基于Openpose的三维人体姿态比对

Data-driven 3D Human Pose Estimation

随着人工智能和计算机视觉等计算机技术的发展,我们的生活越来越便利。于此同时,相关人体姿态识别的技术发展,在各个领域起到越来越到多的作用。本文针对人体运动姿态的对比问题开展研究。基于Openpose来获取二维的骨骼数据坐标,转化为三维骨骼数据。但通过该方法获取的骨骼坐标点,时常会因噪声产生的抖动,本文利用了卡尔曼滤波器来消除噪声连续坐标点的抖动。随后,利用余弦相似度和欧式距离的方法结合起来,来判断学员和标准动作偏离的大小,根据得出的分数,来判断教练与学员运动姿态的相似程度。这种方法,为图像识别在动作教学方面上提供了新的思路。

With the development of computer technology such as artificial intelligence and computer vision,our life is more and more convenient.At the same time,the development of human posture recognition technology plays a more and more important role in various fields.This paper focuses on the comparison of human motion posture.Based on openpose,two-dimensional bone data coordinates are obtained and transformed into three-dimensional bone data.However,the skeleton coordinate points obtained by this method often jitter due to noise.In this paper,Kalman filter is used to eliminate the jitter of noise continuous coordinate points.Then,cosine similarity and Euclidean distance are used to judge the deviation between the trainees and the standard movements.According to the score,the similarity between the coaches and the trainees is judged.This method provides a new idea for image recognition in action teaching.