3D Skeleton Action Recognition for Security Improvement

Most of action recognition methods allow achieving high action recognition accuracy, but only after processing the entire video sequence, however, for security issues, it is primordial to detect dangerous behavior occurrence as soon as possible allowing early warnings. In this paper, we present a human activity recognition method using 3D skeleton information, recovered by an RGB-D sensor by proposing a new descriptor modeling the dynamic relation between 3D locations of skeleton joints expressed in Euclidean distance and spherical coordinates between the normalized joints, A PCA dimension reduction is used to remove noisy information and enhance recognition accuracy while improving calculation and decision time. We also study the accuracy of the proposed descriptor calculated on limited few first frames and using limited skeleton joint number, to perform early action detection by exploring several classifiers. We test this approach on two datasets, MSR Daily Activity 3D and our own dataset called INDACT. Experimental evaluation shows that the proposed approach can robustly classify actions outperforming state-of-the-art methods and maintain good accuracy score even using limited frame number and reduced skeleton joints.