A Robust Deep Model for Human Action Recognition in Restricted Video Sequences

In this paper, we propose an action recognition algorithm in noisy data conditions with Convolutional Neural Network (CNN) as the front end and Deep Bidirectional Long Short Term Memory (DBi-LSTM) as the backend. The deep features are extracted from the input frames using a VGG16 model. The sequential information among frames is learned using the DBi-LSTM part, which is composed of three layers stacked together in both forward and backward directions to increase the learning depth. The proposed algorithm achieved 96.77% vs. 96.76% and 95.83% vs. 91.60% accuracy of the baseline methods on KTH and YouTube datasets, respectively. Moreover, the proposed algorithm has shown significant robustness in noisy training data as the accuracy drops only 1% down.