摘要
为减少传统行为识别方法中光流场特征提取的计算量,提取光流区域的关键点,分析关键点的频率域多尺度轨迹,与运动方向以及形状信息进行融合,得到关键点特征;为凸显与重要运动部分相关的局部特征能够产生更加有区分度的行为表示,提出一个运动部分规则框架识别每个运动部分的动作,将已得运动部分组合成每一个行为样本的区分度输入分类器进行人体行为识别。将实验结果与其它算法进行比较,验证了该算法具有更好的识别率和实时性。
To reduce the computation of the feature extraction of the optical flow field in the traditional action recognition, criti-cal points of motion flow field were extracted, and multi-scale trajectories were generated from those points and characterized in the frequency domain. A sequence was described by fusing this frequency information with motion orientation and shape informa-tion. To highlight the local features which were associated with important motion parts led to a more discriminative action repre-sentation ,a motion part regularization framework was introduced to recognize the motion part, and the learned motion part weights were further utilized to form a discriminative weighted vector representation for each action sample for final classifica-tion. Compared with other algorithms？ the results show that the algorithm has better recognition and real-time performance.
出处
《计算机工程与设计》
北大核心
2017年第9期2546-2550,共5页
Computer Engineering and Design
基金
国家自然科学基金项目(50808025)
中山市科研计划基金项目(2014A2FC388)
关键词
光流关键点
多尺度轨迹
运动部分识别
局部特征
人体动作识别
critical points of flow field
multi-frequency trajectory
motion parts regularization
local features
human action recognition
