Edge-Joint Assisted and Salient Enhanced Self-Supervised Action Recognition

We propose an algorithm that focuses on edge joints and salient features in human skeleton data. The edge joint consistency module filters strong connectivity to emphasize weakly connected joints, capturing subtle action differences, especially for similar actions. The self-attention-enhanced salient feature module learns local salient features through attention masks, highlighting crucial action-specific details while maintaining global context. This approach improves action recognition by better utilizing joint connectivity and enhancing feature distinction between classes.