| Abstrakt: |
Automatic Human Action Recognition (HAR) using RGB-D (Red, Green, Blue, and Depth) videos captivated a lot of attention in the pattern classification field due to low-cost depth cameras. Feature extraction in action recognition is an important aspect. As compared to Depth Motion Maps (DMM), Depth Motion Maps–Local Binary Pattern (DMM–LBP) provides compact representation of features. After extracting features using DMM–LBP, Principal Component Analysis (PCA) is used for dimensionality reduction. For classification task, randomly generated input weights of Extreme Learning Machine (ELM) can lead to non-optimal results. Therefore, in this paper, we have used an improved learning algorithm named Self-adaptive Differential Evolution (SaDE) ELM for action classification. In the proposed approach, DMM–LBP is used for feature extraction and SaDE–ELM is used for action classification. To evaluate strength of the proposed approach, experiments are performed on four public datasets, namely, MSRAction3D, MSRDaily Activity3D, UTD–MHAD, and MSRGesture3D. The proposed approach gives better accuracy compared to existing approaches Kernel ELM (KELM), l2-Collaborative Representation Classifier (l2-CRC), and Probabilistic CRC (ProCRC) methods. We have also presented statistical significance of results in the paper. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink