Human Action Recognition Based on Sparse LSTM Auto-encoder and Improved 3D CNN
Qi Fu, Shiwei Ma, Lina Liu, Jinjin Liu · 2018
In computer vision area, extracting effective information from video data to realize automatic human action recognition has become one of the most active research points. In order to deal with the high dimensional challenge from massive amount of video data and extract effective features by feature self-learning model based on deep learning, a method based on a sparse long-short term memory (LSTM)auto-encoder and improved 3D convolution neural network (3D CNN)is proposed in this paper. We first construct a sparse LSTM auto-encoder to extract the key frames. And then, to extract lower dimensional features for approximating the upper limit of these features classification ability, the 3D CNN model is improved by combing with the feature engineering method, that is recursive feature elimination (RFE)algorithm. Finally, we extract high-level features by the improved 3D CNN for human action recognition. The experimental results on typical datasets UCF101 demonstrate that our algorithm has significant effect on data dimensionality reduction and outperforms some competing hand-craft feature and single feature methods.