Action Recognition Using Temporal Evolution And Feature Learning

As one of the most important subjects in the field of computer vision,more and more attention has been paid to human behavior recognition in video.This technology is closely related to the social life of mankind.Its application scope has covered many fields such as intelligent video surveillance,human-computer interaction,motion analysis,virtual reality,video retrieval,medical assistance and so on.With the related technology matures,the application prospect is more clear.The study of human behavior recognition technology has achieved some results recently,although some results can better recognize the behavior in some existing data sets,can not avoid the light,jitter,background clutter,occlusion and other noise interference,which make it still difficult to use in the actual production of life.In the same pattern recognition process,human behavior recognition includes three main steps: the representation of human behavior in the video,the learning of the human behavior model,and the recognition on the behavior.The emphasis is on the representation of the behavior,whether it can extract the feature vector which has high discrimination and robustness to the original video data,will seriously affects the recognition results of the behavior.The existing research methods are mostly for the study of the characteristics of different airspace characteristics in the video,and the research on the time domain characteristics is less.In this paper,we mainly study the differences in the behavior of different behaviors temporally aiming to extract the temporal discriminant characteristics of different behaviors with time and realize the effective recognition of human behavior.The main contributions of this paper are as follows:1.A method of human behavior recognition based on positive and negative timing feature learning is proposed.The proposed method is based on the assumption that the behavioral video is different in the forward and reverse play,and the similarities and differences of the similar behavior are similar,while the differences between the different behaviors are different.This method firstly extracted spatial local feature from the video,then learned the behavior of forward and reverse timing characteristics from the spatial local features,the forward and reverse timing feature is extracted from the difference between the forward and reverse timing by the classifier,and used it as the video feature representation to achieve fast and efficient recognition behavior.2.A method of human behavior recognition based on temporal regression feature learning is proposed.Behavior is an orderly arrangement of atomic movements in time dimension,which is often continuous in time dimension.This method focuses on the timing characteristics of behavior,and proposes a time domain feature extraction method that studies the behavior of time from the forward video sequence.Firstly,learn the timing evolution characteristics from the behavior,and then learn the evolution of the characteristics of the evolution using regression,finally regard the regression coefficient as the timing feature of the behavior.3.A method of human behavior recognition based on behavioral framework is proposed.Video is susceptible to a variety of noise disturbances in the acquisition process,and this interference is also presented in the feature vector that characterizes the behavior with the encoding of the behavior in the video.Aiming at the noise interference in the feature vector,a method of learning the main frame of the behavioral behavior from the time series feature is proposed.By implementing the pyramid in the time domain and the frequency domain,the feature of the low frequency main frame is extracted to realize the more robust behavior Discriminant representation.Compared with the timing feature proposed in the previous work,not only the accuracy precision is improved,but the behavior dimension of the job is greatly reduced,which improves the recognition speed.