Feature-level combination of skeleton joints and body parts for accurate aggressive and agitated behavior recognition

This paper presents a novel and practical approach for aggressive and agitated behavior recognition using skeleton data. Our approach is based on feature-level combination of joint-based features and body part-based features. To characterize spatiotemporal information, our approach extracts first meaningful joint-based features by computing pairwise distances of skeleton 3D joint positions at each time frame. Then, distances between body parts as well as joint angles are computed to incorporate body part features. These features are then effectively combined using an ensemble learning method based on rotation forests. A singular value decomposition method is used for feature selection and dimensionality reduction. The proposed approach is validated using extensive experiments on variety of challenging 3D action datasets for human behavior recognition. We empirically demonstrate that our proposed approach accurately discriminates between behaviors and performs better than several state of the art algorithms.