People Detection In Still Images

People detection has become a hot topic in computer vision and patternrecognition in recent years, and it has extensive application foreground in intelligentvideo surveillance, human-computer interaction and driver assistance system. Up tonow, various algorithms based human detection have been proposed, but because ofthe complex backgroung, non-rigid body, mutual occlusion, detection velocity ofpeople and detection accuracy ect, it does not form consummate human detectionalgorithm.The architectural form consists of feature extraction,classifier training andhuman detection. In this paper, we choose INRIA Person Dataset,which is recognizedby researchers, in this dataset,people dressed differently and they had differentpostures in different shooting scenes, most of images are of very high resolution,sothe dataset is also a relatively high degree of difficulty. The static image detectionbased sliding window is often used. Firstly,images are scaled many times by sameratio, and then the same size sliding windows scan images in fixed step from left toright. When scanning window, feature vectors should be extracted and extraction andclassification results are forecasted by the classifier, all results which are labeled1are saved, finally the results are fused.In this paper, we analysis the shortcoming it is that hog feature are of highdimensional and SVM takes many time to train sample. During the feature extraction,we study PHOG and LBP which can effectively describe people, but a single featureextraction method has its limition. PHOG is based on HOG, it focused on therepresentation of spatial information and shape information, the images are ofpyramid structure,in each layer image is characterized by gradient orientationhistogram. However, PHOG is sensitive to noise, when it is used to describe imageswith complex background, the positive images are often labeled0. LBP operator ischaracterized by the kind of texture description within grayscale,and local textureinformation of gray image can be described and extracted. Uniform LBP caneffectively reduce noise jammer, because Non-Uniform LBP mode include somenoise region, LBP features can effectively compensate anti-disturbance of PHOG.Based on this, a new kind of skin feature extraction is proposed by color histogram,we combine it with PHOG and LBP to form a new kind of feature, it can highlight skin color information and enhance the ability to identify people. In this paper, wechoose extreme learning machine method to detect people. Compared with SVM,ELM does not need to adjust the input weights and the bias, it can generates a uniqueoptimal solution, it has high learning speed, well generalization and high stability.The experiment shows that the algorithm not only can improve the accuracy,but alsoeffectively improve the training speed of feature extraction and classifification.