View Invariant Gait Feature Extraction using Temporal Pyramid Pooling with 3D Convolutional Neural Network

Due to the accessibility of vast volume of security camera data, there is an opportunity for identifying people using gait profiling for social security and person identification. But the gait sampling process is highly dependent on the environmental conditions of the surroundings. In addition to that different condition of the person to be identified like different clothing conditions also adds significant complexity to the problem. In this paper, we have presented a solution in form of a deep learning network for dealing with the above issues while gait feature extraction and person identification. The proposed method uses Temporal Pyramid Pooling with 3D Convolutional Neural Network to extract the temporal features from the silhouettes of the gait sequence. Then a series of 2D Convolutional is applied on these spatio-temporal features to get an embedding for the gait sequence of the person. This embedding is then used as an unique identifier for each person. After obtaining the unique identifier each gait sequence is assigned an identifier which uniquely identifies the person. Our method has been tested on CASIA-B dataset consisting of 124 persons gait sequences from different viewpoint and in different conditions. The experimental results shows that our method has outperformed available gait identification approaches.