ℓ1-Norm based reconstruction error evaluation for human action recognition

Automatic human action recognition is a core functionality of systems for video surveillance and human object interaction. Sparse representation-based classification (SRC) has been widely used for face recognition (FR) and extended for video classification. Reconstruction error is an important scalar for object classification. If the reconstruction error of intra-class is small and the reconstruction error of inter-class is large, the classifier is robustness. In this paper, we propose an improved sparse representation-based classification method with ℓ
₁
-norm based re-construction error evaluation and local spatio-temporal features for human action recognition. The main idea in our method is to make the reconstruction error sparse and stable in order to improve the robustness of classifier. Because of L2-norm is sensitive to the singular samples, the reconstruction error with L2-norm scale is unstable for training and recognition. Against to L2-norm, the ℓ
₁
-norm makes the model more compact and sparse. So, we restrict the reconstruction error by a ℓ
₁
-norm constraint in sparse representation over-compete dictionary training and classifying. The performance of the proposed approach was evaluated on two public available datasets, KTH Action Dataset and UCF Sports Action Dataset, and obtained a higher recognition accuracy compared to other state-of-art methods in the literatures.