Research And Design Of Dynamic Human Detection System Under Static Background

The human body detection, identification and tracking technology has been notonly a hot issue on the field of computer vision and security, but also a wide range ofintegrated topics. In addition, the detection of human body is the prerequisite andbasis to identify and track human body, and the accuracy of the human bodydetection has some direct relationship with the performance of identification andtracking. Generally, the human body detection should divide to the detection underdynamic background and under static background according to surveillance cameramotion or not. Under a dynamic background, the background maybe movedsignificantly with the moving camera; under a static background, it is essentiallyunchanged with some factors change such as noise and illumination change. In thispaper, we consider the research and design of dynamic human detection systemunder static camera, that mainly including the following aspects: preprocessing ofthe input frame sequence, moving target extraction and moving object recognition.The system firstly preprocess the input frame sequence, including noisefiltering, image gray and image enhancement, and so on, and then extract themoving target from the image frame sequence after preprocessing. Under staticbackground, there are time difference, background subtraction and optical flowmethod to extract the moving target. In this paper, based on the comparing amongthe three methods we proposed the detection method combining the time differencemethod and background subtraction method adaptively, in which the time differencemethod using three adjacent color pixels difference and consider the impact ofillumination change, and the background subtraction method using an improvedGaussian mixture mode(GMM). The combination of the two detection methods canimprove the detection accuracy without increasing the computational complexity.After extracting the moving target from moving frame sequence, we shouldrecognize the target with some biological characteristics follow that namely humanidentification. Taking into account the background may appear multiple movingtargets, so we should carry out image segmentation firstly to separate targets andidentify every moving target orderly. The identification process include featureextraction and classifier recognition step. As the non-rigid nature of human walkingand some factors, such as camera angles can cause it difficult to obtain fixedbiological characteristics, so we use the body's head and shoulder parts that arestable relatively as characteristics of human identification. In the classifier section,the design based on the adaptive integration of the Support Vector Machine model and Bayesian model, and this two-classification method based on different thinking:the SVM based on the structural risk minimization, while the Bayesian model basedon the relationship of probability optimization. The experiment results show that theclassification accuracy of this fusion classification is super to both independentclassification methods.