Gait Representation: from Vision-Based to Floor Sensor-Based Gait Recognition

Footstep recognition is a form of biometric authentication that verifies or identifies individuals by their unique underfoot pressure patterns. The selection of an appropriate gait representation (i.e., feature extraction technique) therefore plays an important role in its performance. This investigation explores twenty state-of-the-art 2D gait representations for foot-step recognition, including several appearance representations established for vision-based gait recognition that are novel to the field. In this study, the focus is on access control scenarios (such as high-security facilities, airport checkpoints, or sacred places) where only a small amount of footsteps are available for model training and users may be expected to be barefoot. While peak pressure image features showed the best performance for identity verification across many operating points (an average balanced accuracy of 93% across 86 user models), appearance representations provided meaningful and non-redundant information and should therefore be considered in combination with traditional pressure and time-based representations for developing footstep authentication systems. This paper informs the design and selection of gait representations for floor sensor-based gait recognition, and discusses some possible solutions for developing verification systems with limited training samples.