Multimodal Human Action Recognition Base on Sparse Joints Graph Convolutional Network and Dynamic Image ROI Sampling

Human action recognition has broad applications across various domains, yet accurately capturing discriminative features for reliable action classification in videos remains a challenging task. Traditional methods often rely heavily on skeletal data; however, they struggle to differentiate between actions with similar local motion patterns, such as drinking and eating, where misclassification can occur due to similar hand-to-head movements. To overcome these limitations, we propose a sparse joint graph that enhances global feature extraction of graph convolution by facilitating long-range information propagation across non-adjacent joints. Additionally, a dynamic image sampling strategy is introduced to select key joints based on motion magnitude, thereby reducing the influence of irrelevant background information and improving image-based feature quality. Extensive experiments on multiple benchmark datasets demonstrate that our proposed multimodal framework consistently outperforms existing state-of-the-art methods in human action recognition.