HAR-Depth: A Novel Framework for Human Action Recognition Using Sequential Learning and Depth Estimated History Images

Human action recognition (HAR) is a challenging task due to the presence of pose and temporal variations in videos. To address these challenges, HAR-Depth proposed this paper with sequential shape learning along novel concept depth history image (DHI). A deep bidirectional long short term memory (DBiLSTM) constructed for model relationship existing between frames. Action information each frame extracted using pre-trained convolutional neural network (CNN). The estimated projected onto X-Y plane form DHI. During learning, through DHI used train CNN network. By leveraging trained knowledge network, overfitting issue handled. finetuned recognize actions from query images. Data augmentation adopted avoid by virtually increasing training set. work evaluated on publicly available datasets like KTH, UCF sports, JHMDB, UCF101, HMDB51 achieves performance accuracy 97.67%, 95.00%, 73.13%, 92.97%, 69.74% respectively. results suggest that performs better terms overall accuracy, kappa parameter precision compared other state-of-the-art algorithms present earlier reported literature.