User gait biometrics in smart ambient applications through wearable accelerometer signals: an analysis of the influence of training setup on recognition accuracy

Gait recognition can exploit the signals from wearables, e.g., the accelerometers embedded in smart devices. At present, this kind of recognition mostly underlies subject verification: the incoming probe is compared only with the templates in the system gallery that belong to the claimed identity. For instance, several proposals tackle the continuous recognition of the device owner to detect possible theft or loss. In this case, assuming a short time between the gallery template acquisition and the probe is reasonable. This work rather investigates the viability of a wider range of applications including identification (comparison with a whole system gallery) in the medium-long term. The first contribution is a procedure for extraction and two-phase selection of the most relevant aggregate features from a gait signal. A model is trained for each identity using Logistic Regression. The second contribution is the experiments investigating the effect of the variability of the gait pattern in time. In particular, the recognition performance is influenced by the benchmark partition into training and testing sets when more acquisition sessions are available, like in the exploited ZJU-gaitacc dataset. When close-in-time acquisition data is only available, the results seem to suggest re-identification (short time among captures) as the most promising application for this kind of recognition. The exclusive use of different dataset sessions for training and testing can rather better highlight the dramatic effect of trait variability on the measured performance. This suggests acquiring enrollment data in more sessions when the intended use is in medium-long term applications of smart ambient intelligence.