Research On Human Behavior Recognition Algorithm Based On Deep Learning

In recent years,the field of human behavior recognition has become a research hotspot in the field of deep learning computer vision.Human behavior recognition refers to the category of human behavior in a video determined by a machine according to algorithm rules,which is another challenge for a machine to have visual discrimination.If the machine has the ability of human behavior recognition,it will be regarded as a basic function of robot,thus extending to other fields of research,such as robot behavior imitation,behavior detection and so on.Due to the rapid development of deep learning and the significant performance of convolutional neural network in picture problems,some methods use two-dimensional convolutional neural network to extract video features for human behavior discrimination.However,two-dimensional network structure cannot make good use of video temporal information,and some methods propose to use three-dimensional network structure to process video samples,but the existing three-dimensional network model has a large number of parameters,which is difficult to train.How to make use of the advantages of 3D structure and reduce the number of network parameters is a problem to be solved at present.First of all,the ITDNet-2D network is proposed based on the two-stream convolutional network and the Dense Net network,which aims at how to utilize the motion information in video and effectively fuse it with the appearance information of video.The network features add the motion features of video clips represented by difference between video frames or optical flow features,and the fusion module is designed to add these features to the main network for behavior recognition.The fusion module in this paper is flexible and can change its position in the network.Moreover,this module does not participate in the structure of the two branch networks and is specifically responsible for controlling the amount of motion information,so it is beneficial to optimize the weight of network parameters.In the second place,the ITDNet-3D network is established on the basis of ITDNet-2D,because ITDNet-2D ignores the time-domain feature of video,and extracts the features in the form of picture collection,that is,multiple pictures were processed separately.The convolutional neural network model uses the temporal and spatial characteristics of video clips to recognize human behavior.In order to maintain the spatial-temporal relationship of video samples and not destroy their internal relations,3D convolution and pooling are used to extract feature information.In general,3D networks often result in a large number of parameters,while the ITDNet-3D network in this paper greatly reduces the number of parameters without affecting the network performance.Finally,the experiment and analysis of the two networks on the UCF101 benchmark show the effectiveness of ITDNet models in human behavior recognition.