Pedestrian Identification Based on Spatio-temporal Feature of Plantar Pressure

Gait is an important biometric feature to identify individuals, which has a wide range of needs in the fields of anti-terrorism security inspection and criminal investigation. Previous studies have proved that using the platform with pressure sensors to record Ground Reaction Forces (GRF) when pedestrians walk is an effective way to collect gait Spatio-temporal information. However, it is difficult to extract behavior features directly from such high-dimensional Spatio-temporal sequences, which leads to unsatisfying classification accuracy. Here, we propose a dual-channel convolutional neural network based on the Gramian angular field (GAF), which provides an efficient and simple way to recognize Spatio-temporal information and realize the identification of pedestrians' identities. In the proposed method, there are two independent channels composed of convolutional blocks. The input of the first channel is the peak pressure image (PPI) that is used to extract spatial features. By encoding the GRF sequence using the GAF algorithm, the GAF image is obtained as the input of the second channel and used for the temporal features extraction of the walking pattern of individuals. Then, the spatial and temporal features are merged and used for pedestrian identification by a fully connected network. An adaptive optimizer is used to train the dual-channel network model. In a classification task for 33 pedestrians, the recognition rate of the proposed method achieves 98%. Compared with the models only using spatial features, the accuracy of recognition of the proposed method is significantly improved. The results indicate that the recognition of pedestrians is well realized through the hierarchical extraction and fusion of Spatio-temporal features by a dual-channel model.