Search Results (1,194)

With the continuous growth in motor vehicle ownership, traffic safety risks caused by dangerous driving behaviors remain an important concern. This paper presents RMDD, a Raspberry Pi-based multimodal dangerous driving monitoring feasibility prototype using YOLO26 visual models, personalized facial fatigue estimation, emotion2vec-based speech recognition, and hierarchical reliability-gated BFV fusion (HRG-BFV). HRG-BFV treats object detection as positive-only support for behavior evidence and scales its contribution by development-fold reliability, thereby avoiding hard rejection when the object detector fails. On locked module tests, behavior classification achieved 96.77% Top-1 accuracy, auxiliary detection reached 0.916 mAP@0.5, personalized facial calibration reduced the false-positive rate from 0.877 to 0.211, and speaker-disjoint speech recognition achieved 0.890 accuracy at 0 dB vehicle noise. The historical fixed BFV baseline achieved 0.601 balanced accuracy and 0.506 macro-F1 on P03–P08 (96 clips). In a stricter three-fold leave-one-external-participant-out evaluation on P04/P05/P08 (48 clips), HRG-BFV achieved 0.726 balanced accuracy, 0.562 macro-F1, 0.833 specificity, and 0.677 ROC-AUC, versus 0.655, 0.469, 0.833, and 0.595 for the historical frozen hard gate on the same cohort. Target support reliability was low (0.097–0.194), so the adaptive term appropriately reverted toward behavior evidence rather than producing an unsupported object detection gain. A sealed P08 behavior test exposed substantial cross-subject degradation (0.320 frame accuracy). A cooled 30 min Raspberry Pi 5 run completed without thermal throttling at 0.927 processing windows/s. The evidence supports controlled prototype feasibility but not population-level or on-road generalization. Full article

Airside vehicles operate close to aircraft, service equipment and safety-critical surface markings, yet driver-assistance functions developed for road traffic do not directly transfer to airport aprons, taxiways and service roads. This article presents a vision-based driver-assistance and warning prototype for recognizing airport surface types and horizontal markings in video recorded at Poznan Airport. Two manually annotated segmentation datasets were prepared from GoPro HERO8 video acquired from a vehicle perspective: a four-class surface dataset covering asphalt, concrete, paving blocks and grass, and a three-class marking dataset covering red, white and yellow lines. The study compares You Only Look Once (YOLO) variants YOLOv8 and YOLOv11 with U-Net, DeepLabV3 and SegFormer under a common 512-by-512 input resolution and evaluates both model-level quality and complete video-application behavior. For semantic segmentation, SegFormer achieved the highest validation results, with Intersection over Union (IoU)/Dice of 0.7657/0.8624 for surfaces and 0.8852/0.9380 for markings. Among YOLO models, YOLOv8m obtained the highest surface mean average precision at an IoU threshold of 0.50 (mAP@50) of 0.7847, whereas YOLOv8s obtained the highest marking mAP@50 of 0.8449. On video recordings, paired YOLO configurations processed approximately 15–16 frames per second (FPS) on a personal computer (PC), while U-Net, DeepLabV3 and SegFormer processed approximately 10–11 FPS. A YOLOv8n pair compiled for Raspberry Pi 5 with Raspberry Pi AI HAT+ Hailo-8 reached 10.05 detection FPS and 18.15 processing FPS without GUI rendering. Under the adopted evaluation protocol, SegFormer achieved the highest mask quality in the conducted comparison, while the paired YOLOv8n configuration demonstrated embedded throughput feasibility. Full article

Food–medication interactions can occur when medicines are taken with foods, drinks, herbs, or supplements that influence drug absorption, exposure, or activity. Screening these combinations is challenging because the available evidence is imbalanced, prescription text may be recognized incorrectly, unsupported inputs may produce unreliable predictions, and altered software artifacts may change the recommendation presented to the user. iMediFood-Shield addresses these concerns through an evidence-first edge-AI framework that combines structured diet–drug interaction evidence, prescription-assisted medication confirmation, coverage-aware rejection, calibrated five-class prediction, false-safe-aware confidence gating, and software-based tamper-evident verification. The DDID preparation process began with 23,950 evidence records and produced 16,644 canonical medication–food/herb pairs, including 16,165 single-effect model-eligible pairs and 479 multi-effect conflict pairs. A leakage-free 70%–15%–15% split was applied after canonicalization, and the deployed lookup was restricted to training-supported and conflict records. On the operational locked-test AI branch of 2259 supported unseen pairs, the final calibrated LinearSVC with the validation-selected MedSafe-GATE threshold of 0.65 achieved 91.72% accuracy, 80.57% balanced accuracy, and a macro F1-score of 0.8359. The gate reduced calibrated false-safe predictions from 55 to 28, corresponding to a 49.09% reduction and a final false-safe rate of 1.35% among interaction-bearing AI-branch pairs. RxOCR-Guard achieved 94.67% candidate recall and 100.00% candidate precision on a controlled synthetic prescription benchmark, while mandatory user confirmation was retained because top-1 candidate accuracy was 51.33%. The unchanged baseline and all ten adverse software-bundle conditions produced the expected verification outcomes for artifact-modification, missing-file, key-mismatch, manifest-alteration, and rollback cases. Raspberry Pi deployment reproduced all 2259 reference predictions without mismatch, completed covered AI inference in 1.737 ms on average, and verified the protected software bundle in 80.249 ms on average. These results show that iMediFood-Shield can combine evidence-grounded screening, conservative AI decision control, prescription confirmation, and software-integrity verification within a resource-constrained edge research prototype. Full article

Environmental conditions have become more unstable; therefore, innovative and eco-friendly methods of food production are urgently required. Most existing hydroponic systems lack the capacity for real-time responses and decision-making based on integrated data, similar to contemporary farms. This document outlines the creation of an advanced hydroponic farming system that utilizes Internet of Things (IoT) sensors and a digital twin (DT) simulator to address these challenges. A completely monitored and continuously assessed hydroponic farming simulator operating on a Raspberry Pi, employing various sensors, data management and processing, and automated environmental regulation. The development of this intelligent hydroponic farming system employs a dual-model machine learning pipeline: one that identifies plant diseases through image analysis, and another that assesses plant nutrient levels based on sensor data. The data from the two models are combined using a cloud-based DT, enabling remote access to the DT and offering closed-loop control for irrigation, nutrient dosing, and management of all environmental factors related to crop growth in a hydroponic setting. This research showcases the capability to develop scalable, data-focused precision agriculture solutions that can adapt to the demands of today’s agricultural environment by combining all elements of IoT sensing, machine learning, and DT simulations into one functional hyperphysical system. Full article

The automatic inspection of airport infrastructure, specifically horizontal surface markings and Airfield Ground Lighting (AGL), is a critical task for maintaining aviation safety and operational efficiency. As the aviation industry shifts from manual surveys to automated visual inspections utilizing unmanned aerial vehicles (UAVs) and smart service vehicles, the demand for robust, real-time computer vision systems has surged. However, deploying computationally intensive deep learning models in the field introduces severe Size, Weight, and Power (SWaP) constraints. This paper presents a comprehensive framework for the semantic segmentation of runway/taxiway markings and the point-localization of AGL lamps, specifically focusing on the deployment paradigm shift from the expensive, GPU-accelerated heterogeneous system on chip (SoC) to highly efficient, dedicated Neural Processing Units (NPUs). We evaluate the performance of U-Net, LinkNet, U-Net-Point and HRNet-Lite-Point architectures trained on a custom dataset from the Poznań-Ławica Airport. Crucially, this study conducts a rigorous comparative hardware analysis between the flagship NVIDIA Jetson AGX Orin and a highly cost-effective heterogeneous setup comprising a Raspberry Pi 5 augmented with an NPU Hailo-8 AI accelerator. Experimental results demonstrate that while both platforms achieve real-time inference, the Hailo-8 integration fundamentally disrupts the traditional cost-to-performance ratio. Furthermore, the Hailo-8 configuration consumed less electrical power and memory footprint required by the Jetson, proving that dedicated NPUs are vastly superior for continuous, battery-operated edge deployment in autonomous airport maintenance systems. Specifically, the Raspberry Pi setup with the Hailo-8 accelerator demonstrated superior energy efficiency, requiring a significantly lower energy consumption per processed video frame compared to the Jetson platform. Full article

Small-scale and urban indoor farms typically rely on manual observation, which delays stress detection and yields inconsistent crop quality. While large language models (LLMs) and retrieval-augmented generation (RAG) have been explored for agricultural advisory systems, their integration into a single cloud-free indoor-farming platform that couples multimodal symptom interpretation with autonomous environmental control remains largely unexamined. This study presents an integrated platform built around an agentic AI advisory pipeline that runs entirely on-device on commodity hardware. The pipeline couples a RAG-grounded Mistral 7B language model with a LLaVA 7B vision-language model through condition-based routing, intent classification, multi-step reasoning, and an LLM validation gate, delivering context-aware text and image-based symptom-level guidance from a conversational interface. The advisory layer operates alongside vision-based plant monitoring and a deliberately isolated threshold-based control layer, in which an ESP32 microcontroller autonomously actuates irrigation and lighting against predefined thresholds while a Raspberry Pi 5 performs continuous plant detection and browning monitoring. On Cos lettuce, the advisory pipeline achieved 82.00% weighted accuracy across 50 queries spanning health, symptom, watering, pest, root-health, and growth-stage categories, scored against established plant pathology and postharvest references, with no incorrect responses recorded. The study contributes the design of an agentic advisory pipeline and its integration into a working, cloud-free indoor-farming platform, providing an on-device foundation for intelligent small-scale farming. Full article

Quality control in green coffee bean production is critical for food safety and economic sustainability. A persistent gap in existing automated inspection systems is their inability to capture both sides of each bean, which leads to systematic under-detection of surface defects. This work presents three contributions: (i) a novel mechatronic apparatus that mechanically guarantees dual-sided imaging of every bean, (ii) a public 12-class dataset of green coffee bean defects, and (iii) an embedded, real-time inspection pipeline validated on low-cost hardware. The apparatus sequentially presents each bean, from a standard 350 g sample, to two 16-megapixel cameras under controlled LED illumination. A dataset of 9600 images spanning 12 classes (11 defects and 1 normal) was generated from expert-classified samples and enriched through data augmentation. Four convolutional neural network (CNN) architectures, VGG-16, VGG-19, ResNet-50 and YOLOv8, were trained and benchmarked using precision, recall, F1-score and mean average precision. YOLOv8 achieved the best overall performance, with a precision of 97.4%, a recall of 99.6%, an F1-score of 0.930 and a mean average precision of 96.5%, outperforming VGG-16 (accuracy 86.07%), VGG-19 (accuracy 67.03%) and ResNet-50 (accuracy 87.76%). Dual-sided acquisition raised mean per-class detection accuracy from 0.727 to 0.908, a relative gain of 25.7% over an equivalent single-sided configuration. Deployed in real-time “track” mode on a Raspberry Pi 4, the system simultaneously classifies defects and counts beans by category, processing a 350 g sample in approximately 38 min. Combining mechanical innovation with lightweight deep learning enables practical, scalable, and cost-effective quality control for laboratories specialized in coffee analysis. Full article

Falls are a major cause of severe injury and mortality among older adults, requiring rapid state recognition and alerts in healthcare and caregiving environments. However, many existing fall-recognition approaches primarily focus on post-fall detection or assume access to cloud/GPU computing, leaving limited evidence for short-term pre-fall recognition on low-end CPU-based edge devices. This study proposes a lightweight CNN–TCN–Attention framework for recognizing fall and pre-fall states on resource-constrained edge devices. Using MediaPipe, three-dimensional coordinates and visibility scores of 33 human body landmarks are extracted from input videos. Raw RGB frames are not directly used by the classifier, thereby reducing the processing of visually identifiable information. A total of 2300 fall-related videos from the AI-Hub dataset were reorganized into three classes: normal, pre-fall, and fall. The pre-fall onset was defined as the point at which the downward vertical velocity of the nose landmark exceeded a dataset-specific threshold. The model uses a CNN to extract frame-level skeletal patterns, a temporal convolutional network (TCN) to learn temporal changes in posture, and an attention module to aggregate temporal features. It was deployed on a Raspberry Pi Zero 2 W using ONNX Runtime. The proposed model achieved 94.71% accuracy and a macro F1-score of 93.63%, with an average state-decision latency of 325.88 ms/decision. It improved accuracy by 2.83 percentage points over the lightweight CNN–LSTM–Attention baseline while maintaining comparable latency and achieved approximately 1.75× faster processing than the MobileNetV3–TCN–Attention model. Full article

Unstable rock blocks and cliffs pose a widespread geohazard, and their mechanical state can be tracked through their ambient-vibration resonance frequencies, whose decrease anticipates progressive failure. Such monitoring is usually performed with expensive broadband instrumentation, limiting spatial and temporal coverage. Here we assess whether a low-cost, IoT-enabled node—built around a Raspberry Pi single-board computer, a 24-bit sigma-delta digitiser and a force- balance accelerometer (Geobit FBA-200)—can identify the resonance of an unstable coastal rock block at the celebrated “moving rock” of Kounopetra (Paliki peninsula, Kefalonia, Greece), a site historically renowned for visually perceptible rocking boulders. We stress that the low-amplitude 7.7 Hz structural eigenvibration characterised here is a distinct phenomenon from the historically documented ∼0.3 Hz macroscopic, quasi-rigid rocking of the boulder: the former is the ambient–vibration resonance of the fractured rock mass, the latter a large-amplitude rigid-body oscillation. Two identical nodes recorded ground acceleration simultaneously for nine hours: one on the fractured Kounopetra rock mass and one on stable ground 25 m away, used as a reference. The rock station exhibits a clear, temporally stable fundamental resonance at f₀ = 7.7 Hz (Q ≈ 50, damping ζ ≈ 1%), amplified by up to an order of magnitude relative to the reference and entirely absent from it, whereas a narrow 20.5 Hz line present on both nodes is identified as instrument-related and discarded. A simultaneous two-station analysis further shows that the ambient sources are extremely local (only 0.4% of transient activity is common to the two nodes 25 m apart), quantifying a design constraint for differential schemes. The results demonstrate that a low-cost force-balance node is sufficient to establish a resonance baseline for an unstable rock block, opening the way to dense, affordable early-warning networks; the main limitations are the single vertical component and the short record, which preclude polarisation analysis and long-term tracking of f₀. Full article

Deploying human pose estimation on low-cost, accelerator-free hardware remains a key open question for smart home activity recognition, where affordability and wide availability often matter as much as raw inference speed. This work addresses that question by evaluating thermal-image pose estimation on a Raspberry Pi 5, a CPU-only single-board computer with no dedicated neural-network accelerator. Building on a previously collected single-person thermal dataset of 2500 images annotated with 17 body keypoints, YOLOv8-pose, YOLO11-pose, and YOLO26-pose models were trained across five model scales (n–x) at three input resolutions ($640\times 512$, $320\times 256$, and $160\times 128$ px) and deployed on the Raspberry Pi 5. Models were evaluated in native PyTorch, ONNX, and NCNN export formats, with per-image latency, power consumption, energy per frame, and pose mAP50–95 measured for each configuration. Pose accuracy scaled consistently with model size and input resolution, reaching up to 95.6% pose mAP50–95 for YOLO26-x at $640\times 512$ px, with diminishing returns for the largest model scales. Export-format conversion to ONNX or NCNN preserved pose accuracy almost exactly relative to native PyTorch, with mean degradation below 0.6 percentage points, in contrast to the accuracy loss associated with numerical-precision reduction on GPU-accelerated platforms. NCNN consistently achieved the lowest per-image latency and energy consumption among the three formats, delivering up to 2.1–4.3× speed-up over PyTorch while keeping power draw up to 10 W across all configurations. The findings indicate that, unlike GPU-based edge deployment where precision reduction is the dominant efficiency driver, export-format optimization governs the latency–power–energy trade-off on CPU-only hardware, and that NCNN-based deployment on low-cost, widely available single-board computers offers a practical alternative for smart home thermal pose estimation. Full article

Federated learning (FL) enables training on edge devices, but update-based defenses may overlook runtime tampering, control-flow manipulation, and microarchitectural attacks. RAFALE strengthens integrity through verifiable runtime evidence collected during local training. Control-Flow Integrity (CFI) evidence and Hardware Performance Counter (HPC) measurements obtained through the normal-world performance-monitoring interface are transferred to an OP-TEE Trusted Application, where they are bound to the training round and model update, digested, and signed for server verification. RAFALE was implemented on Raspberry Pi 3B+ clients and evaluated using convolutional neural networks on MNIST and CIFAR-10. Physical-device trials distinguished benign execution from loss-function omission and indirect-target redirection. At $T=2$ and $K=2$, HPC-based detection achieved 83.88% accuracy, 91.94% precision, 76.39% recall, an F1-score of 83.44%, and a 7.60% false positive rate. Combined CFI/HPC monitoring required $29.653\mathrm{s}$ per Raspberry Pi training run. Authentication of one physical-device evidence and model record averaged $1.513\mathrm{ms}$, and each evidence record occupied $588\mathrm{bytes}$. A 64-client simulation showed that admitting redirected adversarial updates reduced balanced CIFAR-10 accuracy from $68.59\%$ to $25.49\%$ at 50% malicious participation. These results show that authenticated runtime evidence can distinguish the evaluated execution deviations without revealing local training data. Full article

Handwriting is a fundamental fine motor skill in early childhood development, yet between 10% and 30% of school-age children experience significant difficulties in its acquisition. Existing automated assessment approaches predominantly classify whether the correct character was produced, rather than evaluating the morphological quality of the stroke itself—the level at which handwriting difficulties are believed to manifest. This article addresses this gap by presenting the design, implementation, and technical evaluation of an interactive social robot shaped like a capybara, developed to support Spanish-language handwriting learning in preschool children through stroke-level, rather than character-level, assessment. The system integrates a Raspberry Pi 5, a 15.6-inch touchscreen, and a stroke morphological comparison algorithm implemented with the Open Source Computer Vision Library. The evaluation engine performs preprocessing, region-of-interest masking, and pixel-level coverage analysis based on the standard recall formulation, translated into 1-to-5-star multimodal feedback. A controlled technical evaluation of 360 trials, conducted by four trained adult evaluators, yielded an overall recognition rate of 82.78% (95% CI: 78.54–86.33%) and a mean response time of 0.68 s, well below the threshold identified in the literature as critical for sustaining engagement in preschool children. Recognition was statistically equivalent across character categories (p = 0.547) but differed markedly across stroke-quality levels (p < 0.001), evidencing the formative sensitivity of the algorithm. Exploratory observations in two Peruvian preschools indicated operational stability and children’s spontaneous engagement with the system. These results position the prototype as a technically validated, replicable foundation—based on general-purpose embedded hardware—for future pedagogically oriented research on child–robot interaction in handwriting instruction. Full article

A control that cannot be audited is a control that does not yet exist operationally. Commercial off-the-shelf (COTS) Internet of Things (IoT) devices in Ambient Assisted Living (AAL) environments expose this problem sharply: they export raw behavioural telemetry rather than the security-auditable event records required by the logging and monitoring controls of ISO/IEC 27001:2022. This study formalises that deficit as the Admissibility Gap, a weighted, field-level measure of the mismatch between native device output and the evidentiary requirements of Annex A. An audit of four publicly available AAL datasets (CASAS, SPHERE, UCI HAR, OPPORTUNITY) confirms that identity attribution, integrity, firmware version, and privacy-minimisation governance fields are universally absent, establishing that the gap is systemic. To close it, this study proposes the Compliance Transformation Layer, an edge middleware applying three rules: LDAP-based identity attribution, keyed HMAC-SHA256 integrity sealing with firmware baseline injection, and privacy-preserving GPS truncation. An experimental campaign on 10,000 synthetic records reduced the weighted Admissibility Gap deficit from 57.0% to 4.7% (an illustrative figure under the authors’ weight vector; because the transformation rules apply deterministically, this is a demonstration of sufficiency rather than an independent validation, and its direction is robust to the weighting), with outputs mapped to the Microsoft Sentinel Common Event Format schema and the BSI IT-Grundschutz OPS.1.1.5 logging requirements. Benchmarking on Raspberry Pi 4 hardware yielded a mean per-record latency of 1.574 ms at idle, demonstrating that audit-ready logging is achievable from the edge gateway inward on commodity hardware without hardware or firmware modification. The integrity and identity guarantees are enforced from the point of gateway ingestion; the device-to-gateway segment is treated as a declared trust boundary. Full article

Wildfires are a major environmental hazard with severe consequences for ecosystems, air quality, infrastructure, and public safety. The rising incidence and severity of wildfire events worldwide have increased the need for reliable early detection and monitoring systems. Remote sensing technologies, such as satellite and unmanned aerial vehicle (UAV) imagery, along with ground-based Closed-Circuit Television (CCTV) cameras, provide valuable geospatial data for large-scale wildfire monitoring. Recent advances in deep learning, particularly Convolutional Neural Networks (CNNs) and Transformer-based architectures, have significantly improved the accuracy of wildfire detection systems. Despite these advances, balancing local feature representation with global contextual modeling remains challenging. CNNs effectively capture local spatial features but have limited receptive fields, whereas Vision Transformers (ViTs) model long-range dependencies but often overlook fine-grained local details and require substantial computational resources. Consequently, accurately detecting small, occluded, and visually ambiguous fire regions remains difficult, particularly for real-time deployment on resource-constrained edge devices. To address these challenges, this study proposes EdgeNeXt-Attn, an enhanced EdgeNeXt-based framework that effectively integrates local feature learning and global contextual modeling through channel and spatial attention mechanisms. The proposed model improves the detection of small, occluded, and visually ambiguous fire regions while maintaining the computational efficiency required for real-time edge deployment. The proposed framework is evaluated on four multi-platform benchmarks spanning ground-based CCTV (DFAN, Complex-Fire), aerial drone (FLAME), and mixed drone–satellite (ADSF) imagery, achieving 92.09%, 95.16%, 96.65%, and 87.81% accuracy, respectively, and outperforming recent state-of-the-art baselines. With only 5.3M parameters, the model achieves real-time inference at 85.9, 27.3, and 8.4 FPS on GPU, CPU, and Raspberry Pi, respectively. Furthermore, ablation studies and Grad-CAM analysis validate its effectiveness and accurate fire localization. These results demonstrate an accurate and computationally efficient framework for real-time wildfire monitoring using multi-platform remote sensing and ground-based imaging systems. Full article

Resource-constrained edge devices require efficient and adaptive health monitoring to ensure reliable operation under dynamic workloads. This paper proposes an LLM-driven, context-aware framework that integrates real-time system metrics, historical data, and task-specific importance levels for anomaly detection and prediction. Specifically, the framework forecasts the semantic health state of the edge device five minutes ahead based on recent monitoring observations, enabling proactive intervention before critical operating conditions are reached. Multidimensional metrics, including CPU, memory, temperature, network load, and process information, are transformed into structured time-series representations and used as input to pre-trained machine learning models. A large language model (LLM) acts as an orchestration layer, dynamically selecting the most appropriate predictive model based on system context and resource constraints. This enables adaptive trade-offs between accuracy, latency, and computational cost. Experimental results on Raspberry Pi devices show that the proposed approach achieves comparable or improved performance while reducing resource usage compared to static methods. Full article