Boosting Soft-Margin SVM with Feature Selection for Pedestrian Detection

We present an example-based algorithm for detecting objects in images by integrating component-based classifiers, which automaticaly select the best feature for each classifier and are combined according to the AdaBoost algorithm. The system employs a soft-margin SVM for the base learner, which is trained for all features and the optimal feature is selected at each stage of boosting. We employed two features such as a histogram-equalization and an edge feature in our experiment. The proposed method was applied to the MIT CBCL pedestrian image database, and 100 sub-regions were extracted from each image as local-features. The experimental results showed fairly good classification ratio with selecting sub-regions, while some improvement attained by combining the two features, histogram-equalization and edge. However, the combination of features could to select good local-features for base learners.