A Robust Human Activity Recognition Approach Using OpenPose, Motion Features, and Deep Recurrent Neural Network

With the emerging advancements in computer vision and pattern recognition, methods for human activity recognition have become increasingly accessible. In this paper, we present a robust approach for human activity recognition which uses the open source library OpenPose to extract anatomical key points from RGB images. We further use these key points to extract robust motion features considering their movements in consecutive frames’. Then, a Recurrent Neural Network (RNN) with Long Short-term Memory cells (LSTM) is used to recognize the activities associated with these features. To make the approach person-independent, different subjects from different camera angles are used. The proposed method shows promising performance, with the best result reaching an overall accuracy of 92.4% on a publicly available activity data set, which outperforms the conventional approaches (i.e. support vector machines, decision trees, and random forests) which achieve maximum accuracy of 78.5%. The proposed activity recognition system can contribute in prominent research fields such as image processing and computer vision with practical applications such as caregiving for older people to help them live more independently.