Pedestrian detection based on adaboost algorithm with a pseudo-calibrated camera

Abstract

This paper presents a new algorithm for pedestrian detection for a fixed camera using the cluster boosted tree (CBT) structure of Wu and Nevatia for building a multi-view tree classifier based on edgelet features. The main advantage of this structure is that it is less sensitive to camera view changes compared to the cascade structure of Viola and Jones. The approach presented in this paper uses geometrical information in the image to estimate pedestrian size for a given pixel position. This we call pseudo camera calibration. Thereby, we combine the CBT classifier trained on the INRIA datasets and the pedestrian size estimator to detect pedestrians. The performance of this algorithm is also evaluated on images captured at a real metro station for several camera positions.