Pedestrian detection via fusing LBP and GSS features

Abstract

In the field of pedestrian detection, HOG feature and LBP feature are very important. Due to the symmetry of the appearance of pedestrian, this paper proposes a pedestrian detection method based on symbiotic CoLBP features and GSS features. Firstly, we calculate the pairwise gradient self similarity GSS between the local gradient blocks of the image, and get the the symbiotic CoLBP features according to the LBP features. Secondly, the FGM (Feature Generation Machine) is used to remove the irrelevant GSS features, and then the discriminant gradient self similarity (DGSS) feature is obtained. Finally, the performance of pedestrian detection is evaluated by two cascaded classifier, classifier of the first level uses linear SVM which is trained based on HOG feature and symbiotic CoLBP feature to remove the most negative samples which is easy to distinguish; for the classifier ofthe second level, considering the HOG characteristics is the premise to generate the GSS/DGSS feature, so we use the calculated HOG features again to generate the corresponding GSS/DGSS features, and then use the Real-AdaBoost classifier which is trained based on HOG and DGSS feature to detect the true positive in the regions of the candidate image to achieve complete pedestrian detection. The experimental results show that the proposed method is superior to the current advanced target detection methods.