Randomized Gaussian Posterior Ratio-based Ensemble for Human Detection

Abstract

In this study, we use probabilistic generative model to explore values that represent how close a given window is human or nonhuman. We believe that the ratio of the two values obtained from the model gives good information to identify and score object in detection problems. Although the true ratio of the window is unknown, we assume that the ratio can be estimated by averaging of observed posterior ratios obtained from multiple generative models, which is constructed by using bagging. The prior for the posterior is determined by searching prior space with our random greedy search method. In this paper, we use a transformed HOG feature vector as the input of the model. A transform function is obtained by using PCA algorithm on a bootstrap sample. Experimental results show that the proposed method achieves better performance on human detection than the previous works.