叶嘉庆  苟兴悦  张启仁　陈俊屹　郑斯敏　曹东

（广州中医药大学医学信息工程学院，广东 广州 510006）

摘要：针对俯卧撑自动化考核中代偿动作隐蔽、评分维度单一导致的评估准确率较低问题，提出一种多维运动学特征与SVM融合的俯卧撑动作质量评估方法。该方法采用MediaPipe BlazePose提取上半身8个关键点的二维坐标并计算肘关节与身体角度序列；经Savitzky-Golay滤波和波峰检测将俯卧撑动作视频分割为动作周期，在每个周期内计算下压幅度、上肢对称性、躯干姿态、躯干对称性和核心稳定性五维运动学特征；以SVM RBF分类器对动作进行合规/异常二分类，对合规动作采用多维FastDTW算法与标准模板对齐并输出百分制质量评分。在Kaggle公开俯卧撑数据集上的实验表明，本文方法在实现动作计数的同时完成质量评估与百分制质量评分，推理速度小于0.1 s/视频，具备快速动作视频评估的应用潜力。

关键词：俯卧撑；动作质量评估；运动学特征；支持向量机；人体姿态估计

中图分类号：TP391.41           文献标志码：A           文章编号：1674-2605(2026)04-0007-06

DOI：10.12475/aie.20260407                            开放获取

Push-up Movement Quality Assessment via Multi-dimensional Kinematic Features and SVM

YE Jiaqing  GOU Xingyue  ZHANG Qiren  CHEN Junyi  ZHENG Simin  CAO Dong

(School of Medical Information Engineering, Guangzhou University of Chinese Medicine, Guangzhou 510006, China)

Abstract: To address the problems of low evaluation accuracy caused by subtle compensatory movements and a single scoring dimension in automated push‑up assessment, this paper proposes a push-up movement quality assessment via multi-dimensional kinematic features and SVM. The method employs MediaPipe BlazePose to extract 2D coordinates of eight upper‑body keypoints and compute the angle sequences of elbow joints and body. After Savitzky-Golay filtering and peak detection, the push‑up video is segmented into action cycles. Within each cycle, five kinematic features are computed: pushing depth, upper‑limb symmetry, trunk posture, trunk symmetry, and core stability. An SVM with RBF kernel is used to perform binary classification of actions as compliant or abnormal. For compliant actions, the multi‑dimensional FastDTW algorithm is applied to align with a standard template and output a quality score on a 100‑point scale. Experiments on the Kaggle public push‑up dataset demonstrate that the proposed method simultaneously achieves action counting, quality assessment, and 100‑point scoring, with an inference speed of less than 0.1 s per video, showing its potential for rapid video‑based action evaluation.

Keywords: push-up; movement quality assessment; kinematic features; support vector machine (SVM); human pose estimation