Kazi Newaj Faisal

Natural and reliable hand gesture recognition is becoming a key enabler for intuitive human-computer interaction in internet of things (IoT) environments, especially for wearable and assistive systems where camera-based sensing is unsuitable. This paper presents an inertial measurement unit (IMU) based framework that converts raw tri-axial accelerometer and gyroscope signals into structured time-frequency images and classifies them using a lightweight convolutional network. First, we propose a stacking-based multivariate time-frequency representation method termed TFR-StackGrid. In this method, the smoothed pseudo Wigner-Ville distribution (SPWVD) is computed for each of the six IMU-MMG channels. The resulting time-frequency maps are then systematically arranged into a fixed 2×3 kinematic grid. This structured configuration preserves the spatial relationship between translational acceleration and rotational motion, thereby integrating multivariate information into a single stacked time-frequency representation, referred to as S-TFR. A compact Gesture Net architecture is then designed to operate directly on these images, combining stacked 2D convolution blocks with a shallow classification head for 11 dynamic hand gesture classes. We conduct a comprehensive evaluation on the large-scale hand gesture accelerometer and gyroscope Dataset (HGAG-DATA), containing 11 gestures, 23,650 samples from 43 participants, and compare seven different time-frequency schemes, including wavelet and other quadratic distributions. The proposed SPWVD-based TFR-StackGrid with Gesture Net achieves 96.21% micro accuracy and 96.28% macro accuracy, with balanced per-class precision, recall, and F1-scores, while relying only on a single wrist-worn IMU based mechanomyography signals. These results demonstrate that structured time-frequency imaging combined with a lightweight convolutional backbone can effectively mitigate the limitations of template matching, handcrafted feature pipelines, and heavier deep sequence models for practical IMU-MMG-based hand gesture recognition.

Driver drowsiness represents a significant risk to road safety, being a major contributor to traffic accidents. This study introduces hybrid adaptive temporal sequential network HatSeqNet, an effectively designed architecture for cross-subject driver drowsiness detection using single-channel electroencephalogram (EEG) signals across different subjects. Leveraging the Mexican hat wavelet transform, EEG signals are transformed into time-scale representations (TSRs), capturing intricate neural patterns and transient dynamics. These TSRs are processed by a hybrid convolutional neural network and recurrent neural network architecture, which effectively models spatial and temporal dependencies. The model is evaluated using leave-one-subject-out cross-validation on a publicly available EEG dataset from a sustained-attention driving task. HatSeqNet achieves a state-of-the-art average accuracy of 88.26%, with precision, recall, and F1-score of 88.87%, 87.63%, and 88.10%, respectively. These metrics surpass those of prior studies on the same SAT dataset. The employment of single-channel EEG enhances the system’s portability, rendering HatSeqNet suitable for practical vehicular applications. The superior performance of the proposed architecture underscores its potential to enhance road safety through reliable and non-invasive driver drowsiness detection, even without personalized calibration.

Radar-based human activity recognition (HAR) offers noncontact sensing, environmental robustness, and privacy preservation, but its performance is limited by the quality of nonstationary micro-Doppler signatures in frequency-modulated continuous-wave (FMCW) radar signals. Conventional time–frequency methods, such as short-time Fourier transform and Wigner–Ville distribution, suffer from spectral blurring and cross-term interference, reducing sensing resolution. This letter introduces an improved sensor-oriented time–frequency enhancement framework for FMCW radar HAR using the synchroextracting transform (SET), which selectively extracts energy along instantaneous frequency ridges to suppress some artifacts and enhance micro-Doppler concentration. Applied to clutter-suppressed slow-time signals, SET produces highly concentrated and interpretable time– frequency representations (TFRs), directly improving radar sensing fidelity. The proposed approach is evaluated on the NJUST FMCW radar dataset, where SET-derived TFRs show improved concentration and activity separability over conventional methods. Using a lightweight time–frequency anisotropic radar network, the framework achieves 98.71% accuracy and a 96.14% F1-score with low computational cost (0.119 s per sample), demonstrating that sensor-level time–frequency enhancement leads to measurable gains in FMCW radar HAR systems for improved sensor-based performance.

This paper presents a 2D extension of the successive variational mode decomposition (2D-SVMD) method, which decomposes 2D signals into band-limited intrinsic mode functions without requiring prior knowledge of the number of modes. The proposed 2D-SVMD addresses the limitations of existing 2D variational mode decomposition (2D-VMD) and 2D variational mode decomposition with total variation (2D-TV-VMD) methods, which need initialization of the number of modes and can produce redundant or mixed modes if the number is not accurately specified. 2D-SVMD decomposes in a sequential manner which enhances the convergence rate of the decomposition process and also mitigates the extraction of extraneous modes, collectively resulting in reduced computational overhead. The proposed 2D-SVMD is also evaluated on synthetic and real-world images and compared to 2D-VMD and 2D-TV-VMD, demonstrating its superior performance in achieving natural decomposition of 2D signals.

Early age identification and prompt intervention play a crucial role in mitigating the severity of neurodevelopmental disorders in children. Traditional diagnostic approaches can be lengthy, but there is growing research potential in using electroencephalogram (EEG) signals to detect attention deficit hyperactivity disorder (ADHD) and intellectual developmental disorder (IDD). By recording the electrical activity of the brain, EEG has emerged as a promising technique for the early identification of these disorders. This research proposes a novel integrated method for identifying multiple neurodevelopmental disorders from the EEG signals of children. The approach combines successive multivariate variational mode decomposition (SMVMD) for analyzing multi-component non-stationary signals and a machine learning (ML)-based classifier, addressing the issue of inconsistent numbers of extracted features by introducing an energy-based feature integration approach. By integrating enhanced features from SMVMD with a K-nearest Neighbor (KNN) classifier, the unified approach successfully detects two separate neurodevelopmental disorders from normal subjects. The proposed method demonstrates perfect classification scores in detecting IDD under three different scenarios and achieves 99.17% accuracy in classifying ADHD subjects from normal subjects. Evaluation against different ML-based classifiers confirms the effectiveness of the proposed feature extraction algorithm and highlights its superior performance compared to recent methods published on similar datasets.

The fluctuations in the stock market serve as an indicator of societal and economic progress. To enhance the forecasting of financial time series data, this paper introduces a novel hybrid model that integrates successive variational mode decomposition (SVMD) with a long short‐term memory (LSTM) network. SVMD is used as a data‐decomposition technique to facilitate the division of complex original series into its decomposed intrinsic mode functions in order to reduce nonstationarity. This approach has been analyzed for both synthetic signals and financial time series data. LSTM is utilized to analyze each signal and take into consideration the influence of historical data in order to enhance prediction accuracy. The final predicted results are derived by reconstructing each predicted time series. To evaluate the model, financial time‐series data from global stock market have been selected for analysis, i.e., the Hong Kong Hang Seng Index (HSI), Stock Exchange Sensitive Index (SENSEX), Standard and Poor’s 500 (S&P500), and North America West Texas Intermediate (WTI). The proposed model has shown improved prediction results based on the performance metrics when compared with single models and other existing hybrid models.