Publication

Abstract : Human Activity Recognition (HAR) is the process of predicting the activity of a person based on a trace of their movement using sensors. Wearable smart devices such as body-worn sensors and smart phones are currently used to detect human activity recognition. By wearing these devices, it is easy to collect data and perform further analysis to obtain necessary information on human activities. The main goal of this project is to recognize patterns from the raw data and extract useful information about the daily activities of the human. Existing strategies, like deep learning, made progress for explicit recognition, yet the recognition of the transition from one activity to the other is poor. In this project, the data corresponding to human activity recognition is presented to convolutional neural network (CNN) model and transitions between activities is presented to the long short-term memory (LSTM) model. The combined model of CNN and LSTM is implemented to improve the HAR accuracy. The HAPT and HAR datasets are used to train the model. The performance analysis of the model is evaluated by tuning the model hyperparameters, such as number of neurons, batch size, learning rate, batch normalization, and dropout. Further the evaluation metrics namely, precision, F1-score and recall are computed. It is observed from the results that better HAR accuracy is obtained with optimized hyperparameter values.