Research And Implementation Of Video Action Recognition System Based On Multi-feature Fusion And Integrated Learner

Video recognition is one of the important research fields in computer vision,and it has important engineering application value in the current society.In recent years,with the development of network hardware and software and the trend of popular culture,more and more video data is circulating on the network,and these data can almost be described by massive amounts.For many video data platforms,it is impossible to rely on manual management for proper management.These data need a good video management system to automate its operation.The research of video motion recognition system is to solve this kind of problem.Motion recognition is a hot research direction in video recognition.Motion recognition in video is easy to be affected by lighting,scale and other factors.Building a good video motion recognition system is a problem that has been studied by the scientific community.In the current research,the scheme adopted by the mainstream video motion recognition system is mainly based on the concept of "extracting motion features of video objects + constructing classification models by machine learning".A good feature extraction scheme and an excellent machine learner can guarantee the ability of the finally generated video motion recognition model.In the existing video motion feature extraction methods,there are video-based and single-frame image-based feature extraction methods,and the features extracted by these methods are used to express the motion information of objects in the video.For the extracted motion features and categories of video objects,machine learning is used to learn,and then a powerful video motion recognition model can be generated.Among machine learning methods,neural network is a commonly used deep learning method in recent years,in which convolutional neural network is a powerful network structure proposed on the basis of neural network.Integrated learning is a new machine learning method proposed in recent years,and its purpose is to make up for the shortcomings of the weak learning ability of traditional machine learning methods.In this paper,through a lot of literature review and experimental testing,we studied the common video recognition methods used in different environments,analyzed a variety of action feature extraction methods used in different environments,and combined them in depth to propose a new type of video action Feature extraction method,and through experimental design,the integrated learning system of extremegradient lifting tree is used in the field of video motion recognition to build a good video motion recognition model.The main work of this article is as follows:(1)Aiming at a noisy video environment,a spatiotemporal dual-stream feature fusion scheme based on histogram of directional gradient,optical flow direction information map and convolutional neural network is proposed.The scheme extracts these two features based on single-frame images and video streams.In space,it extracts the direction gradient histogram feature of the action based on the single-frame RGB image where the action is.As the spatial feature of the current action in the video;in time,based on the optical motion flow image from the initial frame of the action video to the current frame,the convolutional neural network is used to extract its features as the time feature of the current action.Finally,these two features are regarded as the motion features of the object in the video,and the machine learning method is used to generate the action recognition model.Through the simulation experiment,the effectiveness of the video action recognition system based on the fusion feature method in the noisy video environment is finally proved.(2)Aiming at the small noise video environment,the feature extraction method based on motion history image is used and improved,and a multiple motion history image method is constructed.The video motion stream extracts multiple motion history images as the shallow features of the action,and on this basis,the Hu moment feature and the Zernike moment feature of the geometric invariant moment features of the image are extracted as the deep feature of the action,and then the two deep layers are fused Feature and use multiple machine learning methods for simulation experiments.The experimental results prove that the improved multiple motion history images have better support for geometric moment features compared to the motion history images,and also prove the effectiveness of the video motion recognition system constructed based on the fusion feature method in a small noise video environment.(3)A specific XGBoost integrated classification system is constructed to support the action fusion features of video extraction in two noise environments with different intensities.Combined with the two fusion feature schemes proposed above,two different video action recognition systems were constructed,which acted on large-noise video data and small-noise video data,respectively.These two systems mainly use the built XGBoost integrated learning system to generate the final video action recognition model.Through simulation experiments,the model generated by this classification method is compared with themodels generated by several other machine learning methods.The test results show that the final video action recognition model of XGBoost integrated learning system constructed in this paper achieves91.667% recognition accuracy on the test samples of KTH,97.000% on some test samples of UCF101,and performs well on the time efficiency of model construction,which proves the integration of XGBoost The method has a good application prospect in the field of video motion recognition.