摘要
为实现全景视场下人体行为特征的有效提取,在原始形状上下文特征匹配算法的基础上,提出一种基于自适应分块思想的金字塔匹配核算法.结合光学成像原理及全景视场下人体投影特点,计算图像二阶中心矩对人体轮廓主轴方向进行补偿.然后对轮廓点进行均匀采样,对各采样点提取形状上下文特征,在匹配过程中分析高维特征空间中采样点的分布特点,采用自适应分块的思想对金字塔匹配核函数的收敛策略进行改进,根据各维度上数据的分布范围自适应地调整收敛系数,以保证各个维度上的点集收敛速度一致.最后通过室内摔倒检测实验来验证算法的可靠性,使用K均值聚类方法进行识别,识别率可达92.9%.该特征提取算法为智能监控系统的稳定性提供了保障.
In order to extract human action features effectively under the panoramic view,we propose a novel PMK algorithm with adaptive partitioning based on the original shape context feature matching algorithm.First we combine optical imaging principle with the projection characteristics of human body under the panoramic view,and utilize the second moment to correct the principal axes direction of human contour.Then do the uniform sampling with the edge points,and extract the shape context feature at each sample point.In the process of matching sampling points,through the analysis of the distribution characteristics of sampling points in the high dimensional feature space, we introduce adaptive partitioning to Pyramid Match Kernel（PMK）algorithm to improve the convergence strategy.According to the range of data points at each dimension,the improved algorithm can adjust convergence coefficient the data adaptively,thus achieving the consistent convergence speed of the points set at each dimension.Finally we perform an experiment on fall indoor detection to verify the reliability of the proposed algorithm,and the K-means clustering algorithm is used for classification,the recognition rate can reach 92.9%.The results show that the improved feature extraction algorithm can provide the guarantee for the stability of the intelligent monitoring system.
出处
《光子学报》
EI
CAS
CSCD
北大核心
2017年第12期197-206,共10页
Acta Photonica Sinica
基金
国家自然科学基金青年基金(No.61605016)
吉林省科技发展计划项目(No.20160520018JH)资助~~
