Benchmarking the performance of SVMs and HMMs for accelerometer-based biometric gait recognition

Abstract

Support Vector Machines (SVMs) and Hidden Markov Models (HMMs) have been in use for numerous classification tasks in pattern recognition. HMMs can be considered as a de-facto standard in speaker recognition. For accelerometer- based biometric gait recognition these methods have also shown good classification results, which are, however, not comparable as different data sets and features have been used. The contribution of this paper is a comprehensive benchmarking of the stated methods on a single database composed using a standard cell phone. In total, more than 19 hours of accelerometer data from 36 subjects were collected during two sessions. We analyze the influence of time on the recognition rates and state the results for normal and fast walk. In addition, we compare the results obtained when different amounts of training data are used. We show that SVMs are slightly superior to HMMs yielding an Equal Error Rate (EER) of around 10%.