Incremental Human Gait Prediction without Catastrophic Forgetting

Human gait prediction is an important task in predictive exoskeleton control. However, if static models are used to facilitate this task, two problems arise. First, the models cannot adapt to new environments and terrains during deployment, and second, the models cannot be personalized to any given end user without costly involvement of a human expert. Incremental models can alleviate these shortcomings, but they usually are prone to catastrophic forgetting, which can be dangerous during live deployment. In this work, we introduce an incremental model, that can learn human gait from scratch without outside interference, but does not fall prey to catastrophic forgetting. We test and evaluate our model on a real world gait database and show, that it delivers competitive results with regard to other standard approaches.