采用动态压缩与约束搜索的视频动作比对算法
Video Action Matching Algorithm Based on Dynamic Compression and Constraint Search
庄丽萍 1陈锦 1蒋锦华 1姚洪泽 1蔡志明2
作者信息
- 1. 福建理工大学 电子电气与物理学院,福建 福州 350118
- 2. 福建理工大学 电子电气与物理学院,福建 福州 350118;福建理工大学 电子信息与电气技术国家级实验教学示范中心,福建 福州 350118
- 折叠
摘要
以计算机视觉的动作比对方法为基础,提出一种基于人体姿态识别算法、动作序列数据压缩和约束搜索路径的动态时间归整(DTW)算法的视频动作比对方法.首先,利用人体姿态识别网络获取人体动作关键关节点的二维坐标数据并转化为关节角度特征;其次,使用优化的双阈值动态数据压缩方式获取动作视频的关节特征短序列;最后,使用动态时间规整的优化算法对两个压缩后的短视频序列进行相似度计算,实现人体动作各关节比对.实验结果验证了优化算法的高效性,在八段锦视频数据下,比对时间从 16.353 s降至 0.079 0 s,凸显出在计算时间节省方面的明显优势.
Abstract
This study conducted a research based on the action alignment method of computer vision,and proposed a video action alignment method based on human pose recognition algorithm,action sequence data compression,and constraint search path dynamic time warping(DTW)algorithm.This comparison method first utilized a human posture recognition network to obtain two-dimensional coordinate data of key joint points in human motion and converted it into joint angle features.Secondly,an optimized dual threshold dynamic data compression method was used to obtain short sequences of joint features in action videos.Finally,the optimization algorithm of dynamic time warping was used to calculate the similarity of the two compressed short video sequences so as to realize the comparison of human motion joints.The ex-perimental results confirmed the efficiency of the optimization algorithm,reducing the matching time from 16.353 seconds to just 0.079 0 seconds for Baduanjin video data,highlighting a significant advantage in terms of time savings in computation.
关键词
人体姿态识别/动作比对/动态时间规整/约束搜索/动态压缩/时间序列Key words
human pose recognition/action comparison/dynamic time regulation/constrained search/dy-namic compression/time series引用本文复制引用
基金项目
福建理工大学科研启动基金(GY-Z21064)
出版年
2024
NETL
NSTL
万方数据