基于Kinect的健身动作识别与评价

Fitness Movement Recognition and Evaluation Based on Kinect

摘要: 随着健康问题越来越受到人们的重视，运动健身越来越被广大人民所接受。如何更有效的运动健身并合理减少身体伤害，是现今科学健身运动的研究热点之一。本文利用Kinect收集健身动作数据，对动作进行自动分类，与标准动作进行比对评分，计算出易受伤程度。具体的，首先通过对骨架点的收集和预处理，从骨架点数据中提取特征值，计算权重并对特征值归一化，得到健身动作的动作测试数据集和模板数据集。通过KNN算法对测试数据进行分类与识别，得到测试动作的分类结果。对应分类结果，利用评价和易受伤害计算模型，最终得到健身动作的评分和建议。实验表明，提取的骨架数据特征对所有动作识别仅配合微调的KNN算法就能有较好的效果，并比神经网络算法更具有广泛性。动作评价和易受伤分析可以减少运动伤害，提高健身运动趣味性。

Abstract: As health issue enrolling a more important role in daily life, sports and fitness movements are accepted increasingly by the majority of people. How to exercise more effectively and reduce physical injuries reasonably is one of the hottest researches in the current scientific fitness movement. In this paper, we use Kinect to collect fitness movement data, then automatically classify the movements, and compare them with standard movements to calculate the degree of vulnerability. Specifically, we collect and preprocess the skeleton data of person’s movements first, and then we extract the feature and normalize the weight to form a representative feature among all movements. Second, we classify all the movements by little tuned KNN algorithm, and form the movement’s template model by similarity calculation. Third, we evaluate the movements based on the identified template and the injury model to guide the performance of movement. The experiments show the features we extract for all movements can represent the movement well that only little tuned KNN algorithm can get a competitive recognition result. We has compared neural net algorithm and KNN that the former is less generalized than the latter. The score and suggestion of fitness movement can reduce sports injuries and improve the enjoyment of fitness sports.