Object detection in outdoor environments is a challenging task. One is not only confronted with the problem of acquiring a sufficient amount of training images, but also the issue of huge variation in the objects appearance due to changing weather and light conditions. When using appearance-based object detection algorithms, such as in this paper, dimensional reduction of input data is an integral component to reduce computational costs and improve reliability. Based on the probabilistic classification method of Gaussian classifiers this paper examines the effect different dimensional reduction approaches have on the classification performance of thermal infra-red object images with respect to incomplete training data. It is shown that in the detection task at hand, which is to find the rear end of a truck in a thermal infra-red image, a reduction approach that combines principal component analysis (PCA) and linear discriminant analysis (LDA) is less sensitive to missing data.

No abstract