Improved training algorithm for tree-like classifiers and its application to vehicle detection

We propose a new training algorithm for tree classifiers and cascades for object detection and compare it to a standard algorithm for cascade training. Our experiments show that the proposed algorithm significantly reduces the number of features needed per stage by incorporating the output of the previous stage as a weak learner into the next stage. This approach also speeds up the classification while maintaining the same detection accuracy. The analysis of the features selected by the algorithm provides further insights into its functioning.