Research On Convolution Neural Network Behavior Recognition Based On Optical Flow Characteristic

The rapid development of computer vision and the gradually improvement of modern society's security awareness,video-based behavior recognition research has important significance and application value in all aspects of human daily life.Scholars at home and abroad has achieved some achievements in the field of behavior recognition.Before the emergence of deep learning,the research route usually first extract the behavior characteristics of video samples,and then classify the behavior recognition.The emergence of Convolutional Neural Network(CNN)technology has brought about a new chapter in the field of behavior recognition.CNN can automatically learn characteristics from the input data layer by layer,and it has invariance for translation,scaling,rotation and other transformations,and widely used in various fields.After having a deep understanding of the research status in the field of behavior recognition and the CNN structure and principle,a method based on the CeLiu algorithm to solve the optical flow field image was proposed based on the behavior recognition of the Time Segment Network in order to improve the real-time and accuracy issues of the behavior recognition process.After experimental comparison and verification,this algorithm has improved the accuracy of the final behavior recognition,but the real-time performance has decreased.In order to further improve the real time and accuracy of the behavior recognition process,the algorithm FlowNet 2.0,which can automatically extract the optical flow field by CNN,was introduced into the behavior recognition process.The smooth flow image with sharp motion boundary can be obtained by the method of the bilinear sampling and the closed-solution of the boundary in this process.Then,the optical flow field is format converted?Standardization and grayscale to reduce redundancy.The experimental results show that the optical flow field image obtained by this method is combined with the time segment network to identify the behavior achieved a very good recognition effect.At the same time,the real-time performance of the entire recognition process is greatly improved.