Skeleton Based Human Activity Recognition

Skeleton-based human activity recognition (HAR) focuses on identifying human actions from skeletal data, aiming to build robust and accurate systems for applications like healthcare and autonomous driving. Current research heavily utilizes graph convolutional networks (GCNs) to model the spatial and temporal relationships within skeletal sequences, with a focus on improving the representation of skeletal data and addressing issues like over-smoothing in GCNs and adversarial attacks. A significant area of investigation involves enhancing the robustness of these systems against adversarial attacks, which seek to subtly manipulate skeletal data to mislead the recognition system, with recent work exploring both attack and defense strategies in black-box and transfer-based settings. This field is crucial for advancing the reliability and security of HAR systems in real-world applications.