Human Activity Recognition in Videos Using Deep Learning

Human Activity Recognition (HAR) is a challenging classification task. In the past, it traditionally involved the identification of the movement and activities of a person based on sensor inputs, apply signal processing to receive features and fit the features into a machine learning model. In recent times, deep learning methods have shown good results in automatic Human Activity Recognition. In this paper, we propose a pre-trained CNN (Inception-v3) and LSTM based methodology for Human Activity Recognition. The proposed methodology is evaluated on the publicly available UCF-101 dataset. The results show that the proposed methodology outperforms recent state-of-art methods in terms of accuracy (79.21%) and top-5 accuracy (92.92%) on the HAR task.