Human Action Recognition Based On Semi-Supervised Learning

Considering the different speed of human action in video,this paper proposes a multiscale human behavior recognition algorithm framework based on semi-supervised learning for the problem of not using the information of unknown tag data in the current video human action recognition field.Firstly,the improved dense trajectory features with video lengths of L and 2L are extracted,and the trajectories are described by Hog,Hof,and MBH descriptors.Then,the trajectories of different scales of video are encoded to obtain corresponding video scales.The video representation corresponds to different views under semi-supervised learning.Finally,the distributed constrained k-means is used to effectively utilize the information of the video data of the unknown tag to improve the adaptability and recognition efficiency of the model.Extensive experiments are conducted on two standard action datasets,KTH and HMDB51 to verify the effectiveness of the algorithm.The experimental results show that compared with the latest technology of HMDB51,the proposed method is superior to other existing methods on KTH dataset,and has achieved competitive results.