Human Activity Recognition Using LSTM with Feature Extraction Through CNN

Abstract

Human activity recognition is important for detecting anomalies from videos. The analysis of auspicious activities using videos is increasingly important for security, surveillance, and personal archiving. This research paper has given a model which can recognize activities in random videos. The architecture has been designed by using BiLSTM layer which helps to learn a system based on time dependencies. To convert every frame into a featured vector, the pre-trained GoogLeNet network has been used. The evaluation has been done by using a public HMDB51 data set. The accuracy achieved by using the model is 93.04% for ten classes and 63.96% for 51 classes from same data set only. Then, this network is compared with other state-of-the-art method, and it proves to be a better approach for the recognition of activities. Abstract should summarize the contents of the paper in short terms, i.e. 150–250 words.