Traffic accidents are a major public safety challenge around the world and are often influenced by the cognitive and physiological states of drivers. Among the multiple in-vehicle factors, listening to music has shown complex effects on driver behavior, particularly in relation to music tempo. This study proposes TransBiNet, a novel deep learning architecture that integrates Transformer-based attention mechanisms with Bidirectional Long Short-Term Memory layers to classify driver states under different auditory conditions using internal biometric signals. Data were collected from 26 participants driving in a simulated environment, where each subject completed scenarios involving fast-tempo music, slow-tempo music, and no music. Physiological signals (heart rate, breathing rate, galvanic skin response, and skin temperature) and head motion data (gyroscope and accelerometer) were gathered via wearable sensors and used as input to the model. The architecture was optimized through Hyperband-based hyperparameter tuning and showed a test accuracy of 97.62% as well as strong precision and recall across all classes. The results showed that internal physiological and motion-based signals are sufficient for robust classification of music-induced driver states, supporting the potential for real-time, sensor-driven driver monitoring systems in intelligent transportation.

Human Activity Recognition (HAR) is one of the essential building blocks of so many applications like security, monitoring, the internet of things and human-robot interaction. The research community has developed various methodologies to detect human activity based on various input types. However, most of the research in the field has been focused on applications other than human-in-the-centre applications. This paper focused on optimising the input signals to maximise the HAR performance from wearable sensors. A model based on Convolutional Neural Networks (CNN) has been proposed and trained on different signal combinations of three Inertial Measurement Units (IMU) that exhibit the movements of the dominant hand, leg and chest of the subject. The results demonstrate k-fold cross-validation accuracy between 99.77 and 99.98% for signals with the modality of 12 or higher. The performance of lower dimension signals, except signals containing information from both chest and ankle, was far inferior, showing between 73 and 85% accuracy.

Human walking is a complex activity with a high level of cooperation and interaction between different systems in the body. Accurate detection of the phases of the gait in real-time is crucial to control lower-limb assistive devices like exoskeletons and prostheses. There are several ways to detect the walking gait phase, ranging from cameras and depth sensors to the sensors attached to the device itself or the human body. Electromyography (EMG) is one of the input methods that has captured lots of attention due to its precision and time delay between neuromuscular activity and muscle movement. This study proposes a few Machine Learning (ML) based models on lower-limb EMG data for human walking. The proposed models are based on Gaussian Naive Bayes (NB), Decision Tree (DT), Random Forest (RF), Linear Discriminant Analysis (LDA) and Deep Convolutional Neural Networks (DCNN). The traditional ML models are trained on hand-crafted features or their reduced components using Principal Component Analysis (PCA). On the contrary, the DCNN model utilises convolutional layers to extract features from raw data. The results show up to 75% average accuracy for traditional ML models and 79% for Deep Learning (DL) model. The highest achieved accuracy in 50 trials of the training DL model is 89.5%.

Human Activity Recognition (HAR) is one of the fundamental building blocks of human assistive devices like orthoses and exoskeletons. There are different approaches to HAR depending on the application. Numerous studies have been focused on improving them by optimising input data or classification algorithms. However, most of these studies have been focused on applications like security and monitoring, smart devices, the internet of things, etc. On the other hand, HAR can help adjust and control wearable assistive devices, yet there has not been enough research facilitating its implementation. In this study, we propose several models to predict four activities from inertial sensors located in the ankle area of a lower-leg assistive device user. This choice is because they do not need to be attached to the user’s skin and can be directly implemented inside the control unit of the device. The proposed models are based on Artificial Neural Networks and could achieve up to 92.8% average classification accuracy.

Human Activity Recognition (HAR) is a crucial technology for many applications such as smart homes, surveillance, human assistance and health care. This technology utilises pattern recognition and can contribute to the development of human-in-the-loop control of different systems such as orthoses and exoskeletons. The majority of reported studies use a small dataset collected from an experiment for a specific purpose. The downsides of this approach include: 1) it is hard to generalise the outcome to different people with different biomechanical characteristics and health conditions, and 2) it cannot be implemented in applications other than the original experiment. To address these deficiencies, the current study investigates using a publicly available dataset collected for pathology diagnosis purposes to train Machine Learning (ML) algorithms. A dataset containing knee motion of participants performing different exercises has been used to classify human activity. The algorithms used in this study are Gaussian Naive Bayes, Decision Tree, Random Forest, K-Nearest Neighbors Vote, Support Vector Machine and Gradient Boosting. Furthermore, two training approaches are compared to raw data (de-noised) and manually extracted features. The results show up to 0.94 performance of the Area Under the ROC Curve (AUC) metric for 11-fold cross-validation for Gradient Boosting algorithm using raw data. This outcome reflects the validity and potential use of the proposed approach for this type of dataset.