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

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.