Decision-Level Fusion for Vehicle Detection

This paper deals with the problem of decision-level fusion for vehicle detection from gray-scale images. Specifically, the outputs of some classifiers are simply ”distances”, that is, they represent ”distance measurements” between a query pattern and a decision boundary. We argue that the distance component is very helpful for decision fusion. Unfortunately, some of the most popular statistical decision fusion rules, such as the Sum rule and Product rule, do not take advantage of the ”distance” property. Even worse, these rules make assumptions about data independence and distribution models which do not hold in practice. Motivated by these observations, we propose a simple decision-level fusion rule in the context of vehicle detection. Our fusion rule takes advantage of ”distance” information and does not make any assumptions. We have applied this rule on a vehicle detection problem, showing that it outperforms well known statistical fusion rules. Key–Words: Decision fusion, Haar wavelet transform, Gabor filter, Support Vector Machines