Abstract
This paper investigates the effectiveness of probabilistic decision-level fusion in the context of vehicle detection from gray-scale images. Specifically, we demonstrate that probabilistic fusion rules often disregard the "physical" meaning of classifier outputs and make assumptions that might not hold in practice. As a result, their performance is seriously affected. To support our argument, we have trained two Support Vector Machine (SVM) classifiers to perform vehicle detection using two types of features: Harr Wavelet and Gabor. To classify an input pattern, the output of the SVM classifier needs to be thresholded first. In this case, each output represents a class. When considering the raw SVM outputs (i.e., without thresholding), however, each output represents a "distance" between the input pattern and the decision boundary. Unfortunately, some popular probabilistic decision-level fusion rules, such as the Sum and Product rules, disregard the physical meaning of the raw SVM outputs. Moreover, they make assumptions about data independence and distribution models which might not hold in practice. Motivated by these observations, we propose a simple but effective decision-level fusion rule which exploits the physical meaning of the SVM outputs and does not make any assumptions about the data. We have evaluated the proposed rule on real data sets, showing that it outperforms traditional probabilistic fusion rules.