Dynamic long short-term memory network for skeleton-based gait recognition

As an important biometric, gait has been extensively used for human recognition. Currently gait analysis based on 3D skeleton data are more commonly studied due to its accuracy, succinctness, and view-invariant representation. Traditional methods generally extract geometric features of human body and temporal dynamics characteristic. However, designing features from ad-hoc schemes mainly is based on experience, and thus it is difficult to obtain high-dimensional features beyond the limits of human interpretability. In this paper, considering that Long Short-Term Memory(LSTM) recurrent neural network can model the long-term contextual information of time sequences well, we first propose a dynamic approach to LSTM network for gait recognition from skeleton data. These skeleton data is essentially the discriminative information without extracting any hand-crafted features in advance. The framework is able to process sequences with variable length. Our experiments based on three publicly available datasets demonstrate that the proposed gait recognition method is effective and outperformed those in literature.