Research On Human Motion Posture Recognition Based On Multi-feature Fusion

With the development of a new era of artificial intelligence,computer vision plays an increasingly important role in people's life and work.As a hotspot of computer vision,human motion posture recognition has been widely used in video monitoring,smart home,motion analysis and other fields.However there are many limitations in practical application because of the complex and changeable posture in human body movement.Therefore,the accurate description of the motion posture through the effective characteristic information is the key factor to improve the recognition effect of human motion posture.First of all,this thesis analyzes four kinds of classical background modeling method,and this four background modeling methods are used for moving target detection experiments.ViBe algorithm is found to have better accuracy and real-time performance in this moving target detection experiment.In practice,when the motion target appears in the first frame,ghost shadows will be generated in the ViBe algorithm.To solve the problem of ghost shadow,the improved ViBe algorithm combining the Wronskian function is proposed in this thesis,which also reduces the noise interference,and could get a clear and complete goal of human movement.Secondly,based on multi-feature fusion,the human motion posture model is constructed.In this thesis,a method based on multi-feature fusion is proposed to describe the characteristics of motion posture,which is used to obtain more abundant characteristic information.The eight star model,Hu moment,Zernike moment and Wavelet moment are selected as feature descriptor of human motion posture.The extracted feature information is optimized and fused through the genetic algorithm,and the fitness function is constructed based on the mean variance ratio to select the feature information with better separability among multiple classes.The line graph is drawn with the fusion feature,it is found that the fusion feature reduce redundancy and ensure inter-class separability and intra-class stability.In addition,the global and local feature information are complemented by extracting SIFT features.Finally,building the classifier baesd on deep learning requires a large number of training samples and high hardware requirements.K-NN classifier has a heavy computation in the classification.Therefore,this thesis chooses support vector machine as classifier,which has good performance in small sample,and constitutes a multi-classifier in a one-to-one way to realize the recognition of human motion posture.This thesis adopts the standard video database and self-built video database to experiment.Video is pretreated tosolve the problem of blank video occupation resource.The multi-classifier constructed in this thesis is used for the human motion posture recognition test,the experimental results show that the human motion posture recognition based on multi-feature fusion algorithm can effectively improve the recognition accuracy and perform well in practical applications.