文摘
Our proposed methods can be viewed as a simplified one-stage CNN, which can also achieve state-of-art results. We proposed four different types of pixel neighborhood differential features, which aim to mine the discriminative information of the intrinsic structure for the pedestrian. We proposed an unsupervised feature pattern learning method, which can reduce the redundancy of the feature and discover discriminative differential statistic patterns. We proposed a supervised feature pattern learning method, which is utilized to get more compacted and informative feature for pedestrian detection.