Research on Gradient Local Binary Patterns Method for Human Detection

I-Abstract Human detection is a key problem in computer vision, which is widely used in image analysis, intelligent vehicle and visual surveillance. However, the task of human detection is rather challenging because of high variations of clothing, pose, occlusion, scale and illumination. In human detection systems, feature extraction and learning method are two important parts and hot research topics. The target of this dissertation is to develop robust feature extraction algorithms that encode image regions as high dimensional feature vectors that support high accuracy human/non-human decisions. In recent years, Histogram of Oriented Gradients (HOG) (N.Dalal, CVPR2005) is an efficient method in human detection as a standard. And local pattern based methods have attracted increasing interest in object detection and recognition systems. Recognition, 1996) is widely used in texture classification, face detection and recognition. But the original definition of LBP is not suitable for human detection because of bigger human size. Some LBP based improved methods are proposed for better detection results than LBP (Semantic Local Binary Patterns (S-LBP) by Y.Mu, CVPR2008, and Histogram of Template (HOT) by S.Tang, ICASSP2010). The HOG method only use the values of 4 neighbor pixels to calculate the edge direction and gradient value, which make HOG not powerful to classify some curves with different radians. Secondly, HOG and HOT methods are disturbed from the noise. Thirdly, the gradient values of pixels are voted to the histogram as weight in HOG, but a fixed weight value are voted in S-LBP and HOT. The parts with big gradient values always play important part in the classification. In this dissertation, an intra combined feature extraction method named Gradient Local Binary Patterns (GLBP) with gradient and texture information is proposed. The experimental results on INRIA Dataset show that, GLBP Linear Detector achieves the best detection rate (92% Hit Rate in 10-4 False Positive Per Window (FPPW)) Abstract-II-compared with other linear detectors, and GLBP-Polynomial GLBP Linear Detector achieves the best detection rate (95.7% Hit Rate in 10-4 FPPW) compared with other non-linear detectors. Secondly, A patterns cutting algorithm for reusing the lost information from non-uniform patterns, and a new formula for gradient value calculation using binary values of LBP are proposed in this dissertation to make GLBP method efficient in Radial basis function (RBF) kernel for high detection rate cases. The experiment result shows the developed GLBP-RBF detector using these two additional parts can get the best detection rate (96.7% …