Pedestrain Detection Based On Hybird Local Binary Pattern For Fast Feature Pyramid

Pedestrian detection technology becomes more and more important to analyze semantic object in the scene. It has been widely applied to various fields, like smart car, intelligent traffic monitoring system and so on. In reality, due to the diversity of pedestrian pose, pedestrian size, and the interference caused by changing of the light intensity, there are still many problems that need to be solved in the pedestrian detection technology.Training a classifier with strong distinguishing ability is one of the key problems for pedestrian detection. In typical methods, different pedestrian descriptors such as histogram of oriented gradient (HOG), local binary pattern (LBP), and SIFT features are used to obtain the pedestrian classifier with a certain degree of detection ability. Although classifiers based on these features have high detection rate, their operational efficiency cannot meet requirement of most practical applications.Thus, in this thesis we propose an image coding of binary pattern mode for transform features of images into binary image coding. The proposed approach transforms complicated calculation process into a relative simple calculation process to get strong distinguishing ability feature maps. Meanwhile, we also adopt the method of fast feature pyramids for efficient feature calculation.The main research contents of this work are as follows:1) Thesis put forward a new process of feature solving. After comparing the details of HOG feature and the LBP feature, we found that on the premise of keeping direction of gradient orientation, it is still possible to improve the system operation efficiency. Therefore, we introduce a new feature, the BPG feature, to accelerate the feature calculation process. In order to divide the gradient orientation into 8 bin range and sum up the corresponding gradient values in different gradient directions bin, the mean value of each 8 orientation bin is used as the threshold value. By comparing with the threshold value, the BPG features are obtained by the corresponding two value coding. The experimental result demonstrates that the new feature has a strong pedestrian distinguishing ability.2) The BPG feature and LBP feature were combined to form a feature pool which was used in the process of training and classification. As the pyramid technique is used in image processing, each image has been zoomed in and zoomed out several times. After an image was zoomed out to a certain level, the BPG feature might become worse and its recognition ability can be affected. In this case, the LBP feature can capture some missing information of BGP to enhance the accuracy and efficiency of feature extraction.3) Thesis uses dynamic threshold Adaboost algorithm to train the strong classifier. The problem of uncertainty of random selection of threshold can be avoided by computing dynamic threshold for each level of strong classifier. Therefore, we can find the optimal threshold that meets the requirement of strong classifier training.4)Thesis, the feature calculation process of each zoom scale images is improved. We use the neighboring power exponent law to calculate the estimated parameters between two adjacent zoom images. The feature map of each zoom scale images are computed based on a fast parameter calculation formula. As a result, the running time of computing feature for each zoom images has been reduced significantly, which leads to a much efficient pedestrian detection system.