Real-time visual detection of vehicles and pedestrians with new efficient adaBoost features

This paper deals with real-time visual detection, by mono-camera, of objects categories such as cars and pedestrians. We report on improvements that can be obtained for this task, in complex applications such as advanced driving assistance systems, by using new visual features as adaBoost weak classifiers. These new features, the “connected controlpoints” have recently been shown to give very good results on real-time visual rear car detection. We here report on results obtained by applying these new features to a public lateral car images dataset, and a public pedestrian images database. We show that our new features consistently outperform previously published results on these databases, while still operating fast enough for real-time pedestrians and vehicles detection.