responsive to such demands, ensemble learning methods have gained importance over the recent years. These involve aggregating several models in order to minimize bias, variance, and enhance robustness. This paper investigates whether ensemble methods outperform solo techniques in enhancing the accuracy of software effort estimation. We did empirical analysis by evaluating popular ensemble models, including ExtraTrees, HistGB, BootstrapNN, and Random Forest against solo machine learning approaches. The evaluation of these models has been performed using standard performance metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and the R-squared (R2) statistic. Findings from this study shows that ensemble techniques generally deliver higher estimation accuracy than solo models, demonstrating their strength in handling heterogeneous and uncertain software project datasets. On the other hand, some single models, like BootstrapNN and Linear Regression, performed better than ensemble models on specific datasets. This shows that ensemble models tend to be more reliable and accurate, solo models can still deliver strong results, especially when the dataset has certain characteristics.
Physical Artificial Intelligence (AI), or embodied AI, represents a paradigm shift from purely virtual intelligence to systems with a physical presence capable of perceiving, reasoning, and acting upon the world. This survey provides a comprehensive review of the foundational concepts of Physical AI, highlighting the critical role of foundation models and integrated development platforms. As a case study, we conduct a deep vertical analysis of the NVIDIA ecosystem, examining how simulation (Omniverse, Isaac Sim), synthetic data generation (Cosmos), robot learning (Isaac Lab), perception (Isaac ROS), and edge computing (Jetson) create a self-reinforcing development cycle. We further analyze state-of-the-art architectures, comparing language-grounded planners to generalist Vision-Language-Action policies. Our analysis reveals a clear convergence towards hybrid architectures that combine deliberative reasoning with reactive control, heavily relying on standards like OpenUSD for interoperability. The survey concludes by discussing key limitations, such as the sim-to-real gap and proprietary lock-in, emphasizing the convergence of learned models and classical techniques in shaping the future of autonomous systems.
This paper presents a series of measurement campaigns evaluating a outdoor hybrid integrated 5G GNSS RTK positioning system with 5G networks operating in Frequency Range 2 (FR2, mmWave) and for comparison also on Frequency Range 1 (FR1). The experimental setup combines high precision GNSS receiver with mobile terminals equipped with Quectel RG530F-EU 5G FR2 compatible modem. Field measurements conducted on the TalTech campus area provide insights of horizontal positioning accuracy, RTK fix stability and the influence of FR2 specific signal characteristics on the reception of correction data. The results demonstrate the feasibility of achieving centimeter level positioning accuracy over 5G mmWave links in favorable conditions while also highlighting key challenges such as limited FR2 coverage.
The fusion of advanced sensing technologies with medical imaging has revolutionized diagnostic capabilities, particularly for neurodegenerative diseases like Alzheimer’s. Magnetic Resonance Imaging (MRI), a prominent tool for sensing and visualizing brain structures, is often impaired by noise artifacts such as Salt-and-Pepper noise. This study introduces a robust approach to denoising MRI scans by optimizing the Median Filter—a widely used method for impulse noise reduction. Our work proposes key algorithmic improvements to enhance performance under varying noise densities, preserving critical details required for accurate segmentation and diagnosis. Experimental evaluation demonstrates superior performance compared to standard methods, showcasing the integration of sensing and image processing technologies for improved clinical outcomes.
Multihorizon time series forecasting (TSF) tasks are a crucial component of smart city systems, enabling decision-makers to take proactive measures in traffic, energy, transportation, and environmental management. Although existing deep learning models are effective in capturing hierarchical temporal dependencies, they often assume static inter-station dependencies, rely on single temporal operators that fail to capture both short-term transients and long-term cycles, and use autoregressive decoding that amplifies early errors via exposure bias. To tackle these limitations, we introduce a Station-Aware Patch-Gated Temporal Transformer (SAPGTT), an end-to-end framework for robust multi-horizon forecasting. Specifically, we propose a station-aware spatial mixer that applies self-attention across stations at each time step, dynamically learning time-varying dependencies without relying on fixed dependency assumptions. Then, we introduce a hybrid temporal encoder that incorporates bidirectional long-short-term memory (Bi-LSTMs) for short-to-medium dynamics, along with temporal self-attention to capture long-range seasonal patterns. In addition, a patchwise gated convolutional forecaster segments sequences into motifs and employs gating to suppress noise, producing direct multi-horizon outputs that eliminate error accumulation from iterative rollouts. Extensive experiments on a benchmark urban air quality fore-casting task demonstrate that SAPGTT consistently outperforms state-of-the-art deep learning baselines, achieving up to a 17.0% reduction in MAE and 18.8% in RMSE, delivering accurate, stable, and horizon-consistent forecasts.
Safe, well-connected infrastructure is essential to increasing the share of cycling as an urban mode of transport. Yet cities often face budgetary and scheduling constraints when deciding where to expand their cycling networks. This paper presents a data-driven framework for prioritizing cycling infrastructure investments using GPS records from Tartu’s public bicycle-sharing system (PBSS) collected between 2020 and 2023. The dataset comprises nearly three million trips and approximately 330 million location points with trip identifiers. After map matching and spatial aggregation using H3 hexagonal grids, we analyze cycling activity at fine spatial resolution. We develop three complementary prioritization strategies, current demand, potential demand, and network connectivity, to identify critical network gaps and compare them with the city’s existing master plan. Finally, we apply the MULTIMOORA multi-criteria decision-making method to integrate these strategies and produce a ranked list of candidate bikeway segments. The combined analysis highlights several high-priority corridors not captured in current planning documents. Our results illustrate how large-scale micromobility data, combined with multi-criteria decision analysis, can inform efficient, user-oriented, and evidence-based cycling infrastructure planning in smart urban environments.
Deepfake images generated by GANs pose serious security and ethical concerns. Most detection models rely on learning GAN fingerprints, which can be suppressed through cross-domain translation. In this study, we evaluated the robustness of an Xception-based deepfake detector using the 140K Real and Fake Faces dataset, where artifact-free fakes were produced via Stable Diffusion image-to-image translation. A three-class dataset of real, original fake, and translated fake images was constructed, and FFT analysis confirmed reduced high-frequency artifacts after translation. Experimental results show the model achieved over 98% F1-score on real and original fake images and only 0.679 on translated fakes due to a recall drop to 53%. These findings highlight the vulnerability of current CNN-based detectors to domain-shifted, artifact-free deepfakes.
Quantum computing is a new approach to data processing based on the quantum mechanics principles. Its unique capabilities, such as superposition and entanglement, open new horizons for the development of intelligent distributed systems, including the Internet of Things (IoT). This article discusses the basic principles of quantum computers, the current state of technology, and their impact on the future of the IoT. Quantum algorithms can significantly improve the efficiency of data processing, optimization of network resources, and cybersecurity in the IoT environment. In addition, quantum cryptography provides a higher level of protection for information transmitted between IoT devices. The development of quantum computing can change existing paradigms in the IoT field, contributing to the creation of more scalable, secure, and intelligent systems. This opens new prospects for the implementation of quantum solutions in areas such as smart cities, healthcare, industry, and autonomous transport.
Network coverage planning is a critical challenge to ensure seamless communication while minimising deployment costs, particularly in heterogeneous networks with multiple cell types and technologies. We introduce the Be.Neutral HetNet Planner, a novel graph-based framework that converts a spatial street-level graph into a logical connectivity graph, enabling flexible and efficient network design. The methodology supports any metric, such as coverage, cost, density, or latency, and the logical connectivity graph creation can be accelerated using GPU parallelisation with PyCUDA for large-scale scenarios. As a use case scenario, we optimise Wi-Fi access point placement to complement a 5G network, using coverage as metric. Our proposal combines heuristic placements with mixed-integer programming to create network deployment solutions. The process is illustrated with step-by-step visualisations, and the results show that Be.Neutral HetNet Planner is an effective, and extensible solution for network planning and optimisation in urban environments.
Reliable evaluation of AI-generated programming code remains a critical challenge due to the dual necessity of ensuring semantic correctness and maintaining structural integrity. Existing evaluation methods often fail to capture both these aspects comprehensively, limiting the understanding of AI code generation quality. To address this gap, we propose a multi-dimensional evaluation framework that integrates semantic similarity metrics derived from CodeBERT with structural complexity analyses based on abstract syntax trees (AST), program dependence graphs (PDG), and control flow graphs (CFG). Our experimental setup involved fine-tuning the CodeT5 model (220 million parameters) on 232 programming language problems sourced from the CodeForces platform, generating a total of 2,320 code samples in C++ language. Semantic evaluation revealed a median CodeBERT similarity score of 81.53, indicating consistent semantic alignment between generated and reference codes. Structural complexity metrics exhibited significant variability, with complexity scores having a median of 91.34. These results illuminate both the strengths and limitations of current AI code generation models, underscoring the need for combined semantic and structural assessments.
Large Language Models (LLMs) are increasingly being applied in qualitative analysis for tasks such as theme extraction, frequency analysis, and impact evaluation. However, their adoption raises privacy and GDPR compliance concerns when transcripts are processed using commercial LLMs such as ChatGPT or Gemini. Existing studies highlight these risks but provide little systematic evidence for comparing local and commercial LLMs. This study evaluates the performance of local LLMs such as LLaMA-3.1 (8B), LLaMA-3.2 (1B +/- 3B), LLaMA-3.3 (70B), Gemma-2 (2B +/- 27B), and Phi-3.5 (3.5B +/- 6.6B) against commercial LLMs (ChatGPT-4o and Gemini-2.5 Flash) using 82 anonymized transcripts for qualitative analysis tasks. A structured prompt design was applied, and the results were benchmarked against ground-truth coding using cost, throughput, hallucination rate, and accuracy rate. The findings indicate that the small local LLMs (about 3B) performed comparably close to Gemini, medium models (6-9B) performed close to ChatGPT, and large LLMs (27B +/- 70B) consistently outperformed both commercial LLMs. Hallucination reduction of up to 85% was observed with local LLMs at negligible recurring costs. Furthermore, local LLMs help with GDPR compliance and privacy preservation. It also minimizes cost while delivering accuracy that is comparable, or better than the commonly available commercial LLMs.
This paper introduces and validates Affective-Emotive Resonance (AER), a phenomenon of sustained, layered emotional connection observed in a digital human. While contemporary Large Language Models (LLMs) can simulate reactive empathy, a critical gap remains in their ability to foster enduring, relationship-like dynamics. This study addresses this gap by proposing that such relationality is not an inherent or automatic byproduct of the model scale but an emergent property that can be intentionally cultivated through design. We define AER and present a four-dimensional framework for its evaluation: Persistence, Belonging, Continuity, and Layered Emotional Care. Using this framework, we conducted a controlled experiment comparing two empathic digital human personas against a neutral baseline, all of which were built on the same foundational LLM. The results demonstrate that while the baseline model exhibits a negligible AER, the scaffolded personas manifest high degrees of resonance in distinct affective registers. This study provides empirical evidence that persona scaffolding is a primary mechanism for architecting relational AI, shifting the research focus from whether AI can be empathic to how specific, beneficial, and safe relational dynamics can be designed reliably. This finding underscores the profound ethical responsibility and "duty of care" inherent in creating emotionally resonant AI systems.
LoRaWAN is increasingly considered for mobile IoT applications, yet its performance under dynamic conditions is still subject to thorough evaluation. This study investigates the impact of mobility on LoRaWAN communications in vehicular scenarios, focusing on key performance indicators such as packet loss, signal strength, and inter-packet gap. Experiments were conducted on an isolated road segment using a LoRaWAN field tester mounted on an e-scooter and a car, with motion speeds of 40, 60, and 80 km/h. Three spreading factors (SF7, SF9, SF12) were evaluated under identical conditions. Results show that a trade-off is necessary between good communication range and mobility support, since lower SFs provide reliable connectivity at higher speeds, while higher ones, despite better sensitivity, suffers from significant degradation due to Doppler effect, changing environmental conditions, and longer time on air. Findings emphasize the importance of selecting transmission parameters based on mobility profiles and environmental conditions. The work contributes to the understanding of LoRaWAN behavior under mobility and supports the development of more resilient and adaptive communication strategies for dynamic IoT deployments.
Body-focused repetitive behaviors (BFRBs) are self-directed habits that can cause physical harm and psychosocial challenges. Early detection of these behaviors offers opportunities for timely intervention and improved treatment. In this work, we investigate the use of deep learning models applied to data collected from the Helios wrist-worn device, developed by the Child Mind Institute, which integrates an inertial measurement unit (IMU), thermopile sensors, and time-of-flight (TOF) sensors. We develop and evaluate a variety of predictive models including multilayer perceptron (MLP) and random forest models whose inputs are frequency domain representations of sensor data, a CNN-BiLSTM model that effectively leverages TOF data, and ensemble models that combine MLP and BiLSTM models. Among these methods, we find that a late fusion based ensemble approach achieves the highest binary F1-score of 96.86% in distinguishing BFRBs from non-BFRB gestures. We also find that models trained on IMU data in conjunction with data from thermopiles and TOF sensors significantly outperform single-modality baselines with the exception of the CNN-BiLSTM model trained on TOF data that outperforms the MLP trained on all sensor data. Finally, we find that distinguishing between different types of BFRBs is significantly more challenging than distinguishing BFRBs from innocuous gestures.
Energy-aware task scheduling in large robotic swarms is challenging due to computational complexity, communication overhead, and limited onboard energy. Traditional centralized schedulers struggle to scale, while fully distributed approaches often lack global coordination. To address this, we propose a hybrid scheduling framework in which a central coordinator performs global task allocation using a Multi-Verse Optimizer (MVO), and the swarm itself participates in parallel local schedule refinement. After receiving their assigned task subsets, individual robots refine execution order using Particle Swarm Optimization (PSO) or Genetic Algorithms (GA), allowing computation to be distributed across the swarm rather than concentrated at the center. This significantly reduces central processing demand by offloading while adapting task execution to local energy conditions. The framework is implemented in a ROS +/- Docker environment with explicit energyaware scheduling. Experimental results show that the approach reduces scheduling computation time by up to 250% and lowers mission makespan by 5 +/- 7% compared to six state-of-the-art methods, while improving overall energy efficiency and coordination. These findings demonstrate that combining centralized global insight with distributed parallel refinement enables scalable and energy-efficient swarm task scheduling suitable for large-scale deployments.
The scarcity of diverse and abundant image data poses a significant challenge to training robust vessel classification models for maritime security applications. This study examines the application of synthetic data generation to address this limitation. We evaluate and compare three state-of-the-art generative models, Denoising Diffusion Probabilistic Model (DDPM), Deep Convolutional Generative Adversarial Network (DCGAN), and Variational Auto-Encoder (VAE) for synthetic vessel image generation. The generated images are used to augment a real dataset at various ratios (10% to 1000%) for training a YOLOv12n classifier. Results demonstrate that synthetic data, particularly from DDPMs, which produced the highest quality images (lowest FrAechet Inception Distance (FID) score), significantly enhances model performance. The DDPM-augmented approach achieved a maximum accuracy improvement of 5.47%, outperforming both classical data augmentation techniques and the other generative models. This work confirms that high-quality synthetic data is a powerful tool for improving the accuracy and robustness of vessel classification systems. All the source code is publicly available
