Pedestrian Detection with Geometric Context from a Single Image

摘要

We address the problem of pedestrian detection in still images. Many current pedestrian detection systems limit their performance by ignoring the underlying 3D geometric context in the image. We can estimate the geometric context by learning appearance-based method. We propose a novel context feature and local integrable features. These features are used for building many candidate weak classifiers by using linear SVM. Finally, MPL-Boost method selects the best weak classifiers suited for detection and construct the rejector-based cascade detector. We provide a thorough quantitative evaluation of our method on TUD-Brussels dataset and demonstrate that it outperforms the state-of-the-art pedestrian detector in recall rate, meanwhile, shows faster speed than other context incorporation method.