In this study, we propose a novel method, named ActJOLO, which builds upon the existing JOLO model by incorporating an advanced self-supervised learning technique as an upstream guide for posture recognition. Our approach emphasizes the analysis of high-intensity motion features within the human body, thereby enhancing the efficiency of action modeling.
Experimental results on the NTU RGB+D dataset demonstrate that our framework improves processing speed compared to the original model, while maintaining high ccuracy. This work offers a new perspective on skeleton-based human action recognition and highlights its potential for deployment on low-performance processors.