Action Recognition of Human Skeletal Data Using CNN and LSTM

Abstract

Human action recognition recognizes an action performed by human beings in order to witness the type of action being performed. A lot of technologies have been developed in order to perform this task like GRN, KNN, SVM, depth maps, and two-stream maps. We have used 3 different methods in our research first method is a 2D CNN model, the second method uses an LSTM model and the third method is a combination of CNN+LSTM. With the help of ReLu as an activation function for hidden and input layers. Softmax is an activation function for output training of a neural network. After performing some epochs the results of the recognition of activity are declared. Our dataset is WISDM which recognizes 6 activities e.g., Running, Walking, Sitting, Standing, Downstairs, and Upstairs. After the model is done training the accuracy and loss of recognition of action are described. We achieved to increase in the accuracy of our LSTM model by tuning the hyperparameter by 1.5%. The accuracy of recognition of action is now 98.5% with a decrease in a loss that is 0.09% on the LSTM model, the accuracy of 0.92% and loss of 0.24% is achieved on our 2D CNN model while the CNN+LSTM model gave us an accuracy of 0.90% with the loss of 0.46% that is a stupendous achievement in the path of recognizing actions of a human. Then we introduced autocorrelation for our models. After that, the features of our models and their correlations with each other are also introduced in our research.