A Multi-Branch CNN-LSTM Based Human Activity Recognition Using Wearable and Smartphone Sensors
Mst. Alema Khatun, Mohammad Abu Yousuf, Mohammad Ali Moni · 2025
Monitoring human activity in various circumstances is the aim of Human Activity Recognition (HAR) systems. Several inertial sensors are widely used in this context. These sensors are being used more frequently in smart devices like smartphones. In this chapter, we collected raw data from smartphone sensors, namely H-Activity , using an application for accelerometers, gyroscopes and linear acceleration. Deep learning models are used to compare the performance of several smartphone sensor-based data types. In order to enhance recognition performance, we also developed the CNN-LSTM model, a multi-branch hybrid network that has been tested on H-Activity datasets. The performance of our model was further demonstrated using a public dataset, WISDM. When using the H-Activity data that we have gathered, the suggested model surpasses rival models by accurately identifying human activity with a 99.05%, and a 99.23% recognition rate when using data from the WISDM database. We believe that the data we collected will be relevant for further study, and we hope that the model we constructed will be useful in a clinical environment.