Horizontal graph connections for skeleton-based human action recognition on UAV-Human

Unmanned Aerial Vehicles (UAVs) are finding increasing use in surveillance, search, rescue, and many other applications. Human action recognition (HAR) using UAV data is an essential task that enables real-time detection of human behavior from visual data. The altitude and mobility of UAVs make HAR difficult. Graph Convolutional Network (GCN) architectures have achieved remarkable performance for skeleton-based HAR in recent years. While GCN is quite good at learning spatiotemporal data, it strongly relies on the dependencies of the graph definition. GCNs’ capacity to represent the relationship between large-distance joint pairs is limited. Joints at horizontal positions in the skeleton model have strong dependencies in many typical actions such as walking, running, etc. The dependencies come from the need to keep the human body balanced while moving. In this work, these dependencies are exploited by defining a graph with horizontal edges for action representation. In this way, the adjacency matrix can capture implicit dependencies of symmetric joint pairs. The performance of the GCN using horizontal graph connections is evaluated on the dataset UAV-Human. Experiments show that the proposed method outperforms SOTA methods.