HDS-SP: A novel descriptor for skeleton-based human action recognition

Jingjing Liu, Zhiyong Wang, Honghai LIU

Abstract: 3D skeleton-based human action recognition has attracted much attention due to a wide spectrum of promising applications in terms of depth sensors. This paper, based on the principle of the better viewpoint the better action recognition performance, devises a HDS-SP descriptor for skeleton-based human action, i.e., the histogram of distributed sectors based on their specific projections. The HDS-SP descriptor consists of both spatial and temporal information from specific viewpoint, in which projecting 3D trajectories on specific planes and creating reasonable histograms using proposed way are two primary contributions. Inspired by the nature of human action, the spatial information is incorporated into the histogram of the displacement of one joint between two successive frames voting for corresponding bins following weight-based rules over an undefined plane while the specific projection planes are optimized by both local search algorithm and Particle Swarm Optimization (PSO); the temporal information is captured by temporal hierarchical construction. The proposed method is evaluated in five widely researched datasets for skeleton-based human action recognition with results significantly outperforming the state-of-the-art.