Research On Human Detection And Action Recognition Based On Convolution Feature Deformable Part Model

Human detection and action recognition are the research hotspots and key problems in the field of computational vision.It involves many advanced technologies together,such as artificial intelligence,machine learning,pattern recognition and so on.And human detection and action recognition are widely used in intelligent monitoring,human-computer interaction and intelligent robot.However,most of the current object detection methods still have serious problems,such as the use of manually specified features,error and missed detection.Generally,researchers were used to take the body detection and behavior recognition account into two independent tasks for design.However,a complete process of human action recognition is to firstly detect the human object in the image,then extract the action characteristics by analyzing the detected human position,finally classify and recognize actions.Therefore,this paper proposes a new method to integrate human detection and action recognition into a unified framework,which to map the low-level response to generate human action features.Recently,the deep convolution neural network has been widely used in visual tasks due to the advantages of automatic learning of image features,and has made great progress.Therefore,in the present work,the convolution neural network was used in the unified task of human detection and action recognition.First,the application background of the human detection and action recognition was introduced,and development of this subject by foreign researches are summarized,including the relevant theory of convolution neural networks,and the current mainstream of the object detection methods and action classification methods.Secondly,deformable part models was used in human detection in the present study.Because the traditional HOG feature needs to be manually extracted,and it only concerned the characteristics of edge,the computation is of complexity and high cost.It is proposed to replace the HOG feature by using CNN to extract the convolution feature,which realizes the automatic learning of image features,solves the problem that traditional methods are difficult to extract effective features and computational overhead,and which can improve the accuracy of DPM human detection effectively.Finally,because the process of generating a convolution feature pyramid with image pyramid and traditional CNN was a complex image preprocessing,which is less efficient,a multi-scale convolution neural network is proposed to improve the traditional CNN by designing multi-scale modules.The multi-scale convolution feature extracted by this network effectively improves the computational efficiency and improves the performance of DPM human detection.At the same time,the relationship between the whole body and the parts of the human body is analyzed,and the action feature are extracted from the bottom response of the detection information.This feature can effectively describe the attitude information of the human body.This method combines hman detection and action recognition into a unified framework,so as to achieve a set of human detection,attitude estimation and behavior recognition as a whole system.