Skeleton-Based Posture Estimation for Human Action Recognition Using Deep Learning

The recognition of human actions in real-world scenarios is crucial for smart systems like automatic security monitoring and medical monitoring. Achieving high accuracy in these recognition systems often requires significant resources, including costs and human resources. This is a challenge for all researchers as well as companies around the world. Some recent research has focused on analyzing sequenced postures to recognize human actions using deep learning techniques, which demonstrates promising results. This paper presents new study results on human posture analysis based on skeleton instead of relying on optical flow approach. This method involves analyzing two stages: (1) evaluating the effectiveness of extracting skeletons from videos using various approaches such as YOLOv8 and MediaPipe, and (2) analyzing efficiency of human action inferring from skeletons-based posture sequences using deep learning techniques such as Long short-term memory (LSTM). Additionally, the proposed method is experimented with a KTH dataset. The experimental results compared the skeleton-based human action recognition task between the two pose estimation models. The results demonstrate that the proposed method is more effective in real-world human action recognition tasks with a not-too-complicated implementation process while the accuracy is improved quite well, about 95% to 97%. The results show that solutions based on MediaPipe to extract skeletons outperformer YOLOv8 pose based approach.