Search

Deep learning methods have recently attracted attention in solving various tasks through their learning and determining capability for powerful features. It has been known that the absence of a pooling layer in the deep learning architectures made the network performance deteriorate significantly. Existing pooling methods assume that each value in the small region of the local feature map contributes equally to generate the pooled feature map. However, some values in the pooling region may contribute partially for generating the map. In addition, existing methods usually focused on the spatial dimension and ignored the inter-channel relationship during the pooling process. Besides, most of the pooling methods utilize a static pooling strategy which is designed based on the experts' knowledge rather than learnt from the data. This study proposes a pooling method that is able to learn from the training data by utilizing a learnable attention mechanism, called Normalized Attention Inter-Channel Pooling (NAIP). The mechanism of the pooling process mainly focuses on the most important feature region for generating the pooled feature map. The proposed method is compared with the state-of-the-art works which use graph convolutional neural network and convolutional neural network architectures for skeleton-based human action recognition and image classification tasks. The experiment demonstrates that the NAIP approach outperforms the existing methods under the same circumstances. [ABSTRACT FROM AUTHOR]

Common approaches in the cross-domain sensor-based human activity recognition are based on the homogeneous domain adaptation which relies on the assumption that the training and testing data are of homogeneous feature space. In reality, such an assumption does not always hold. For example, although two devices may share common sensors, it is possible that each of them has its own specific sensors. In this case, the homogeneous domain adaptation approaches cannot be used directly. Although heterogeneous domain adaptation approaches have been proposed to handle such feature space heterogeneity, most of them require some label information in the target domain, which is often difficult to obtain. As a compromise, the hybrid domain adaptation has been recently proposed to address feature space heterogeneity. Instead of using the target domain label information, it exploits the common features between domains as additional information, so the adaptation can be performed in an unsupervised manner. However, it still neglects the possibility of the common features between domains having different distribution, which may lead to the negative transfer of the domain-specific features. In this work, we introduce a domain-invariant latent representation of the common features to enhance the specific feature transfer in the hybrid domain adaptation approach. The latent representation learning and the domain-specific feature transfer are performed jointly using an autoencoder-based framework. The experimental result shows that the performance improves when the common features are furthermore aligned in the latent space. It is also shown that, in overall, our model outperforms existing approaches, yielding up to 9.48% accuracy improvement. [ABSTRACT FROM AUTHOR]

Continuous exposure to stress leads to many health problems and substantial economic loss in companies. A lot of attention has been given to the development of wearable systems for stress monitoring to tackle its long-term effects such as confusion, high blood pressure, insomnia, depression, headache and inability to take decisions. Accurate detection of stress from physiological measurements embedded in wearable devices has been the primary goal in the healthcare industry. Advanced sensor devices with a high sampling rate have been proven to achieve high accuracy in many earlier works. However, there has been a little attempt to employ consumer-based devices with a low sampling rate, which potentially degrades the performance of detection systems. In this paper, we propose a set of new features, local maxima and minima (LMM), from heart rate variability and galvanic skin response sensors along with the voting and similarity-based fusion (VSBF) method, to improve the detection performance. The proposed feature set and fusion method are first tested on the acquired dataset which is collected using the wrist-worn devices with a low sampling rate in workplace environments and validated on a publicly available dataset, driveDB from PhysioNet. The experimental results from both datasets prove that the LMM features can improve the detection accuracy for different classifiers in general. The proposed VSBF method further boosts the recognition accuracy by 5.69% and 2.90% in comparison with the AdaBoost, which achieves the highest accuracy as a single classifier on the acquired, and the DriveDB dataset, respectively. Our analyses show that the stress detection system using the acquired dataset yields an accuracy of 92.05% and an F1 score of 0.9041. Based on the analyses, a soft real-time system is implemented and validated to prove the applicability of the proposed work for stress detection in a real environment. [ABSTRACT FROM AUTHOR]

The sensor-based human activity recognition (HAR) using machine learning requires a sufficiently large amount of annotated data to realize an accurate classification model. This requirement stimulates the advancement of the transfer learning research area that minimizes the use of labeled data by transferring knowledge from the existing activity recognition domain. Existing approaches transform the data into a common subspace between domains which theoretically loses information, to begin with. Besides, they are based on the linear projection which is bound to linearity assumption and its limitations. Some recent works have already incorporated nonlinearity to find a latent representation that minimizes domain discrepancy based on an autoencoder that includes statistical distance minimization. However, such approach discovers latent representation for both domains at once, which causes sub-optimal representation because both domains compensate each other's reconstruction error during the training. We propose an autoencoder-based approach on domain adaptation for sensor-based HAR. The proposed approach learns a latent representation which minimizes the discrepancy between domains by reducing statistical distance. Instead of learning representation of both domains simultaneously, our method is a two-phase approach which first learns the representation for the domain of interest independently to ensure its optimality. Subsequently, the useful information from the existing domain is transferred. We test our approach on the publicly available sensor-based HAR datasets, using cross-domain setup. The experimental result shows that our approach significantly outperforms the existing ones. [ABSTRACT FROM AUTHOR]

Similarity measures for analyzing human behavior patterns are inseparable part of the intelligent environment, with the assistive functionality as its core value. The measure must represent the contexts properly which characterizes the users' environment. Recent studies attempted to formulate similarity measures for the intelligent environment by incorporating relevant contexts. Yet, they are lacking the integration of multiple inter-related important contexts, which leads to model underestimation and possibly the wrong interpretation. This work proposes a context-based similarity measure for analyzing human behavior patterns. The proposed similarity measure extends and combines the commonly used contexts (i.e., activity, location, and time) into a holistic measure. To avoid the biased representation of activity context similarity, we add one more aspect, namely process context, which describes a wide range of interval relations among the activities of a user. The proposed approach is compared with state-of-the art similarity measures by evaluating both real and simulated data. The result shows that our approach yields the better result in terms of robustness toward noises. In addition, our approach also shows a better reliability compared to previous works in the case of anomaly detection. [ABSTRACT FROM AUTHOR]