ARRNET: Action recognition through recurrent neural networks
Kumaresh Krishnan, Nikita Prabhu, R. Venkatesh Babu · 2016
In this paper, we propose a novel method for recognition of human actions from joint points. Our approach utilizes Long Short Term Memory (LSTM), a Recurrent Neural Network (RNN) variant to keep track of and train the network on joint information across an ordered sample of 15 frames from a video. We ensure that important properties of actions like left right invariance are learnt by the system through data augmentation. Our experiments on sub-JHMDB and Penn Action datasets provide encouraging results which surpass previous action recognition models on these datasets. We analyse our model on the results obtained for tests on these datasets.