Few-Shot Open-Set Gait Recognition Based on Radar Micro-Doppler Signatures

Radar-based gait recognition, which ensures privacy protection and functionality in all weather conditions, facilitates biometric recognition advancements. However, it encounters difficulties with data availability, particularly in terms of the quantity and variety of samples. To address these concerns, this research introduces a few-shot open-set gait recognition model that makes use of radar spectrograms. We identified that the “hard examples" result in a “metric imbalance", and proposed a strategy to ensure that the metric space establishment is not adversely affected by these examples. In addition, distinguishing between known and unknown samples is essential for recognition in open-set scenarios, thereby complicating the recognition of known samples. In the context of the meta-learning framework, we have emphasized the issue of ‘negative transfer’ and introduced a ‘knowability discrimination’ strategy to improve task relevance across episodes in meta-learning. To evaluate the efficacy of the ‘FO-Gait’, we developed a radar spectrogram gait dataset comprising twenty-four participants, and employed six few-shot open-set recognition algorithms for comparative analysis. Experiments involving variations in openness underscore the robustness of our concept for the concerned problem, while ablation experiments validate the efficacy of our designs.