An Improved Lightweight Human Normative Pose Estimation Algorithm
Danyu Li, Yunci Ma, Jiajun Zou, Zhiming Xu · 2024
Human pose estimation algorithms have high accuracy and speed, but there are still issues of missed and false detections. This article proposed a lightweight human normative posture estimation algorithm for precise target positioning and action norm evaluation in human normative posture recognition. First, the YOLOv4 object detection algorithm and the lightweight OpenPose algorithm were combined to obtain target bone node data. Second, by calculating action angles as features for extraction, an action matching method was constructed to evaluate and correct the training actions. Finally, the extracted training action skeleton image was trained and classified using a deep convolutional neural network (DCNN) model. Relieved the uncertainty and irregularity of human normative movements. Training and validation were conducted on the collected datasets of professional coaches, and the experimental results shown that, our model achieved better performance than other models.