Classification of Human Physical Activities and Postures During Everyday Life

Nowadays Wearable sensor-based Human Activity Recognition (HAR) is gaining popularity for its affordability and low computational demands. These sensors are widely used in healthcare and surveillance. However, using smartphone sensors for HAR can be inaccurate due to their non-fixed position. This study uses the publicly available Human Activity Recognition Trondheim dataset (HARTH) to develop a HAR model. The proposed HAR model is capable to recognize diverse human daily life activities in the free-living environment. Identifying individual natural behaviour in the wild (free living) remains challenging because humans engage in unscripted daily activities. While working with controlled conditions (scripted data) can produce optimal results, this system often struggles when applied in real-life scenario. Multiple machine learning classifiers are tested on time domain features extracted from sensor data, with the Multilayer Perceptron (MLP) classifier achieving an impressive 92.92% accuracy, making a significant contribution to the HAR field.