Efficient vehicle detection and orientation estimation by confusing subsets categorization

Aerial traffic surveillance requires algorithms that can accurately predict the locations and orientations of hundreds of vehicles in a large high resolution aerial image within seconds. Under this constraint, the classical cascaded detection framework based on boosting algorithms still remains an optimal choice. These methods, however, usually use many binary classifiers to enhance the localization performance resistant to orientation variances, which is not effective in distinguishing confusing orientations and subsets. This paper categorizes these confusing subsets automatically by analyzing the correlations between specific orientation angles and location deviations at local detection window regions, makes robust predictions on them by N-nary multi-class classifiers. This helps to reduce the required number of classifiers to less than half and improve both localization and orientation estimation accuracies, making it potential for additional speed optimization.