Research On Human Action Recognition In Video

Human action recognition refers to automatically recognizing the behavior of human performing in videos by computer algorithms through the analysis of human action features.In recent years,it has become a hot topic in the field of computer vision,which has been widely used in advanced human-computer interaction,intelligent driving support system,sports action analysis,intelligent video surveillance and other fields.Due to the non-rigidity of human motion,the complexity of the background and camera movement,human action recognition becomes a very challenging subject.In this paper,the methods of human action recognition in video are studied.The specific research work is as follows:1.Human action recognition based on dense trajectories samples the whole image of every frame densely,which leads to high feature dimensionality,large computational cost and containing the irrelevant background information.A human action recognition method is proposed based on dense trajectories with saliency detection.First,a multi-scale static saliency detection is used to get the action subject positions,which then is combined with the results of dynamic saliency detection to get human action areas.The original algorithm is improved by only extracting dense trajectories in these areas.To enhance adequacy of feature expression,Fisher Vector is used to replace BOW model encoding the features.At last,SVM is learned to get the results of human action recognition.The experimental results conducted on KTH dataset and UCF Sports dataset show that proposed method has improved on the recognition accuracy comparing with the original algorithm by 1.2% and 0.4%,respectively.2.In order to make full use of the video-wide temporal information and reduce the redundant frames and dimensions of features,a method of extracting valid video frames and performing temporal rank pooling them for human action recognition is proposed.First,VLAD is used to encode dense trajectory features of every frame of video to get feature representations.The Cosine similarity analysis of frame features is employed to remove the redundant features and extract feature sequence of valid video frames.Temporal rank pooling is performed to order feature sequence of valid frames temporally and get the feature vectors capturing the evolution of video-wide temporal information.At last,SVM is learned to get the results of human action recognition.The experimental results shows that the recognition accuracy achieved by the proposed method conducting on HMDB51 datasets and UCF101 datasets are 65.2% and 89.4%,respectively.3.Aiming at the problem that dimensionality disaster easily occurs in the processing of dealing with video data,a dimensionality reduction method called linear sequence discriminant analysis is proposed for human action recognition.Firstly,ViBe algorithm is used to subtract the backgrounds of video frames to get action areas,where extracting dense trajectories as video features from,to suppress the noise from background caused by camera movements.Then,Fisher Vector is used to encode the features and linear sequence disciminant analysis is conducted on the encoded features,the sequence class separability is measured by dynamic time warping distance between different classes,learning a linear discriminative projection of the feature vectors in sequences to a lower-dimensional subspace by maximizing the separability of the sequence classes while minimizing the within-sequence class separability to reduce data dimension.At last,SVM is learned to get the results of human action recognition.The experimental results conducted on KTH datasets and UCF101 datasets show that the proposed method has improved the recognition accuracy.