Reseaech On Human Behavior Recognition Algorithm

The recognition of human behavior in static images is one of the popular research directions in image processing.The correct recognition of human behaviors contributes to the classification and retrieval of images,video surveillance and human tracking.In this paper,10 kinds of human behaviors in Imagenet database are selected to study the human behavior recognition algorithm.Firstly,the behavior recognition theory based on shallow feature is studied,and the behavior recognition algorithm based on distributed representation of pose and appearance.The human body is divided into the pose position,and the image area with similar structure is found by the distance measurement method as the data positive sample,and the classification model is trained.The algorithm is compared with the classical gradient direction histogram and the deformable part model,and the algorithm of behavior recognition based on distributed behavior image region has achieved better classification results.Secondly,according to the visual perception mechanism of the human body,a human behavior recognition algorithm based on semantic significant graph is proposed.Through the combination of sliding window and similarity measure,find the behavior area which can show the semantic features of the image,that is,the semantic significant region.The semantic significant region and original images are used as double input source in the research of human behavior recognition.This research enhances the utilization of salient regional information in the image,shows the recognizable region of the image clearly and promotes the analysis of human behaviors.Thirdly,the research put the different parts of human body into different channels according to the characteristics of the non-rigid movement of the human body.Thus,the behavioral semantics of the human body is described in a more detailed way.Through the analysis of recognition in different channels,the 5-channel behavior recognition algorithm is proposed.According to the corresponding division principle of different channels,the D-Imagenet database is formed.CNN feature extraction and SVM classifier training are carried out on five channels and three levels.The algorithm details the characteristics and effectively solves the problem of similarity between category and category of internal differences in human behavior recognition.