基于骨骼点特征的运动视频关键帧提取模型  

Human motion video keyframe extraction modelbased on bone point features

作　　者：高雪雪 谷林[1] Gao Xuexue;Gu Lin(Xi'an Polytechnic University,Xi'an 710600,China)

机构地区：[1]西安工程大学,西安710600

出　　处：《国外电子测量技术》2022年第9期88-94,共7页Foreign Electronic Measurement Technology

基　　金：国家重点研发计划(2019YFB1703801)项目资助。

摘　　要：针对人体运动分析中信息冗余带来的庞大计算和存储压力,提出了一种结合了人体姿态估计技术的视频关键帧提取技术,该技术利用人体骨骼点信息提取运动角度特征和基于角度的衍生特征,同时依据感知哈希算法,引入运动特征矩阵概念来描述并映射视频帧,由此计算出帧间视觉内容的距离特征。通过运动特征的设计来启发集成机器学习,最终构建出提取运动视频关键帧模型。从而去除序列间信息增益较少的冗余,大大降低视频帧的冗余问题,进而提高了视频应用、管理效率。通过实验验证,该模型能够取得较好的提取效果。Aiming at the huge calculation and storage pressure brought about by information redundancy in human motion analysis, this paper proposes a video keyframe extraction technology that combines human posture estimation technology, which extracts motion angle features and angle-based derived features through human bone point information, and introduces the concept of motion feature matrix to describe and map video frames according to the perception hash algorithm, thereby calculating the distance characteristics of visual content between frames. Through the design of motion features to inspire integrated machine learning, and finally build a keyframe model that extracts motion video. Therefore, the redundancy with less information gain between sequences is removed, and the redundancy of video frames is greatly reduced, which improves the efficiency of video application and management. Through experimental verification, the model can achieve better extraction results.

关 键 词：提取关键帧 随机森林 感知哈希 骨骼点特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]