A Novel Human Abnormal Posture Detection Method Based on Spatial-Topological Feature Fusion of Skeleton

Yuefeng Ma, Zhiqi Cheng, Deheng Liu, Shiying Tang

Abstract: Skeletons can clearly represent human postures, while the skeleton-based human abnormal posture detection has been widely used. Previous skeleton-based methods of human abnormal posture detection are often considered from a single perspective such as joint positions, joint distances and bone angles, without fully utilizing the spatial and topological information of skeleton. To overcome these shortcomings, we propose a novel human abnormal posture detection method based on Spatial-Topological Feature Fusion (STFF) of skeleton. In this study, we present a new definition of spatial similarity of two skeletons called ‘Skeleton Keypoints Displacement Metric’, with the minimum total displacement of all skeleton keypoints. Based on the similarity, we introduce an optimal skeleton matching method to select the optimal matching skeleton from a given set of template skeletons which includes skeletons of typical normal human postures. Then the deviation between the target and the optimal matching skeleton can be regarded as spatial feature to be integrated with the topological feature which can be obtained from the topological structure of the target. At last, we employ a classification method, to achieve human abnormal posture detection. Experiments show that our method achieves state-of-the-art performance on our dataset and FallDown detection dataset.

External IDs:dblp:conf/mmm/MaCLT25