Disguising Gait Detection and Recognition Based On Deep Learning

Abstract

In a standard gait recognition model, the disguising gait sequence will lead to error recognition. A disguising gait detection and recognition model is proposed in this paper. We proposed a new gait cycle detection method based on the variation curve of distance between the ankles. And we introduce the polynomial fitting method to filter the noise of the curve. A model-based fusion feature representation is used in the model, which is combined the static and kinetic features extracted from the skeleton of the human body. A LSTM-based autoencoder network is introduced to encode the feature sequences to a single feature vector. The disguising detection is conducted by calculating the similarity of the feature vectors of the contiguous gait cycles. In the normal gait recognition module, a parallel deep learning network is adopted to improve the feature extraction ability and accuracy of the recognition model. The effectiveness of the proposed model is evaluated on two gait datasets, which are acquired by Kinect sensor. The experiment results demonstrate the proposed model is capable of disguising gait detection, and has a relatively high recognition rate.