Actions Recognition Based On Convolution Neurlal Network And Composite Feature

The research of actions recognition in video is mainly to make the computer understand the meaning of human action better.The computer vision can be better used in the fields of human-computer interaction,monitoring security,and artificial intelligence etc.The main difficulty of human actions recognition is focused on two aspects of the action feature extraction and action feature classifier.Due to the projection of objects in three-dimensional space to two-dimensional space,the direction of lights,the changes of background and environmental shelter can cause significant changes in two-dimensional video images,which increases the difficulty of actions recognition.How to extract the human motion information in the video image has become the key point of the video motion feature extraction.It is an important research direction to find a kind of video feature which has stronger information description ability and recognizable ability.In action recognition,it is necessary to improve the learning efficiency and recognition rate of the classifier.The main contributions of this paper are as follows:1.This paper proposes a composite video feature which combines the space-time interest points feature,optical flow feature and human gray image.As a common feature of video analysis,space-time interest points feature has the ability of global motion information describe,while the optical flow feature has strong ability to describe the locality motion information.The gray level image in human body can reflect the instantaneous movement posture of human body.The composite features proposed in this paper have a stronger ability to describe motion information and stronger recognizable ability.2.This paper uses the 3D convolution neural network as the feature classifier in actions recognition.Compared with the traditional linear support vector machine classifier,template matching classifier and other shallow learning classifiers,convolutional neural network classifier can extract abstract information in motion,and has a stronger ability of learning features.Because of the convolution kernel of 3D convolution neural network can not only slide in horizontal and vertical direction of video images,but also can slide on the time axis in the convolution process,the 3D convolution neural network can locate the location of the action in time and space better.