Reconciling of the distances in physical space and latent feature space for complex activity modeling

Abstract

In the field of human activity recognition (HAR), fundamental difficulties in complex activities recognition (CAR) remain. In this work, to lay a general pipeline for CAR, we try to reconcile the distance of atomic actions of the complex activities in the physical space with that in the latent feature space by utilizing the Variational Auto-Encoder (VAE) and the Uniform Manifold Approximation and Projection (UMAP). The results of this research show that our pipeline can project the atomic actions into a latent feature space with a lower dimension while preserving the global structure reconciled with their global structure in the physical space.