Research On Extended Histogram Of Gradients-Based Human Detection And Pose Estimation

Pedestrian detection is a major concern direction frontier and research hotspot in the field of computer vision.Pedestrian Detection and Pose Estimation mainly through intelligent analysis from an image sequence or video stream for the automatic detection of the pedestrian,at the same time estimated the pedestrian location and direction of the respective components in the image information.Pedestrian Detection and pose estimation with a very wide range of applications in intelligent vehicle safety systems, intelligent monitoring, motion analysis,human-computer interaction and so on.In this thesis, pedestrian detection based on the extended gradient histogram features of the image, then using the human detector based on the detection algorithm, at last,using constraint-based tree Pictorial Structure appearance model algorithm to estimate the human pose.The main research work are summarized as follows:(1)Firstly,mainly expounds the pedestrian detection and pose estimation algorithm processes.as well as for pedestrian detection classification SVM classifier, but also describe the tree Picture Structure model in detail.(2)Pedestrian detection plays an important role in human pose estimation,especially pedestrian detection,providing the body location information in the image.This thesis discusses the gradient histogram(HG) and Histogram of Oriented Gradients(HOG)some restrictions proposed extension Histogram of Oriented Gradients(ExHOG) pedestrian detection algorithm. ExHOG alleviates the problem of discrimination between a dark object against a bright background and vice versa inherent in HG. It also resolves an issue of HOG where by gradients of opposite directions in the same cell are mapped into the same histogram bin.(3) Asymmetry in training sets of humans and nonhumans and high dimensionality of existing features are problems plaguing human detection. Furthermore, the high dimensionality of existing features hampers realtime performance of human detection and classification.In this thesis providing the Asymmetric Principal Component and Discriminant Analyses (APCDA) to solve these problems. Our experimental results show that the proposed ExHoG consistently outperforms the standard HG and HOG for human detection.,while improving the detection accuracy.(4)Aims at the problem that the human body appearance model of human parts is vulnerable to background interference in the body pose estimation algorithm of the tree Pictorial Structure model, In order to improve the human body appearance model, we puts forward body pose estimation algorithm of appearance model which is based on priori segmentation and the appearance transfer mechanism. According to PS model, using the human body detector and foreground highlighting by preprocessing, it can find an approximate position and size of the human body, meanwhile remove the background clutter, We can estimate the appearance model of the human body parts based on a priori segmentation and appearance transformation mechanism. Experiments show that using the algorithm of the human body detector and foreground highlighting can not only reduce the search space of the components, but also improve the accuracy of the body pose estimation.