Neural Networks in Human Activity Recognition: Trends, Challenges, and Future Directions

Human Action Analysis is a procedure in tracking individual gestures, motion and actions which finds applications, including healthcare, surveillance, security and human-computer communication. With the proliferation of sensors and deep learning techniques, HAR has marked remarkable development in the past few years. This review paper gives a brief survey of the current studies in Activity Analysis, focusing on deep learning-based approaches. The various sensors employed in HAR, including accelerometers, gyroscopes, and vision-based sensors were used in acquiring signals. The main objective is to design an appropriate model for recognizing/classifying human actions effectively to support elderly and differently abled people. The evolution of deep learning architectures used in detection such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks can be utilized for action recognition. The characterization of these algorithms is evaluated using various metric indices including precision, accuracy, recall, Fl-score, and mean average precision. The challenges and limitations of existing HAR systems, including data purity, imbalance nature of class, and positioning of sensor were also presented. This review aims to provide a clear picture in Action Recognition, highlighting the Pros and cons of different approaches. By identifying the prospects and obstacles, this review seeks to inspire future research in behaviour analysis, ultimately leading to the development of more accurate, efficient, and robust systems.