Human Activity Recognition (HAR) systems are devised for continuously observing human behavior - primarily in the fields of environmental compatibility, sports injury detection, senior care, rehabilitation, entertainment, and the surveillance in intelligent home settings. Inertial sensors, e.g., accelerometers, linear acceleration, and gyroscopes are frequently employed for this purpose, which are now compacted into smart devices, e.g., smartphones. Since the use of smartphones is so widespread now-a-days, activity data acquisition for the HAR systems is a pressing need. In this article, we have conducted the smartphone sensor-based raw data collection, namely H-Activity , using an Android-OS-based application for accelerometer, gyroscope, and linear acceleration. Furthermore, a hybrid deep learning model is proposed, coupling convolutional neural network and long-short term memory network (CNN-LSTM), empowered by the self-attention algorithm to enhance the predictive capabilities of the system. In addition to our collected dataset ( H-Activity ), the model has been evaluated with some benchmark datasets, e.g., MHEALTH, and UCI-HAR to demonstrate the comparative performance of our model. When compared to other models, the proposed model has an accuracy of 99.93% using our collected H-Activity data, and 98.76% and 93.11% using data from MHEALTH and UCI-HAR databases respectively, indicating its efficacy in recognizing human activity recognition. We hope that our developed model could be applicable in the clinical settings and collected data could be useful for further research.

Attention is the mental awareness of human on a particular object or a piece of information. The level of attention indicates how intense the focus is on an object or an instance. In this study, several types of human attention level have been observed. After introducing image segmentation and detection technique for facial features, eyeball movement and gaze estimation were measured. Eye movement were assessed using the video data, and a total of 10197 data instances were manually labelled for the attention level. Then Artificial Neural Network (ANN) and Recurrent Neural Network-Long Short Term Memory (LSTM) based Deep learning (DL) architectures have been proposed for analysing the data. Next, the trained DL model has been implanted into a robotic system that is capable of detecting various features; ultimately leading to the calculation of visual attention for reading, browsing, and writing purposes. This system is capable of checking the attention level of the participants and also can detect if participants are present or not. Based on a certain level of visual focus of attention (VFOA), this system interacts with the person, generates awareness and establishes verbal or visual communication with that person. The proposed ML techniques have achieved almost 99.24% validation accuracy and 99.43% test accuracy. It is also shown in the comparative study that, since the dataset volumes are limited, ANN is more suitable for attention level calculation than RNN-LSTM. We hope that the implemented robotic structure manifests the real-world implication of the proposed method.

Depression and anxiety are prevalent mental illnesses that are frequently disregarded. It is estimated that more than 5% of the population suffers from depression or anxiety. Although there have been a number of influential studies in these fields, the majority of the research focuses on one or two factors for detection purposes, whereas these factors are not mutually inclusive and vary among studies. To mitigate these issues, we must first consider all possible symptoms associated with depression and develop a multimodal diagnosis system that may take into account any number of patient-specific factors. If multiple factors can be addressed within a single learning model, it is advantageous for data collection and future development. To facilitate training with missing modalities, we propose an attention-based multimodal classifier with selective dropout and normalization, which can facilitate the training of various multimodal datasets on one neural network. We have experimented with three multimodal datasets with varying modalities to show the impact of combined training in one neural network and achieved an F1 score of 0.945. However, missing modalities in the model can create uncertainty in the prediction. For uncertainty approximation, the Monte Carlo dropout (MC dropout) and the spectral-normalized neural Gaussian process (SNGP) with the coefficient of variation and S1-Score metrics are implemented to provide important information about multimodal diagnosis processes. In the experiment, selective dropout with SNGP achieved a coefficient of variation in loss of 0.384 and an S1-score of 0.9374.