Video Human Detection Research Based On Gaussian Mixture Model

The human detection in surveillance video is a key technology of the promotion andapplication of the intelligent city, which also is currently one of the hot spots in computervision. This paper studies the human detection technology in surveillance video, which hasimportant theoretical and practical significance for the protection of public transportationand reduction of the accident rates.Firstly, this paper analyzes the research situation of human detecion technology.Because of surveillance video image quality is generally poor, a lot of the video imagepreprocessing methods have been studied to improve the video quality and facilitatesubsequent image analysis, such as image denoising and illumination compensationmethods.Secondly, this paper studies state-of-the-art moving target detection method based onthe video image and focuses on moving object detection algorithm based on Gaussianmixture algorithm. Considering the research scene and the deficiency of classic Gaussianmixture model, this paper proposes an improved Gaussian mixture algorithm withdynamically adjusting the number of Gaussian mixture model and the selection of initialvalues, and adaptive learning rate, which has improves moving target detection algorithmgreatly.Again, this paper studies human recognition algorithm using the detected movingtarget, and mainly analyzes the human detection algorithm based on statistical methods.After analyzing several features for human recognition, the pyramid gradient histogramcharacteristics(Pyramid Histogram of Oriented Gradients, PHOG characteristics) and itscalculation method are mainly discussed. This paper proposes a human detection methodbased on PHOG features and SVM (support vector machine), which is described in detail.Finally, this paper builds a human detection system for video surveillance. The systemdraws on the advantages of both background modeling and statistical learning.Experimental results show that the system has better detection performance in complexscenes with shadows, uneven illumination, and multi-target recognition.