Human Activity Recognition (HAR) using radar signals has gained significant attention due to its non-intrusive nature and robustness in various environments. However, the impact of radar signal preprocessing techniques on the performance of deep learning (DL) models remains an active area of research. This study investigates how different radar domain representations affect HAR accuracy by evaluating four preprocessing methods: Time-Range (TR) maps generated via Range-Fast Fourier Transform (FFT), Range-Doppler (RD) maps obtained through sequential FFTs, and Time-Doppler (TD) features extracted using Short Time Fourier Transform (STFT) and Smoothed Pseudo Wigner Ville Distribution (SPWVD). We employ a baseline Convolutional Neural Network (CNN) and state-of-the-art Transfer Learning (TL) models to assess whether advanced preprocessing or increased model complexity yields greater performance gains. The results reveal that high-resolution TD analysis using SPWVD does not significantly enhance classification performance and incurs substantial computational overhead, limiting its real-time applicability. Conversely, the TR representation offers computational efficiency but struggles to classify complex activities with the baseline CNN accurately. RD and STFT methods provide a favorable balance between classification accuracy and computational efficiency. Notably, transitioning from the baseline CNN to TL models leads to substantial improvements in recognition accuracy: up to 29.36% for TR, 21.42% for RD, 16.66% for STFT, and 11.11% for SPWVD representations. Overall, our findings demonstrate that TL models, when combined with computationally efficient radar preprocessing techniques like RD or STFT, significantly improve recognition accuracy and generalize well across datasets, as confirmed by evaluation on two publicly available radar-based HAR datasets. Among these, the RD representation combined with VGG-19 yielded the best trade-off between accuracy and latency, achieving a total processing time of 0.91 s per sample for a 10 s activity duration, making it highly suitable for latency-sensitive HAR applications.

Federated learning (FL) has gained prominence in energy forecasting applications. Despite its advantages, FL remains vulnerable to adversarial attacks that threaten the reliability of predictive models. This study introduces a stealth attack, Federated Communication Round Attack (Fed-CRA), which increases communication rounds without affecting forecasting accuracy. Increased communication rounds can delay decisionmaking, reducing system responsiveness and cost-effectiveness in dynamic energy forecasting scenarios. Experimental validation on two datasets demonstrated that Fed-CRA increased communication rounds by 574% (from 72 to 485) in the AEP dataset and by 237% (from 92 to 310) in the COMED dataset. This led to a corresponding rise in energy consumption by 573% (from 41.04 kWh to 276.35 kWh) and 237% (from 52.44 kWh to 176.65 kWh), respectively, while preserving forecasting accuracy. To counter this attack, we proposed Federated Incentivized Averaging (Fed-InA), a game theory-inspired framework that rewards honest clients and penalizes dishonest ones based on their contributions. Results showed that Fed-InA reduced the additional communication rounds caused by Fed-CRA by 85% in the AEP dataset and 70% in the COMED dataset, while maintaining forecasting performance. Fed-InA achieves resource efficiency comparable to Federated Averaging (FedAvg) and demonstrates robustness in handling non-IID data.

Non-verbal communication is a crucial element of human interaction, serving as a powerful tool for expressing emotions, establishing connections, and understanding others beyond verbal language. The recognition of non-verbal cues has held significant importance in recent years, particularly for individuals with disabilities like deafness or muteness. This recognition plays a crucial role in enabling effective communication for these groups. However, existing camera-based systems for detecting non-verbal cues have drawbacks, including privacy concerns, difficulties in varying lighting conditions, the need for complex training, and operational range limitations. In this study, we employ contactless sensing technology to detect non-verbal cues such as Good Idea, Thinking, Worried, Normal, Shocked, and OK. This proof-of-concept study demonstrates the efficacy of this privacy-preserving system. The data was first transformed into spectrograms and then processed using deep learning models like ResNet50 and VGG16, achieving remarkable classification accuracy, notably 95.83% using ResNet50.

Head movement holds significant importance in con-veying body language, expressing specific gestures, and reflecting emotional and character aspects. The detection of head movement in smart or assistive driving applications can play an important role in preventing major accidents and potentially saving lives. Additionally, it aids in identifying driver fatigue, a significant contributor to deadly road accidents worldwide. However, most existing head movement detection systems rely on cameras, which raise privacy concerns, face challenges with lighting conditions, and require complex training with long video sequences. This novel privacy-preserving system utilizes UWB-radar technology and leverages Deep Learning (DL) techniques to address the mentioned issues. The system focuses on classifying the five most common head gestures: Head 45L (HL45), Head 45R (HR45), Head 90L (HL90), Head 90R (HR90), and Head Down (HD). By processing the recorded data as spectrograms and leveraging the advanced DL model VGG16, the proposed system accurately detects these head gestures, achieving a maximum classification accuracy of 84.00% across all classes. This study presents a proof of concept for an effective and privacy-conscious approach to head position classification.

This paper presents a simplified automatic modulation classification system based on stepwise regression and Hierarchical Polynomial Classifier. The proposed system applies High Order Cumulants of the received signal to a polynomial network in order to calculate their second order polynomial expansion. Then, stepwise regression model is used as a feature selection system; whereas important features are identified and used for the classification problem and the insignificant features are discarded. With this optimization step, the complexity of the classifier is reduced significantly while maintaining the same classification accuracy as the original Hierarchical Polynomial Classifier with all the expanded features.

In this paper, we propose a glove-based Arabic sign language recognition system using a novel technique for sequential data classification. We compile a sensor-based dataset of 40 sentences using an 80-word lexicon. In the dataset, hand movements are captured using two DG5-VHand data gloves. Data labeling is performed using a camera to synchronize hand movements with their corresponding sign language words. Low-complexity preprocessing and feature extraction techniques are applied to capture and emphasize the temporal dependence of the data. Subsequently, a Modified k-Nearest Neighbor (MKNN) approach is used for classification. The proposed MKNN makes use of the context of feature vectors for the purpose of accurate classification. The proposed solution achieved a sentence recognition rate of 98.9%. The results are compared against an existing vision-based approach that uses the same set of sentences. The proposed solution is superior in terms of classification rates while eliminating restrictions of vision-based systems.

In this paper, a new automatic modulation classification system using a polynomial classifier is proposed. The feature set used in the proposed system is comprised of second, fourth and sixth higher order cumulants of the received signal. The polynomial classifier expands the feature vector into a higher dimensional space in which the different modulation types are linearly separable. The performance of the proposed system is tested against BPSK, QPSK, 8-PSK, 16-QAM and 64-QAM constellations. Results showed a relatively higher classification rate for all types of modulations compared to previously published results using fourth and sixth order cumulants with other classification schemes. Moreover, the proposed classification scheme has low computational complexity and is more suited for real time applications.

With the advent of cheap and readily available body-worn sensors, new electronic pedagogical approaches are emerging. This paper describes the design and development of a prototype electronic training system for the game of Cricket. The system is designed to provide real-time feedback to players and coaches to help improve technique and performance in a game. The system uses a watch with a built-in accelerometer that transmits real-time 3-axis acceleration data from a player using a proprietary wireless protocol. Machine learning techniques are used to analyze this data to provide real-time critique to the player. Player's performance data is also recorded in a Learning Management System (LMS) and a coach or the player can view various reports related to their past performance. The system has been implemented using Dynamic Time Warping (DTW) and shows over 93% accuracy.

In this paper, we propose a novel biometric based on utilizing microwave images constructed from measuring a voltage that is related to the properties of the reflected microwave signal. The proposed technique requires collecting scans of the users' hands using an open-ended rectangular waveguide operating at a frequency of 9.6 GHz in the near-field region. The microwave scan reflects the dielectric properties of the hand affected by the thickness, shape and structure of the hand at every point in the scan. The proposed technique is safe as it uses low power microwave signals in the range of 10-20 dBm. An in-house microwave hand-image database is collected using a microwave scanning system. The database is comprised of 370 images collected from 37 users. Radon and DCT transformations are used to extract distinguishing features from microwave images. Classification is done with polynomial classifiers. In the context of the polynomial classifier, a variety of dimensionality reduction schemes are examined. Amongst these schemes, spectral regression is found to give the best results in terms of accuracy, dimensionality of feature space and classifier stability. A user identification rate of 94.7% was achieved with a feature vector size of 36 elements only.

Base-calling is one of many problems that can be solved using pattern recognition, the act of classifying raw data based on prior or statistical information extracted from the data into various classes. In this paper, we propose a new framework using polynomial classifiers to model electropherogram traces obtained from ABI sequencing machines to perform base-calling. Initially, pre-processing, which includes segmented normalization and peak sharpening, needs to be performed to reduce the imperfections caused in a trace as a result of the chemistry involved. Discriminative feature vectors are then extracted from the chromatogram traces and are expanded to a higher dimensional space by second order polynomial expansion. A linear classifier is then trained and bases are classified respectively. Chromatogram traces that were chosen for analysis belong to Homo sapiens, Saccharomyces mikatae and Drosophila melanogaster. Simulation results indicated an accuracy of up to 99.2% upon testing three different chromatogram traces consisting of about 600 to 800 bases each. The proposed model's performance was compared to the existing standards: ABI and PHRED in terms of insertion, deletion and substitution errors. Simulation evidence indicated that the designed model performs comparably or slightly better than ABI in terms of deletion and insertion errors. Moreover, polynomial classifier resulted in negligible substitution errors compared to ABI. Polynomial classifier was also observed to perform comparable to PHRED in terms of deletion error and substitution errors. The results obtained demonstrate the potential of this model to perform base-calling.

Cognitive radio (CR) is a promising technology for improving the utilization of the scarce radio spectrum by allowing secondary users to regularly sense the spectrum and opportunistically access the under-utilized frequency bands. However, spectrum sensing in CR environment is a challenging task due to varying radio channel conditions and might lead to interference with licensed users. In this paper, we propose a new framework for CR spectrum sensing based on linear and polynomial classifiers. A cooperative CR network is considered in this paper with CR nodes collaborating in making the decision about spectrum availability. Simulation results indicate that both polynomial and linear classifiers provide high detection rate of primary users with a constant false alarm rate at very small signal to noise ratio conditions. For instance, the proposed techniques can achieve above 90% detection probability at Eb/N0= -7dB with observation window of 50 bits and 10% false alarm rate. It is also shown that the performance improves as we increase the sensing time for both schemes.

This paper proposes a two tier feature extraction approach for the recognition of video-based isolated Arabic sign language gestures. In the first tier, the prediction error of the image sequence is binarized and collapsed into two unidirectional accumulated differences images. In the second tier of feature extractions, two approaches are applied to the accumulated differences images: frequency domain transformation, and radon transformation. We apply such feature extractions on each of the accumulated differences images and then concatenate the resultant feature vectors. Alternatively, the accumulated differences images are concatenated prior to the second tier of feature extractions. The paper reports on the classification results of both solutions using Fisher's linear discriminants. Comparisons with existing work reveal that up to 39% of the misclassifications have been corrected.

This paper presents various spatio-temporal feature-extraction techniques with applications to online and offline recognitions of isolated Arabic Sign Language gestures. The temporal features of a video-based gesture are extracted through forward, backward, and bidirectional predictions. The prediction errors are thresholded and accumulated into one image that represents the motion of the sequence. The motion representation is then followed by spatial-domain feature extractions. As such, the temporal dependencies are eliminated and the whole video sequence is represented by a few coefficients. The linear separability of the extracted features is assessed, and its suitability for both parametric and nonparametric classification techniques is elaborated upon. The proposed feature-extraction scheme was complemented by simple classification techniques, namely, K nearest neighbor (KNN) and Bayesian, i.e., likelihood ratio, classifiers. Experimental results showed classification performance ranging from 97% to 100% recognition rates. To validate our proposed technique, we have conducted a series of experiments using the classical way of classifying data with temporal dependencies, namely, hidden Markov models (HMMs). Experimental results revealed that the proposed feature-extraction scheme combined with simple KNN or Bayesian classification yields comparable results to the classical HMM-based scheme. Moreover, since the proposed scheme compresses the motion information of an image sequence into a single image, it allows for using simple classification techniques where the temporal dimension is eliminated. This is actually advantageous for both computational and storage requirements of the classifier.

We apply polynomial discriminant function classifiers for isolated-word speaker-independent Arabic digit recognition. The performance of the polynomial classifier is evaluated for different implementations. We also provide a performance comparison between the polynomial classifier and dynamic time warping (DTW). The polynomial classifier is found to outperform DTW in many aspects such as recognition rate, and computational and memory requirements.

We propose a new approach for classifying speech vs. non-speech, which proves to significantly improve speech recognition performance under noise. The proposed algorithm relies on the energy and spectral characteristics of the signal and applies a 3-level two-dimensional thresholding to determine whether an input frame is speech or non-speech. The algorithm runs in real-time, and offers better immunity to background noise, and to background speech than traditional energy-based word boundary detection. The performance of the endpoint detector is reported here in terms of improvements in speaker-independent (SI) and speaker-dependent (SD) recognition performance using 5 different simulated noise conditions and various signal-to-noise ratios (SNR). The proposed endpoint detection of speech improves the SD recognition accuracy by 24% for office noise, and reduces the false rejection rates for both SI and SD by 45% for babble noise and lobby noise.

Modern speaker recognition applications require high accuracy at low complexity. We propose the use of a polynomial-based classifier to achieve these objectives. This approach has several advantages. First, polynomial classifier scoring yields a system which is highly computationally scalable with the number of speakers. Second, a new training algorithm is proposed which is discriminative, handles large data sets, and has low memory usage. Third, the output of the polynomial classifier is easily incorporated into a statistical framework allowing it to be combined with other techniques such as hidden Markov models. Results are given for the application of the new methods to the YOHO speaker recognition database.

Modern speaker verification applications require high accuracy at low complexity. We propose the use of a polynomial-based classifier to achieve this objective. We demonstrate a new combination of techniques which makes polynomial classification accurate and powerful for speaker verification. We show that discriminative training of polynomial classifiers can be performed on large data sets. A prior probability compensation method is detailed which increases accuracy and normalizes the output score range. Results are given for the application of the new methods to YOHO.

This paper proposes a methodology that uses automatic speaker recognition for evaluating the speaker recognizability of processed speech. This methodology provides an efficient measure (an objective one) that is hoped to replace subjective tests which involve human listeners. Preliminary experiments have been conducted using speech processed through the following coders: LPC-10 2400 bps (FS 1015), CELP 4800 bps (FS 1016), and PSELP 2400 pbs.

A new set of LP-derived features is introduced. The concept of these features is motivated by the power sum formulation of the LP cepstrum. Due to the fact that the LP model implies that the resulting poles are either real or occur in complex conjugate pairs, the power sum of the poles is equivalent to the power sum of their real components. Therefore, the LP cepstrum is associated to the power sum of the real component of the LP poles. This fact is utilized in deriving a new set of features that is associated to the imaginary components of the LP poles. The author refers to this new set of features as the sepstral coefficients. It is found that the sepstral coefficients and cepstral coefficients are relatively uncorrelated. Hence, they can be used jointly to improve the performance of pattern classification applications where cepstral features are usually used. The author presents some preliminary results on speaker identification experiments.

In this paper we introduce a new set of features that provides improved performance for speaker identification. This feature set is referred to as the adaptive component weighting (ACW) cepstral coefficients. The ACW scheme modifies the linear predictive (LP) spectral components (resonances) so as to emphasize the formant structure by attenuating the broad-bandwidth spectral components. Such components are found to introduce undesired variability in the LP spectra of speech signals due to environmental factors. The ACW cepstral coefficients represent an adaptively weighted version of the LP cepstrum. The adaptation results in deemphasizing the irrelevant variations of the LP cepstral coefficients on a frame-by-frame basis. Experiments are presented using the San Diego portion of the King database. The ACW cepstrum is shown to offer improved speaker identification performance as compared to other common methods of cepstral weighting.>

A new set of features is introduced that has been found to improve the performance of automatic speaker identification systems, The new set of features is referred to as the adaptive component weighting (ACW) cepstral coefficients. The new features emphasize the formant structure of the speech spectrum while attenuating the broad-bandwidth spectral components. The attenuated components correspond to the variations in spectral tilt of transmission and recording environment, and other characteristics that are irrelevant to speaker identification. The resulting ACW spectrum introduces zeros into the usual all-pole linear prediction (LP) spectrum. This is equivalent to applying a finite impulse response (FIR) filter that normalizes the narrow-band modes of the spectrum. Unlike existing fixed cepstral weighting schemes, the ACW cepstrum provides an adaptively weighted version of the LP cepstrum. The adaptation results in deemphasizing the irrelevant variations of the LP cepstral coefficients on a frame-by-frame basis. The ACW features are evaluated for text-independent speaker identification and are shown to yield improved performance.>

An evaluation of various classifiers for text-independent speaker recognition is presented. In addition, a new classifier is examined for this application. The new classifier is called the modified neural tree network (MNTN). The MNTN is a hierarchical classifier that combines the properties of decision trees and feedforward neural networks. The MNTN differs from the standard NTN in both the new learning rule used and the pruning criteria. The MNTN is evaluated for several speaker recognition experiments. These include closed- and open-set speaker identification and speaker verification. The database used is a subset of the TIMIT database consisting of 38 speakers from the same dialect region. The MNTN is compared with nearest neighbor classifiers, full-search, and tree-structured vector quantization (VQ) classifiers, multilayer perceptrons (MLPs), and decision trees. For closed-set speaker identification experiments, the full-search VQ classifier and MNTN demonstrate comparable performance. Both methods perform significantly better than the other classifiers for this task. The MNTN and full-search VQ classifiers are also compared for several speaker verification and open-set speaker-identification experiments. The MNTN is found to perform better than full-search VQ classifiers for both of these applications. In addition to matching or exceeding the performance of the VQ classifier for these applications, the MNTN also provides a logarithmic saving for retrieval.