It is evident that there is still scope for enhancement in the domain of human body posture and movement detection systems. These systems are founded upon the principles of deep learning algorithms. The identified areas for improvement encompass the quantity of data sets, the precision of detection, and the real-time detection performance. Firstly, we extracted frames from action videos to obtain a datasets. Then, we used roboflow to annotate joint anchor points on the datasets. Next, YOLOv8 model was employed for training and testing on the datasets, and the performance of two different YOLOv8 models was compared. The findings of the experimental study demonstrate that YOLOv8n-pose outperforms the YOLOv8s-pose model in terms of training speed and mean average precision (mAP). YOLOv8n-pose maintains recognition accuracy while minimizing model size, parameter count, and runtime memory usage. Experimental results show that YOLOv8n-pose has an average accuracy of 80.1% and a precision of 98% when the number of parameters is only 3.3 M and the GPU memory occupancy is only 1.47 G during training. This ensures a favourable balance between training precision and the velocity of the training process.