A Meta Learning Framework for Few Shot Personalized Gait Cycle Generation and Reconstruction

Human gait is a complex biometric signature with significant intra-subject and inter-subject variability. Traditional deep learning models for gait generation and reconstruction often require substantial personalized data, limiting their applicability in scenarios with scarce data for new individuals. This paper introduces MetaGait, a novel framework that leverages meta-learning to enable rapid personalization of gait models from only a few examples. MetaGait employs a Model-Agnostic Meta-Learning (MAML) approach. The core idea is to train a base gait model on a diverse set of gait analysis tasks derived from the Human Gait Database (HuGaDB). Each task involves adapting the model to a specific subject or walking condition using a small support set of gait cycles, and then evaluating its performance on a corresponding query set. This process allows the model to learn an optimal initialization that facilitates quick adaptation to unseen subjects. The base model architecture utilizes a temporal convolutional network (TCN) backbone for its efficacy in capturing temporal dependencies in sequence data. We evaluate MetaGait on both few-shot gait cycle generation and reconstruction tasks using HuGaDB. Quantitative results demonstrate that MetaGait significantly outperforms baseline models trained from scratch or fine-tuned conventionally, especially in low-data regimes (1-shot and 5-shot learning). Qualitative assessments show that MetaGait generates more natural and subject-specific gait patterns and achieves more accurate reconstructions from sparse inputs. Comparison with recent state-of-the-art methods highlights MetaGait's competitive performance in personalized gait modeling. MetaGait offers an effective solution for few-shot personalized gait modeling. It significantly reduces the data dependency for personalization, paving the way for more practical applications in areas such as robotics, and clinical gait analysis.