Research On Human Behavior Recognition In Video

Recognition of human behavior in video is an important research area in computer vision,which has a significant impact on the improvement of people's quality of life.Due to the diversity of backgrounds in the detection environment,changes in brightness,and the diversity of the clothing,there are high requirements for the human behavior recognition robustness,stability,and real-time performance.This article mainly studies the method of human behavior recognition in video under the fixed conditions of the camera.The main tasks are as follows:1.The principle of the original HOG feature is studied.HOG feature has many shortcomings such as large computational complexity and low recognition rate.Combined with motion region detection method,a human behavior recognition method based on 3D-HOG feature is proposed.The experimental results show that the human behavior recognition method based on 3D-HOG features has good robustness in illumination conversion and the computational efficiency has been significantly improved.2.The recognition rate of human behavior recognition in video will be greatly affected by the image's translation,rotation,affine changes,scale changes,and image brightness changes.In this paper,SIFT algorithm is studied based on Gaussian difference pyramid theory.Because it has scale invariance and strong robustness,when extracting features in dynamic video sequences,we need to introduce the third dimension(time).According to the experimental results,it shows that 3D-SIFT features strong stability and good matching performance.3.The characteristics of the human behavior recognition method in video need to combine the characteristics of small amount of calculation,strong robustness and high recognition rate.Based on the theory of HOG and the optical flow estimation method,we extract intensity and direction histogram characteristics.For the defects that the optical flow is sensitive to background interference and susceptible to illumination,we weight the pixels in the optical flow to enhance its robustness.According to the experimental results,the optical flow intensity and direction histogram features have strong robustness and low computational complexity.