2D staircase detection using real AdaBoost

This paper proposes a frontal staircase detection algorithm using both classical Haar-like features and a novel set of PCA-base Haar-like features. Real AdaBoost is used for training a cascaded classifier. The PCA-based Haar-like features are extremely efficient at rejecting background regions at early stages in the cascade. A specifically designed scanning scheme made the algorithm constantly time efficient to different image sizes. An multi-detections integration scheme that is exclusive for staircase detection is extremely useful at further rejecting false positives. A new evaluation metric is proposed to rate each final detection, instead of Boolean classifying it. Experimental results show that the approach can detect staircases accurately at extremely low false positive rate.