Human activity recognition (HAR) systems employing wearable sensors are a promising area of research for tracking human activity. Recently, wearable devices such as smartwatches and sensors have been developed for activity recognition and monitoring. These systems aim to obtain the subject’s state within his or her environment by exploiting heterogeneous sensors attached to the body. With the development of deep learning, new strategies have emerged to facilitate and solve the HAR problems. In this work, a deep multilayer bidirectional long-short memory (Bi-LSTM) architecture has been implemented to detect human activities. Instead of training a single model as in traditional LSTM methods, two models are presented in the Bi-LSTM scheme, one for learning the input data sequence and the other for learning the reverse sequence. Finally, a new novel postprocessing approach has been proposed based on windowing and voting in the last step to improve the average F1 score. Comprehensive investigations on the three publicly available datasets consisting of a different set of activities were used to evaluate the performance of the proposed framework. The empirical results of this paper on AReM, Mhealth and PAMAP2 datasets attained 95.46, 95.79 and 93.41% average F1 score, respectively. The results also revealed that selecting the window size and implementing the appropriate voting method has a significant effect on improving the average percentage of the F1 score.

Human Activity Recognition (HAR) is the process of identifying and analysing activities performed by a person (or persons). This paper proposes an efficient HAR system based on wearable sensors that uses deep learning techniques. The proposed HAR takes the advantage of staking Convolutional Neural Network and Long Short‐Term (LSTM), for extracting the high‐level features of the sensors data and for learning the time‐series behaviour of the abstracted data, respectively. This paper proposed a Fuzzy Soft‐max classifier for the dense layer which classifies the output of LSTM Blocks to the associated activity classes. The authors’ decision for proposing this classifier was because sensor data related to the resembling human activities, such as walking and running or opening door and closing door, are often very similar to each other. For this reason, the authors expect that adding fuzzy inference power to the standard Soft‐max classifier will increase its accuracy for distinguishing between similar activities. The authors were also interested in considering a post‐processing module that considers activity classification over a longer period. Using the proposed Fuzzy Soft‐max classifier and by the post‐processing technique, the authors were able to reach the 97.03 and 85.1 rates of accuracy for the PAMAP2 and Opportunity dataset, respectively.

Fall is one of the most critical health challenges in the community, which can cause severe injuries and even death. The primary purpose of this study is to develop a deep neural network using wearable sensor data to detect falls. Most datasets in this field suffer from the problem of data imbalance so that the instances belonging to the Fall classes are significantly less than the data of the normal class. This study offers a dynamic sampling technique for increasing the balance rate between the samples belonging to fall and normal classes to improve the accuracy of the learning algorithms. The Sisfall dataset was used in which human activity is divided into three categories: normal activity (BKG), moments before the fall (Alert), and role on the ground (Fall). Three deep learning models, CNN, LSTM, and a hybrid model called Conv-LSTM, were implemented on this dataset, and their performance was evaluated. Accordingly, the Conv-LSTM hybrid model presents 96.23%, 98.59%, and 99.38% in the Sensitivity parameter for the BKG, Alert, and Fall classes, respectively. For the accuracy parameter, we have managed to reach 97.12%. In addition, by using noise smoothing and removal techniques, we can hit a 97.83% accuracy rate. The results indicate the proposed model's superiority compared to other similar studies.