Skeleton Moving Pose-based Human Fall Detection with Sparse Coding and Temporal Pyramid Pooling

This paper presents an efficient, yet high-accurate skeleton-based approach for human fall detection, which combines the sparse coding and temporal pyramid pooling techniques. The new method first separates the skeleton joints into five different parts, for each of which a moving pose descriptor is extracted to represent the human sub-actions. The principal component analysis is then employed to reduce the dimensions of the descriptors. Afterwards, sparse coding is invoked to encode each descriptor separately. Finally, these encoded descriptors are treated as a set of time series and then aggregated into the final video descriptors by temporal pyramid pooling, which can acquire temporal tendency to further boost the performance. Experimental results show that the new approach outperforms the state-of-the-art works on some commonly used datasets.