Human posture recognition model integrating YOLOv8 pose and graph multilayer perceptron

Abstract

In complex and real scenes, human posture recognition is often affected by factors such as large scale changes, diverse posture forms, and frequent local occlusions, which in turn leads to insufficient key point prediction stability and structural consistency. Focusing on the goal of improving human posture recognition accuracy and real-time performance, this paper constructs a human posture recognition model that integrates target detection, skeleton structure modeling, and temporal smoothing mechanisms. This model takes key point detection as its core content and introduces an end-to-end attitude prediction framework. Moreover, on this basis, it combines cross-stage feature fusion and lightweight detection heads to optimize model efficiency, and at the same time enhances the overall consistency of human posture through structured skeleton modeling and temporal feature smoothing. Experiments show that the precision reaches 0.95, the recall rate reaches 0.94, the F1 value increases to 0.95, and the keypoint positioning error decreases to 3.10, which is about 26% less than the comparison model error. In terms of inference efficiency, the single-frame inference time is shortened to 0.08 s, the frame rate is increased to 125, and the parameter size is compressed to 7.60. In the multi-pose test, the model accuracy remains above 95 under occlusion and complex pose conditions. This model achieves a good balance between accuracy, stability, and real-time performance, and provides an efficient and reliable technical solution for human posture recognition in complex scenes.