STUFY AND APPLICATION OF AI TECHNIQUES FOR ACCURATE HUMAN ACTIVITY RECOGNITION

Abstract

HAR is the fastest-growing field of Computer Vision (CV) that has a broad range of applications. Recognition of human activities is a very challenging task as we are moving towards automation and AI. In this thesis, we have proposed an intelligent smartphone-based human activity recognition system using deep learning techniques. HAR system can automatically classify and predict the daily human activities gathered using the inbuilt smartphone sensors like gyroscope and accelerometer. The proposed deep learning models are the CNN and the LSTM network. The proposed models are evaluated on WISDM dataset. Moreover, Batch-Normalization (BN) and Dropout layers are used in between the proposed network to reduce the overfitting of models. The proposed models have shown good results with high accuracy and low complexity. ‘CNN’ model achieves accuracy up to 95% and the ‘LSTM’ model achieves accuracy up to 97.5%. The complexity of the proposed network is low with 21,922 as trainable parameters. Experimental results are depicted with ‘Learning curves’ and ‘Confusion matrix’. The results of this research are quite promising and has no limitation to utilize on environment.