Feature Detection And Action Recognition Of Moving Human Body In Video Based On Improved HOG

In today’s field of computer vision,the detection of human targets in dynamic video and the action recognition of the human body has become an important research direction.It has important application in intelligent monitoring,video retrieval and human motion analysis.However,due to the non-rigid nature of the human body and the complexity of the video background,the field is still a challenging research direction.Video movement in the human body action recognition process is divided into three processes,namely,human target detection,human target feature extraction,human behavior identification.In the beginning of the birth of the gradient direction histogram(HOG)algorithm is used for human target detection,not in the human behavior recognition,this paper combines HOG algorithm with other algorithms in human behavior recognition.This paper aims to identify the human behavior by improving the hybrid Gaussian background model and the improved HOG algorithm.The main actions of seven human behaviors are as follows: running,walking,bending,jumping on one hand,waving with both hands,to identify,and in recognition to achieve a better recognition effect.The main work of this paper is as follows:In the detection of human target,firstly,several basic images preprocessing methods are described,and only the input image can be pretreated to get better input image.We introduce the application of image enhancement,image denoising and mathematical morphology in image.By image denoising can remove the noise that can influence the results of the image,through the mathematical morphology processing,eliminating the isolated pixels in the image and the holes in the human target.This paper introduces several commonly used background modeling models,and uses the mixed Gaussian model by comparing their advantages and disadvantages.However,although the model can effectively obtain the background in the video,the method is influenced by the shadow in the image,Therefore,this paper presents an improved mixed Gaussian model,which is a hybrid Gaussian background model based on HSV space and shadow suppression method,which can effectively suppress the shadow by converting the original image into the image in HSV space and passing the value on V vector.In the aspect of human body feature extraction,the center of gravity feature is first proposed,and the center of gravity vector is obtained by extracting the contours of human body and calculating the center of gravity of human contour.Two kinds of histograms,including gradient histogram(HOG)and optical flow field histogram(HOF),were studied.By using two histograms,a single histogram cannot show the human movement feature in video,so this paper presents an improved HOG algorithm,that is based on HOG,HOF,center of gravity characteristics of the human body feature extraction method.In the aspect of human behavior recognition,this paper introduces several commonly used methods of action recognition classification,namely template matching method,direct classification method and state space method.Through the study of several algorithms,this paper chooses the KNN algorithm in direct classification.And the human behavior identification based on HOG,HOF and center-of-gravity features is realized.Six different experiments are carried out on seven different human behavior behaviors in Weizmann database by different combinations of three different characteristics.The correctness of the improved HOG algorithm proposed in this paper is 95.42%.