The paper introduces a method to handle occlusions in self-supervised skeleton-based action recognition for autonomous robotic systems. The proposed strategy pre-trains using occluded skeleton sequences, employs k-means clustering on sequence embeddings, and uses K-nearest-neighbor to fill in missing data. The authors also introduce an Occluded Partial Spatio-Temporal Learning (OPSTL) framework, building on Partial Spatio-Temporal Learning (PSTL) and utilizing Adaptive Spatial Masking (ASM) for better use of intact skeletons. The effectiveness of these methods is demonstrated on occluded versions of NTURGB+D 60 and NTURGB+D 120.
Publication date: 22 Sep 2023
Project Page: https://github.com/cyfml/OPSTL
Paper: https://arxiv.org/pdf/2309.12029
