Predicting soft biometrics of a human by using gait signatures

Abstract

Human walking pattern is believed to be unique in its nature and it is well-known through a number of previous research that soft biometrics of humans are encoded in sparse kinematic information, or in other words, in human walking pattern. These kinematic information can now be recorded using inertial measurement units, thanks to the advancement of modern digital devices. In most cases, body-mounted digital devices are used to capture a subject's movement. The raw signals are used to learn and estimate the encoded soft biometrics, which are present in the captured data. We developed a deep learning based algorithm to learn various soft biometrics about a human, such as gender and age from a single step from the recorded data of the person. In this study, we show how a single step of a human gait may be used to determine gender and age. We acquired 6D angular velocities and accelerations of 86 volunteers with the help of chest-mounted inertial measurement units as they conducted their normal gait tasks. Big sequences of signals were broken down into separate step data by subjecting the data to segmentation before being fed to the model. We were successful in predicting a person's gender with a maximum accuracy of 100 % for males and 99.32 % for females. The root mean squared error for age prediction was only 2.29 years, the mean squared error was only 5.25 years and mean absolute error of 0.29 years. We have got the precision score of 0.96, recall score of 0.975 and F1 score of 0.976 for age estimation.