The Study of Human Action Recognition in Videos with Long Short-Term Memory Model

Human Action Recognition (HAR) requires tracking diverse fields of human activities in healthcare, education, entertainment, visual monitoring, video collection, and irregular behavior identification. Machine learning techniques are widely used to identify an action. These techniques have their limitations; they do not provide automatic feature selection for this purpose. This paper presents a deep learning algorithm, the Long Short Term Memory Model (LSTM) combined with the pre-trained Convolutional Neural Network (CNN) VGG16 that will provide a completely automatic feature selection. Firstly, the features in the video sequence are extracted using a VGG16 convolutional neural network. The video is then classified using an LSTM algorithm. The well-known benchmark, UCF101, which includes all 101 classes, is used to test the network. This research will propose an improved approach for identifying an action from multiple videos using the deep learning method. The presented method for an automated system for recognition of human action classification presents 99% accuracy.