Abstract: A scene model and statistic learning based method for pedestrian detection in complicated real-world scenes is proposed. A unique characteristic of the algorithm is its ability to train a special cascade classifier dynamically for each individual scene. The benefit is that the classifier only focuses on the differences between the positive samples and the limited negative samples of each individual scene, thus greatly reduces the complexity of classification, and achieves robust detection result even with a few classifiers. A highly efficient weak classifier selection method and a novel boosting architecture are presented to speed up feature selection and classifier training. To evaluate the proposed algorithm, we captured pedestrian videos under different weathers, seasons and camera motions, and labeled 4300 positive samples. Moreover, a real-time pedestrian detection system named as background modeling and Adaboost training (BMAT) was developed, which produced fast and robust detection results as demonstrated by extensive experiments performed using video sequences under different environments.