Research On Human Action Recognition Method Based On Multi-view Characteristics

Human action recognition in video is receiving more and more attention and research,it is widely used in various important fields such as intelligent visual monitoring,human-computer interaction,and annotation or retrieval based on content video.Due to the limitation of human action recognition from a single point of view,such as the partial occlusion of the human body,the light factor and the action recognition accuracy in daily life.This paper proposes a human action recognition method based on multi-view features.In the process of execution,an independent human action representation can be obtained from multiple perspectives.In order to capture the human body's movements at different viewing angles,multiple cameras are used for capturing.The movement of the human body is usually described as a sequence of consecutive human postures.The Region of Interest of human motion can be extracted to obtain the body motion posture vector.Such a two-dimensional image can be obtained by applying image cutting technology to the video.The general movement is represented by a series of continuous action posture vectors.This paper discretized the continuous posture vectors of human movements into N basic posture movement vectors in daily life and uses the K-means clustering algorithm to determine N basic posture of human movements.The vector is divided into categories.The human motion is represented by a series of discrete time sequence image frames.Each frame in the discrete post pose sequence represents a special motion gesture in the motion action process.That is,a complete human body motion passes through the basic motion pattern.The unique combination of the two;then the fuzzy vector quantization processing of the obtained action pose vector to obtain the corresponding action instance of the degree of membership vector.Extreme Learning Machine(ELM)is a relatively fast single hidden layer feed-forward neural network training algorithm.The traditional neural network training algorithm needs to adjust the input weights and offsets of the intermediate hidden layer,while the extreme learning machine the input weights and hidden layer bias values arerandomly selected.The training is to obtain the network output weights of the middle hidden layer.In the action training stage,feature extraction is performed on some marked action instances and input into the extreme learning machine to train the output weights of the hidden layer network.In the action test phase,the unmarked test action video is extracted from the features of the above steps and input into the extreme learning machine after the action training to obtain a human action classification result under a single visual angle.At the same time,the principle of voting for the action classification results under each single angle of view is adopted in multiple perspectives to obtain the final action classification results under multiple angles of view,thereby realizing the classification of human actions from multiple perspectives.This paper will also reduce the data dimension of the motion feature image data input into the extreme learning machine,which can effectively reduce the training time and also optimize the output error of the training stage and the test stage.In this paper,the KTH and UCF50 action datasets are respectively subjected to human action classification experiments under single and multiple angles of view.The effectiveness of the human motion recognition method based on multi-perspective features proposed in this paper in the field of human motion recognition is demonstrated through experiments.The classification results of the recognition algorithm are compared to show the superiority of this algorithm.