SL-GCNN: A Graph Convolutional Neural Network for Granular Human Motion Recognition

Abstract

Human motion recognition has significantly advanced, with applications in human-computer interaction, virtual reality, intelligent video surveillance, and athletic training. This paper presents SL-GCNN, a novel Graph Convolutional Neural Network framework specifically designed for granular skeletal motion recognition. SL-GCNN incorporates Space-Time Feature Harmonization (STFH) and Linkage Isolation (LI) blocks to enhance classification accuracy for similar motion categories. By leveraging contrastive learning, the model refines hidden layer features and isolates output layer features, effectively addressing the challenges posed by subtle differences in granular motions. Experimental results on the two different datasets demonstrate that SL-GCNN outperforms existing state-of-the-art methods, achieving accuracies of 92.73% and 97.47% on the X-Sub and X-View benchmarks of NTU_RGB_plus_D, and 89.19% and 90.56% on the X-Sub and X-View benchmarks of NTU_RGB_plus_D_120, respectively. The findings highlight the model's robust performance in distinguishing complex granular motion categories, validating its efficacy in real-world applications.