Late Fusion of Graph Convolutional Networks for Action Recognition Using UAV Bone Data

Unmanned Aerial Vehicle (UAV) is becoming popular for surveillance, search, and rescue operations. Human action recognition (HAR) using UAV data is an important task that allows human behavior to be detected instantly from the visual data. UAVs’ attitudes and motions make HAR for UAVs challenging. In recent years, Graph Convolutional Network (GCN) architectures have been mainstream for skeleton-based HAR. This paper proposes a Late Fusion GCN (LF-GCN) framework by applying late fusion to three graph-based deep learning architectures. The framework combines the advantages of different graph-based deep-learning models. The proposed model outperforms existing GCN schemes on the large-scale UAV-Human dataset.