Competitive and high-pressure gaming conditions increasingly drive players to engage in various negative practices to gain unfair advantage over other participants. Among the least studied are smurfing and boosting, which involve unfair leveraging the experience of more skilled players while concealing it. These practices cause mental and material harm to other players and result in financial losses for the gaming industry. Currently, this issue remains unresolved, although existing research results on player skill identification show significant potential for using sensors to recognize specific player patterns. This study introduces GUARD, a Gameplay-aware inpUt-based cheAteR Detection framework designed to detect and prevent dishonest behavior in the tactical first-person shooter Counter-Strike 2. To develop GUARD, we consider all its stages in detail, carefully addressing the acquisition and pre-processing of multimodal data, the extraction of meaningful features based on player’s in-game actions, and the prediction and verification of game proficiency. Extensive expert assistance is applied to categorize the gaming audience into distinct skill-based groups and characterize them. To thoroughly describe players’ behavior we utilize kinematic concepts on directional mouse cursor movements, along with expert game context-related knowledge. Using machine learning techniques for multi-class classification with subsequent verification process, GUARD correctly predicts a player’s game skill group with an F1-Score of 84 %, while misclassifying an honest player as a cheater with less than 5 %. Being consistent with the natural competitive game conditions by design and demonstrating effective detection capabilities, the developed framework has the potential to be applied to existing gaming platforms to promote fair play and protect the interests and well-being of players.

In this article, we report on an experimentally identified feasibility framework for Bio Fuel Cell (BFC) powered Internet of Things (IoT) wireless sensors. Within the sensor each block (power source, harvester, storage, load) is identified from measurements on the same testbed. The storage model uses two RC branches with a clipped quadratic voltage-dependent leakage and reaches a best-case RMSE of 22 mV (48 mV mean, 71 mV worst across the four initial voltages). The per-event voltage budget decomposes into three competing time-integrals and a hybrid continuous–discrete simulator iterates the periodic orbit on the reduced sampled-voltage map with an experimentally measured local contraction modulus ρ ≈ 0.81, which supports a local Lyapunov certificate near the operating point. End-to-end validation on a 2S4P soil-coupled BFC array reproduces every per-schedule verdict in hardware. Wireless transmitters (ESP32 Wi-Fi/BLE) on the same source crash through Vmin at one transmission. The framework is calibrated on a single 2S4P stack.

Autonomous aerospace systems require architectures that balance deterministic real-time control with advanced perception capabilities. This paper presents an integrated system combining NASA's F' flight software framework with ROS2 middleware via Protocol Buffers bridging. We evaluate the architecture through a 32.25-minute indoor quadrotor flight test using vision-based navigation. The vision system achieved 87.19 Hz position estimation with 99.90% data continuity and 11.47 ms mean latency, validating real-time performance requirements. All 15 ground commands executed successfully with 100

Parkinson's disease (PD) is a debilitating condition that causes loss of physical activity, tremors, and coordination issues. Current studies use complex experimental testbeds to acquire data for PD diagnosis, and performance still requires improvement. This research aims to differentiate between healthy individuals and those in the early stages of PD using deep learning and a left-wrist-worn sensor with an embedded tri-axis Inertial Measurement Unit sensor. Data were collected from 40 patients while they performed nine daily exercises. The acceleration and gyroscope signals were preprocessed, segmented, and transformed into images using Time Series Imaging (TSI) algorithms, including Gramian Angular Difference/Summation Fields (GADF/GASF), Recurrence plots, Markov Transition Fields, and Continuous Wavelet Transform. We report on a PD assessment system based on the equal-weighted RGB images from accelerometer/gyroscope axes and transfer learning. GADF and GASF algorithms demonstrate superior performance, particularly with ResNet, achieving an average F1-score of 0.986 for HC vs. PwPD1 classification, with perfect detection (F1=1.0) in functional tasks like arm-holding and glass-filling. For PD staging (PwPD1 vs. PwPD1.5), the system maintained strong accuracy (F1=0.878), reaching F1=1.0 in dynamic exercises. Specifically, the key exercises including arm-holding and fist-clenching provided well-calibrated predictions Expected Calibration Error (ECE) <= 0.045, Brier Score <= 0.031) for both the detection and staging, ensuring high confidence in the model's outputs. Multiclass tasks yielded F1-scores in the range 0.90-0.91, confirming TSI's potential for scalable, clinical-grade PD monitoring.

Parkinson's disease (PD) is a progressive neurodegenerative disorder in which early motor symptoms are subtle, heterogeneous, and often unilateral, making objective detection and staging challenging. This study proposes a minimally intrusive framework based on bilateral wrist-worn inertial measurement units (IMUs) for early PD screening. Data were collected from healthy controls and PD patients (21-24 subjects per exercise) performing three standardized upperlimb clinical exercises: pronation-supination, finger-nosetapping, and hold the outstretched arms position. Multi-domain kinematic features were extracted and evaluated using nested leave-one-subject-out cross-validation (LOSO-CV). For HC vs. PD1 discrimination, finger-nose-tapping combined with a Support Vector Machine achieved the highest subject-level AUC of 0.972 [0.889, 1.000] with robust calibration (Brier $=0.062$ [0.004, 0.146]). PD1 vs. PD2 classification remained limited, reflecting intrinsic symptom overlap. These findings support wrist-worn IMUs for scalable early PD screening while underscoring the need for improved fine-stage modeling.

Parkinson’s disease (PD) is a disorder that affects movement and worsens over time. Common PD symptoms include tremor and difficulty with coordination. Typical sensor-based systems for PD detection rely on fixed sensor placements, limited subsensor inputs, or narrow task sets. Unlike the state-of-the-art that relies on static sensor configurations, in this study, we assess how sensor placement (wrist vs. dorsum), modality, exercise selection, and data processing affect the upper-limb detection of PD under the Leave-One-Subject-Out validation. Data from 55 participants (34 with PD) performing 11 standardized tasks were processed with a 0.2-13 Hz and 0.2-20 Hz band-pass filtering and a 279-feature multimodal representation (per one Inertial Measurement Unit (IMU) sensor). Statistical comparisons employed the Friedman/Nemenyi tests for paired, unbiased evaluation. Dorsum sensors (notably the left one) surpassed wrists. Fusion of accelerometer (A), gyroscope (G), and magnetometer (M) data outperformed single modalities. The posture-holding exercise, hand-opening/closing exercise, and chest-height postural exercise produced the most robust PD-specific signatures, and performance remained stable across the bandwidths. The best configurations obtained ROC of 0.952 and F1-binary of 0.957 with a low Brier of 0.042, demonstrating strong discrimination and reliable calibration. Sensors-only and Sensors+Age models were statistically indistinguishable, suggesting that the discriminative power primarily arises from movement-derived features rather than demographic differences. Overall, the results identify the minimal yet highly effective sensor–exercise combinations suitable for routine clinical or remote monitoring and are validated on this dataset, for practical, interpretable PD assessment.

Parkinson’s disease (PD) is a slowly progressive neurodegenerative disease which still lacks objective tools for diagnosis. According to recent research results, misdiagnosis of PD may reach up to 25%. In this article, we report on the medical decision support system based on wearable sensors and video cameras with consequent “multimodal” data analysis using Machine Learning (ML) methods. For data collection reasons 169 subjects performed eleven exercises recommended by the neurologists. The proposed smart system is assessed through ML metrics and outperform the state-of-the-art solutions by achieving precision 98.6%, recall 98.1%, and F1-micro 98.3%. This decision support system opens wide vista for its application in hospitals as well as at home settings for controlling the undergoing therapy.

In this work, we aim at realizing an eye tracker functionality on an off-the-shelf low cost web-camera. In particular, we address the problem of appearance gaze estimation employing deep learning methods for solving the regression task of predicting the point of gaze on a display. This solution is validated using a commercial infrared eye tracker: we achieve 73.843 Root Mean Square Error (RMSE) in pixels between the predictions. The proposed solution outperforms a relevant web-cam solution in terms of pixel-normalized RMSE 0.022 against 0.116 within the carried out comparative study, respectively.

Parkinson’s disease (PD) ranks as the second most prevalent neurodegenerative disorder, leading to motor complications, particularly in the elderly. Among its many symptoms, Freezing of Gait (FoG) affects nearly 50% of PD patients globally, leading to an increased risk of falls. Despite the widespread use of uni-modal FoG detection, it does not produce satisfactory results. This work aims to detect FoG using a multi-modal feature set. A publicly available multi-modal dataset was used, comprising data collected from 12 patients using various sensors, including Electroencephalogram (EEG), Electromyography (EMG), tri-axial ACC (accelerometers and gyroscopes), and Skin Conductance (SC) sensors. First, the data were pre-processed and segmented into 3-second windows with 90% overlap. The segmented windows were labeled based on the various percentage of FoG points (PFG) thresholds. Then, time-, frequency-, and time-frequency-based features were extracted and fed to the Support Vector Machine (SVM) model. These features included statistical, wavelet transform-based Shannon entropy, and Fast Fourier transform-based features. The results demonstrated that the multi-modal combination EMG + ACC, with a PFG of 0.75, outperformed all other sensor combinations, achieving the f1 score (binary) of 98.82%, highlighting the significant impact of ACC on FoG detection. It surpasses the state-of-art study, which reported an f1 score of 93.49%. Additionally, the combination of the right shank, waist, and arm within the acceleration sensors yielded the best f1 score of 99.19%. Furthermore, this work shows the power of synergistic concatenation of uni-modal data to enhance the model performance for FoG detection tasks.

Retinal vascular segmentation, a widely researched topic in biomedical image processing, aims to reduce the workload of ophthalmologists in treating and detecting retinal disorders. Segmenting retinal vessels presents unique challenges; previous techniques often failed to effectively segment branches and microvascular structures. Recent neural network approaches struggle to balance local and global properties and frequently miss tiny end vessels, hindering the achievement of desired results. To address these issues in retinal vessel segmentation, we propose a comprehensive micro-vessel extraction mechanism based on an encoder- decoder neural network architecture. This network includes residual, encoder booster, bottleneck enhancement, squeeze, and excitation building blocks. These components synergistically enhance feature extraction and improve the prediction accuracy of the segmentation map. Our solution has been evaluated using the DRIVE, CHASE-DB1, and STARE datasets, yielding competitive results compared to previous studies. The AUC and accuracy on the DRIVE dataset are 0.9884 and 0.9702, respectively. For the CHASE-DB1 dataset, these scores are 0.9903 and 0.9755, respectively, and for the STARE dataset, they are 0.9916 and 0.9750. Given its accurate and robust performance, the proposed approach is a solid candidate for being implemented in real-life diagnostic centers and aiding ophthalmologists.

Computer vision systems have been integrated into facilities dealing with the sorting of household waste. This solution allows for the sorting efficiency improvement and cost reduction. However, challenges associated with the poor annotation quality of existing waste segmentation datasets, unsuitable environment for recognition on a conveyor belt, or limited data for creating an effective and cost-efficient sorting system using visible range cameras significantly limit the application efficiency of computer vision systems. In this article, we report on the data-centric pipeline for enhancing the precision of predictions in multiclass household waste segmentation on a conveyor belt. In particular, we have demonstrated that by employing a pseudo-annotation approach combined with an object-based data augmentation algorithm, it is possible to train a model on a set of 'simple' images and achieve satisfactory results when estimating the model on a set of 'complex' images. We collected and prepared the dataset consisting of 5 k manually labeled data and additionally 10 k pseudo-labeled data by object-based augmentation. The proposed pipeline incorporates data balancing, transfer learning, and pseudo-labeling to improve the mean Average Precision (mAP) of the YOLOV8 segmentation model from 67 % to 83 % for 'simple' use case scenarios and from 42 % to 59 % or 'complex' industrial solutions.

Traditional component-based frameworks (CBFs), such as NASA’s F Prime, promote modularity and reuse but lack native support for formal verification, limiting their applicability in safety-critical and real-time domains. Ensuring deterministic behavior and correct synchronization remains a major challenge, especially on multicore embedded platforms. To address this gap, we propose a novel integration of the Behavior-Interaction-Priority (BIP) framework with F Prime, enabling correctness-by-construction through formal modeling and verification. Our methodology translates F Prime components into BIP models, allowing the use of priority rules and interaction semantics to detect race conditions and enforce synchronization determinism. We evaluate this hybrid framework on a Raspberry Pi 3 Model B platform, under both default Linux and PREEMPT_RT kernels, comparing two execution configurations: SYNC (synchronized mode updates via priority messages) and UNSYNC (sequential thread-level updates). Experimental results reveal that SYNC mode, when coupled with real-time kernel support, significantly reduces execution time variability and latency spikes, demonstrating improved synchronization and predictability. This integration enhances the reliability of modular embedded software while offering a practical pathway to adopt formal methods in industry-grade systems.

Parkinson's disease (PD) is a disorder that affects the central nervous system and causes severe motor and nonmotor problems. PD is currently cureless, but early diagnosis and proper therapy can slow down its progression. The latest research employs sophisticated experimental setups for PD diagnosis and enables the binary classification of subjects as healthy/PD. However, most are only suitable for clinical settings with limited monitoring capabilities, and their performance still needs improvement. This work detects PD from healthy control (HC) subjects while enabling multiclassification of PD stages using wrist-worn sensors on both hands. It also monitors PD progression. We gathered inertial measurement unit (IMU) data from 85 subjects at a hospital, 55 of whom were diagnosed with various stages of PD using the Hoehn-Yahr scale while performing 11 exercises under the neurologists' supervision. We preprocess the data by downsampling and bandpass filtering. Next, we segment the signals into overlapping windows and extract time- and frequency-domain features to input into various machine learning (ML) algorithms. We investigate the impact of the proposed exercises and sensor combinations via ensemble models and transfer learning. The score-based ensemble method achieves an f1-micro of 0.938 for HC/PD using the left-hand sensor, while both sensors can be used to get f1-micro 0.867 for multiclassification tasks. We also found that the left-hand sensor performs better than the right-hand sensor.

Sensing technologies are widely used in the solid waste automation for ensuring the waste detection at a conveyor belt and its consequent separation. However, the state-of-the-art sensing methods face the problem of continuous growth of waste volume, as well as its diversity, which implies severe performance requirements. In this article, we report on the intelligent sensing system based on the networked optical and hyperspectral cameras combined with the widely used machine learning methods to solve the problem of object segmentation. The goal of this research work is to amalgamate different data modalities to improve the accuracy of waste identification in a real-life scenario. This research is centered on a multimodal object segmentation framework. It encompasses the collection of a dataset with distinct training and testing partitions, addressing segmentation challenges and label removal from the optical data. In addition, this study tackles the segmentation issue for hyperspectral data where the hyperspectral models use optical predictions. The proposed approach facilitates the precise classification of various waste objects. The dataset comprises 12 object categories, including the assorted plastics, films, textiles, tetrapacks, and cellulosic materials. The proposed solution is validated on an industrial-scale conveyor belt operating at the speed of 3 m/s enabling the real-life conditions. The experimental results demonstrate an F1 -score of 70% for optical model and intersection over union (IoU) metric of 92% for the multimodal pipeline.

Computer vision methods have been recently integrated into industrial facilities performing the automatic household waste sorting. In this research, we report on an approach based on hyperspectral cameras and consecutive data analysis in recycling plants for improve waste sorting processes in terms of waste detection and classification. The majority of existing methods have been conducted on static conveyor belt, and their findings may not be applicable to recycling systems. The study focuses on developing a real-time hyperspectral object classification system for sorting machines, addressing the challenge of differentiating objects that are unidentifiable to RGB cameras. The linear hyperspectral camera operates within the wavelength range of 900 nm to 1800 nm. The research includes the creation of a dataset with physically separated training and test data, automated pseudo-labeling of hyperspectral data. This allows for rapid expansion to new classes without waste of manual annotation. The dataset contains 18 classes of objects, including various types of plastics, films, textiles, tetrapack, and cellulose products. To improve the accuracy and inference speed, we experiment with normalization, dimensionality reduction, and model calibration. Experimental results show a promising F1-score of 71% on a moving conveyor belt.

Modern Computer Vision (CV) methods heavily rely on neural networks. Models based on convolutional and transformer blocks achieve state-of the-art performance on challenging tasks, including classification, semantic and instance segmentation, object and visual relationship detection, monocular depth estimation, and image reconstruction [He et al., 2016, Dosovitskiy et al., 2020]. One reason behind the rapid growth of neural-based solutions is in the advance of computational resources over last few decades. It allows for training larger models which are capable of recognising more complex patterns [Brown et al., 2020]. Another essential underlying condition is in the accessibility of training data from which the patterns can be retrieved [Birhane and Prabhu, ]. Figure 5.1 shows the typical relation between accuracy and the amount of available training data.

Crying is an inevitable character trait that occurs throughout the growth of infants, under conditions where the caregiver may have difficulty interpreting the underlying cause of the cry. Crying can be treated as an audio signal that carries a message about the infant's state, such as discomfort, hunger, and sickness. The primary infant caregiver requires traditional ways of understanding these feelings. Failing to understand them correctly can cause severe problems. Several methods attempt to solve this problem; however, proper audio feature representation and classifiers are necessary for better results. This study uses time-, frequency-, and time-frequency-domain feature representations to gain in-depth information from the data. The time-domain features include zero-crossing rate (ZCR) and root mean square (RMS), the frequency-domain feature includes the Mel-spectrogram, and the time-frequency-domain feature includes Mel-frequency cepstral coefficients (MFCCs). Moreover, time-series imaging algorithms are applied to transform 20 MFCC features into images using different algorithms: Gramian angular difference fields, Gramian angular summation fields, Markov transition fields, recurrence plots, and RGB GAF. Then, these features are provided to different machine learning classifiers, such as decision tree, random forest, K nearest neighbors, and bagging. The use of MFCCs, ZCR, and RMS as features achieved high performance, outperforming state of the art (SOTA). Optimal parameters are found via the grid search method using 10-fold cross-validation. Our MFCC-based random forest (RF) classifier approach achieved an accuracy of 96.39%, outperforming SOTA, the scalogram-based shuffleNet classifier, which had an accuracy of 95.17%.

Parkinson's disease (PD) causes physical activity loss, tremors, stiffness, and coordination-related issues. The state-of-the-art studies utilize complex experimental testbeds for data acquisition and consequent application of Machine learning (ML) methods for the PD diagnosis. However, their performance is still the subject to improve. This research aims to distinguish PD from healthy control (HC) using four wearable sensors and ML, where two sensors are placed on each hand, wrist, and dorsum. The dataset of 54 patients, 21 with HC and 33 with PD, was collected in a hospital while the patient performed 11 exercises under the supervision of neurologists. The dataset was preprocessed and segmented into overlapping windows, and ML algorithms were used in terms of analysis of frequency ranges, features, and exercises. This study demonstrates a simple and accurate method for detecting PD in clinical or home settings with an average of 0.936 $f 1_{{micro }}$ and 0.935 ROC while the performance of ‘best’ exercises achieves 100% ${f}1_{micro}$ .

In this work, we address the cheating problem in video games and provide instruments for reliable detection of various cheaters based on their in-game behavior in the most popular FPS game Counter-Strike: Global Offensive. For this purpose, we collect more than 2,243 competitive game records through official Valve servers. This corresponds to more than 14,000 players with various gaming skill levels. Thorough application of mathematical processing methods, as well as the eSports expert assistance, allows us to develop meaningful and accurate metrics for measuring in-game player actions. We report that with their help cheater detection has high accuracy of 85%, and outperform VAC, VACNet, and OverWatch in a comparative analysis. The tools are additionally tested on the professional players and show a low false positive rate.

In this work, we present a method for automatically generating a synthetic, labelled computer vision dataset of game-character positions for LoL and the application of this synthetic data for training a deep-learning, single-shot object detection model that can then be applied in pseudo labelling and generating detailed player position data from LoL video streams. Further-more, we investigate the effects of data augmentation and class balancing on the overall performance of the neural network, comparing them to the base network trained on a manually labelled dataset. Using synthetically generated datasets of 10000 images containing 21 object classes, we achieved a mean average precision (mAP@50) of 64.8% and an IoU of 67.7%.