Gait Based Emotion Recognition Using Inertial Data

Abstract

Emotion recognition is an active area of research in the domain of human computer interaction. Computers can interact better with humans if they can explicitly understand human emotions. Most of the existing state of the art methodologies rely heavily on facial expression analysis which has its own drawbacks. In this research work, we present a set of hand-crafted features computed from the inertial data of natural walk of a human, which can be used to predict human emotions with higher accuracy. We recorded 6D accelerations and angular velocities of 40 subjects using an on-board IMU of a smartphone attached at the chest. The subjects were asked to walk in their natural gait by assuming themself in one of the six basic emotions i.e. happy, sad, anger, fear, disgust and surprise. The raw inertial signals were segmented into sequences of steps and strides. A set of hand-crafted features from the time, frequency and wavelet domains for steps and strides was computed. The hand-crafted features were used to train two predictors namely Random Forest and Support Vector Machines. Using 10-fold cross validation strategy on stride features set, a classi cation accuracy of 95% was achieved with two classes of emotions and a classi cation accuracy of 86% was achieved with six classes of emotions. The results have been computed on di erent sets of features and show that the proposed set of hand-crafted features can recognize emotions with reliably and accurately.