Abstract
A Support Vector Machine (SVM) is an effective method for pedestrian detection applications; however, performance of an SVM is closely related to the samples that are used to train it. An SVM classifier trained by samples from well-known pedestrian datasets such as INRIA and MIT is observed to have limited detection capability in practical environments. In this paper, a statistical background-foreground extraction approach is proposed that autonomously generates samples containing pedestrians in real scenes, in order to diversify the basic training set of the SVM. Comparative experiments have shown that the SVM classifier's discriminability and adaptability to a new scene are greatly enhanced by utilizing extracted samples from that scene in the training stage. Here, a pedestrian tracker that combines a Camshift tracker and a Kalman filter is adjoined to the pedestrian classifier; the tracker is proved to be robust against pose and scale changes, abrupt direction of motion changes, and occlusions, in several test scenes.