Behavior Recognition Based on 3D Skeleton Features

Faced with the problem of low efficiency of human behavior recognition, this paper proposes a method of indoor human action recognition research which is based on 3D Kinect skeleton features. The framework of 3D skeleton has the advantage of little data and contains key information about the behavior, plus, it can be good represented by sparse dictionary. Therefore, first of all, we divided the human body motions into global motion, arm and leg motion, which was based on 3D skeleton. By extracting multiple of feature sequences, the 3D skeletal feature set was formed; then, we used online dictionary to learn and SPCA (Sparse Principal Component Analysis) to reduce dimensionality, followed by MT-LMNN (Multi-Task Large Margin Nearest Neighbor) and SVM (Support Vector Machine) fusion scoring mechanism to make the best judgment, to recognize human action recognition behavior. The experimental results show that, this method has higher detection rate.