Feature Learning for Accelerometer based Gait Recognition

Recent advances in pattern matching, such as speech or object recognition support the viability of feature extraction with deep learning solutions for gait recognition. Past papers have evaluated convolutional neural networks for this task, while this work focuses on how autoencoders perform in this context. A biometric pipeline was implemented that is capable of identification when presented with step cycles, while also performing feature extraction that employ autoencoders of various configurations. The results obtained from the ZJU-GaitAcc dataset show that fully convolutional autoencoders are able to learn good representation from any type of gait segment. Measurements also show that representation learning works even better when it is incorporated into an end-to-end model of a discriminative classifier.