Pedestrian Detection by Boosting Soft-Margin SVM with Feature Selection

We present an example-based algorithm for detecting objects in images by integrating component-based classiflers, which automaticaly select the best local-feature for each classifler and are combined according to AdaBoost algorithm. The system employs soft-margin SVM for base learner, which is trained for all localfeatures and the optimal feature is selected at each stage of boosting. The proposed method is applied to the MIT CBCL pedestrian image database, and shows fairly good result with 10 local-features such as full image, upper half, lower half, right half, left half, head, right arm, left arm, legs, and body center.