Adaptive Boosting Classifier for Pedestrian Attributes Identification with Color and Texture Features

Combination of color and texture features along with adaptive boosting classifier is proposed in this paper for attributes identification of the pedestrian images. Color histogram and Multi Block Local Binary Pattern (MB-LBP) is used as a feature extraction method, since both methods do not require the same size of pedestrian images. This is suitable with the dataset that is trained and tested in this research, PETA dataset since this dataset has various size, a point of view, resolution, etc. Color histogram and MB-LBP capture two kinds of information in the pedestrian images, i.e., color and texture information. The Adaboost is used as a classifier for the attribute identification, and for the weak classifier, the decision stump is applied. The performance accuracy of texture, color, and the combination of both features, are compared in the attribute identification process. In the conducted experiments, the number of bins for the color histogram, and MBLBP are 8, 16, and 32. The results showed that the combination of both features, color histogram and MB-LBP, achieved the highest accuracy.