A comparison of FFS+LAC with AdaBoost for training a vehicle localizer

Abstract

This paper describes our recent work on identifying leading vehicles in the context of Forward Collision Warning (FCW) application. Specifically, we aim at detecting and localizing leading vehicles in videos that are captured by a forward-facing camera mounted in a moving host vehicle. To achieve that goal, we propose to seamlessly extend the AdaBoost-based object detection framework beyond Haar features, by integrating in the HOG (Histograms of Oriented Gradients) features. Our experimental results show that we can effectively optimize the training of the vehicle detector, by using a large bank of HOG plus Haar features within the AdaBoost framework. Our approach can also significantly reduce the number of features required for achieving a given accuracy, while the cost of such detector with more complex training can still remain tractable by using an FFS+LAC training scheme. © 2011 IEEE.