基于半监督学习的人体异常行为识别  

Recognition of human abnormal behavior based on semi-supervised learning

作　　者：李知菲[1] 端木春江[1] 

机构地区：[1]浙江师范大学数理与信息工程学院,浙江金华321004

出　　处：《浙江师范大学学报（自然科学版）》2013年第3期258-262,共5页Journal of Zhejiang Normal University:Natural Sciences

基　　金：浙江省教育厅科研项目(Y201016478)

摘　　要：需要进行人体异常行为识别的视频一般都是未标记的序列图像,传统的有监督的识别方法往往不能较好地反映其行为的特征,识别率不高.提出了一种基于半监督学习的人体异常行为识别方法,首先使用基于DTW距离的self-training进行标记数据扩充,然后用此扩充的序列图像样本集合训练对应的HMM,最终进行异常行为识别.实验结果证明该方法有效且识别率较高.The video that needed to carry out the recognition of the abnormal human behaviors were mostly the unlabelled image sequences, the traditional supervised recognition methods usually could not well retrieve the behavior features and thus did not have a high recognition rates. It was presented a human abnormal behavior recognition method based on semi-supervised learning. It first extended the set of the labeling data based on the DTW distance by using self-training, then, the extended set of image samples was used to train the corre- sponding HMM, finally, the recognition of the abnormal behavior was carried out. The experiment results demonstrated that this method was effective with a high recognition rate.

关 键 词：人体异常行为识别 半监督学习 动态时间规整 隐马尔科夫模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]