Combining Adaboost with a Hill-climbing Evolutionary Feature Search for Efficient Training of Performant Visual Object Detectors

This paper presents an efficient method for automatic training of performant visual object detectors, and its successful application to training of a back-view car detector. Our method for training detectors is adaBoost applied to a very general family of visual features (called “control-point” features), with a specific feature-selection weak-learner: evo-HC, which is a hybrid of Hill-Climbing and evolutionary-search. Very good results are obtained for the car-detection application: 95% positive car detection rate with less than one false positive per image frame, computed on an independant validation video. It is also shown that our original hybrid evo-HC weak-learner allows to obtain detection performances that are unreachable in reasonable training time with a crude random search. Finally our method seems to be potentially efficient for training detectors of very different kinds of objects, as it was already previously shown to provide state-of-art performance for pedestriandetection tasks.