An embedded network system for human action recognition based on compound moments
Abstract An embedded network system of human action recognition is proposed for unattended indoor video surveillance. After the establishment of all behavior action datasets, we extract and describe data features using compound moments. Affine moments invariant which are of shift, rotation and scale invariance and can well identify distorted target images are used to describe action features. Compound moments, including affine moments invariant and other moments, are used to describe all the features of the target. Meanwhile, we build multi-class SVM classifier that has good generalization to recognize each action and compare with K-Nearest-Neighbor classifier. Results show that SVM classifier with nonlinear kernel functions based on one-versus-one and one-versus-all construct methods achieves very good recognition accuracy of 95%.
Author's Name: Zhou, S., Zhang, X., Li, W., Zhang, Z., Li, M., Fan, X.
Volume: Volume 9
Issues: Issue 24
Keywords: Action recognition, Affine moment invariant, Support vector machine