The Research Of Human Detection Based On Histogram Of Orientation Gradient

Human detection is an important subfield of Computer Vision and Pattern Recognition, it has widespread application in intelligent video surveillance, content-based image/video retrieval/annotation, assisted living/driving, digital entertainment/living, advance human-computer interaction and human motion analysis, etc. However, detecting humans in images/videos is a challenging task owing to their wide range of poses, variable appearance, and the complicated environment/background, illumination. In addition, the different viewpoints of viewer also improve the difficulty. The results of The Pascal Challenge from 2005 to 2009 and the recent research indicate that sliding window classifiers are presently the predominant method being used in object detection, or more specifically, human detection, due to their good performance.For the sliding window detection approach, each image is densely scanned from the top left to the bottom right with sliding windows (mostly are rectangle windows) in different scales pixel by pixel. For each sliding window, certain features such edges, image patches, wavelet coefficients and color are extracted and fed to a classifier, which is trained offline using labeled training data. The method introduced by us belongs to this kind. Recently, methods basing on histogram of orientation gradient are proved to be good for human detection. We dedicated to the research of it. In order to improve its performance, we worked on the selection and training process of classifier, occlusion handling, multiple detection results merging.For classifier, we have tried two kinds of classifiers, Support Vector Machine and AdaBoost. SVM has been wildly used for object detection and recognition for the past decades, because of its good generalization ability and performance. They find a separating hyperplane that maximizes the margin/gap between the object class and non-object class in either the input feature space or a kernelled version of this. Based on the study of the classification scores of the linear SVM classifiers, we introduce a method to handle the occlusion. AdaBoost algorithm was developed from the boosting algorithm. It combines a collection of weak classifiers to form a stronger one. In vision, it is used particularly to build cascades of pattern rejecters, with at each level of the cascade choosing the features most relevant for its rejection task. Although AdaBoost cascades are slow to train, owing to their selective feature encoding, they offer significant improvement (compared to SVMs) in the run-time of final detectors. For the multiply detection results merging, we introduce a method based on Mean Shift, it calculates the accurate positions of human body from the multi-results in 3D positions and scale space.