Trajectory And Convolution Neural Network Based Human Action Recognition

In recent years,with the rapid development of internet technology and computer vision,human action recognition has been widely concerned.Human action recognition has a wide application prospect in augmented reality,intelligent security,autonomous vehicles and other fields.Therefore,this paper aims to recognize human actions.Although the study of human action recognition has made many achievements with the development of the deep learning.Because of the complexity of the motion scene,the uneven quality of the data and the large amount of noise in the human action video data,it can not be effectively analyzed and identified by the existing human action recognition algorithm.The core process of human action recognition is: feature extraction,recognition and classification.The traditional feature extraction method utilizes the expert prior knowledge to extract the feature,and uses the unsupervised learning to extract the feature descriptor from the video data.However,the feature descriptor extracted by the trad itional method can not describe the whole action process comprehensively.Deep learning is similar to the human brain learning process,it can build the supervised end-to-end model to identification and classification through the neuron interconnection.Although the abstract feature extracted by deep learning has strong discrimination,the network structure is complex,computation overhead is huge,and it will produce a lot of redundancy calculation.Therefore,this paper focuses on the combination of traditional methods and the deep learning their own advantages of feature extraction and action classification.The main contributions of this paper are as follows:1.We propose a human action recognition method based on trajector y and convolution neural network.This method makes use of the supervised feature extraction of convolution neural network to analyze human action video,and combines the traditional expert and computer vision prior knowledge to propose the method of trajectory constrained convolution layer feature extraction to solve human action recognition problem.The method not only utilizes the powerful feature extraction ability of convolution neural network,but also enhances the discrimination and robustness of the feature by using the traditional expert prior knowledge to strengthen the feature learning process.2.We propose a feature extraction method based on local stacked fisher vectors for human action recognition.The traditional fisher vector coding method can only extract the global information of the feature space,while the temporal and spatial domain of the human action video samples exist both global structure information and local motion information.Therefore,we propose a local stacked fisher vectors feature extraction method based on the traditional fisher vector coding.The method can extract the characteristics of local and global information,and can describe the movement process more accurately and improve the recognition of human action ability.3.We propose a human actio n recognition method based on deep neural network.Because of the human action data characteristics of small sample,high feature dimension and sparse motion information,the existing model is difficult to learn the hyperplane to identify and classify a human action video sample.We propose a deep neural network model,which is based on the learning of a deep network model with strongly linear and non-linear feature combination ability to extract a large number of high-order abstract features,and thus enhance the performance of human action recognition.