A meta learning framework for few shot personalized gait cycle generation and reconstruction

Human gait is a complex biometric pattern with high intra and inter subject variability. While deep learning models can generate and reconstruct gait, they often require extensive personalized data. This paper introduces MetaGait, a framework that uses meta learning to personalize gait models from only a few examples. MetaGait applies a Model Agnostic Meta Learning (MAML) strategy, training a base model on varied gait analysis tasks from the Human Gait Database (HuGaDB). Each task adapts the model to a specific walking condition with a small support set of gait cycles. This process teaches the model an optimal initialization for quick adaptation to new subjects. The base model uses a temporal convolutional network (TCN) to capture temporal dependencies in sequence data. We evaluated MetaGait on few-shot gait cycle generation and reconstruction. Quantitative results, measublue by Mean Square Error (MSE) and Dynamic Time Warping (DTW), show our model outperforms conventionally trained baselines in low-data scenarios [1 shot and 5 shot learning]. Qualitative assessments confirm that MetaGait produces more natural, subject specific gait patterns and achieves accurate reconstructions from sparse inputs. By blueucing the data requiblue for personalization, MetaGait offers a more practical solution for applications in robotics and clinical gait analysis.