A human action recognition technique is proposed in this paper that extracts silhouettes from video sequences and uses deep stacked auto encoder to classify them. The silhouettes obtained from the original images are able to represent the action in the image more effectively. The silhouette images are given as input to the deep stacked auto encoder to learn the human actions. Deep learning is widely used to learn features from the silhouette images to predict the target output values. The features learnt from the silhouette images are able to represent the shape information more efficiently. The experimental results are demonstrated on Weizmann and KTH datasets. The proposed method is evaluated using various performance measures. The accuracy values obtained for the two datasets are 88.2% and 93.9% respectively.
Volume 12 | Issue 1