Hypothesis verification based on classification at unequal error rates

We examine the classification of object candidates which are preselected by an automatic segmentation algorithm. The selected candidates are either searched objects (e.g. different traffic signs) or known garbage patterns (e.g. other round objects) or also arbitrary patterns never seen before, since the closed world assumption generally made in classification theory is often violated in practice. Our aim is to keep the false positive rate as low as possible, allowing for a rather high fraction of missed relevant objects. We present two classification approaches, one is a local approximator, namely an RBF network, the other one is a polynomial classifier using global approximation. The RBF net is adapted by bootstrapping and uses dimensionality reduction to yield fast classification cycles. The polynomial classifier is adapted by balancing the classes via moment matrices and uses a reject criterion in the decision space. For real-time traffic sign recognition we achieve a false positive rate of less than 0.5 percent at a rate of 5 percent of traffic signs rejected as garbage, which is tolerable since the overall decision is not made frame per frame but for the whole sequence while passing a traffic sign