<![CDATA[Human activity recognition (HAR) is a widely adopted technique in applications requiring accurate identification of human actions. However, HAR approaches often face challenges in generalizing across complex datasets with multi-view variations, resulting in reduced classification accuracy. Existing classifiers face shortcomings in predicting human activities due to the presence of irrelevant video frames, leading to frequent misclassifications. This research proposes a selective kernel network-2D convolutional neural network with additive angular margin loss for deep face recognition (SKN-2D-CNN with ArcFace loss) to recognize human activity effectively. SKN dynamically adapts the receptive field for learning multi scale spatial features, enhancing the recognition of intricate human activities with varying motion scales. In the embedding space, ArcFace loss introduces an angular margin penalty that improves inter-class separability and intra class compactness for recognition. Together, the proposed method minimizes misclassification in human activity by improving the robustness of feature representation. Feature extraction using visual geometry group 19 (VGG19) captures spatial features like edges, textures and shapes from video frames, enhancing the model’s ability to differentiate between complex human activities. The proposed method achieves high accuracy of 99.16 and 98.75% on the UCF101 and HMDB-51 datasets, respectively, when compared with existing methods such as CNN and bidirectional gated recurrent unit (BiGRU).]]>
Copyrights © 2026
