Open-Set Human Identification Based on Gait Radar Micro-Doppler Signatures

Human identification based on gait radar micro-Doppler signatures has received significant attention in recent years. In real world applications, a recognition system should have the ability to reject unknown identities as well as correctly recognize known identities, i.e. the open-set identification. In this paper, we propose a novel framework to tackle this challenging problem. We base our approach on a probabilistic discriminant model built on a deep discriminative representation network (DDRN). Specifically, we first train a DDRN with the cosine margin (CM) loss to map gait samples into an embedding space where the learned features belonging to the same identity are much closer together and those of different identities are further apart. Then, in the learned embedding space, the class-inclusion probability (CIP) model associated with each known identity can be constructed based on the statistical Extreme Value Theory (EVT) to bound each other's support region, which can then be used to estimate the class-belongingness probabilities of a probe sample. Finally, we threshold on these probabilities to determine whether it belongs to one of the known identities or an unknown class. Experiments on a radar measured gait dataset show the effectiveness of our approach.