Simple Head Trajectory Measurement for Deep Gait Recognition

Biometric-based person identification methods, like fingerprint or face recognition, have been challenged due to cyberattacks using deep learning technology. Against cyberattacks, gait recognition has proven to be a promising alternative for robust person identification. However, compared to conventional biometric recognition systems, gait measurement systems are complicated, obstructing the collection of sufficient gait datasets and, consequently limiting the reliability of person identification applications. In this paper, we introduced a new method of gait recognition using the head trajectory segmented by head peaks on the human longitudinal axis during walking. We aimed to simplify gait recognition systems and processes to efficiently create larger databases for reliable individual identification based on gait patterns. Head trajectories were collected from 586 stride sequences of 12 participants walking on a flat and hard ground floor. We used a deep learning neural network model, requiring only 64 sampled 3-dimensional head positions (64x3) as an input. This paper displays different results with three covariates: sampling rate, the number of trained gait strides, and inclusion of body height, considering various measurement environments. The average performance was measured in 10 repetitions with different training datasets to mitigate biased results due to the small number of samples. The average identification accuracy has reached up to 94.4%, and the equal error rate (EER) was measured from 2.33% to 4.13% in the selected practical scenarios. Since various sensors like cameras and inertial sensors can capture head trajectories, our proposed method can offer widespread usage in diverse environments across the physical and virtual worlds.