A probabilistic, multivariate approach for object recognition in thermal infra-red images

Abstract

For any task that autonomous vehicles may encounter in unstructured outdoor environments a reliable vision system is a key point for success. That is especially true with an autonomous convoy, where each vehicle has to track and follow the one in front. When applying a multivariate based approach for object detection, dimensional reduction of processed data is a vital part of any algorithm. Based on probabilistic classification into two classes (positive and negative) three different approaches for dimensional reduction are examined in this paper: The first method transforms new images in two reduced principal component analysis (PCA) spaces, derived from negative and positive training images respectively. The second approach calculates a mutual PCA space from all training images and the third strategy uses linear discriminant analysis (LDA) for data reduction. In these reduced spaces image classification is done with the Gaussian classifier. Through experiments it is shown that classification in the mutual PCA and the LDA space result in fewer errors and a more reliable class assignment. Furthermore, the use of LDA is more robust if confronted with incomplete training data. Finally it is shown that a confidence approximation using Gaussian processes can, if trained, identify positive and negative images and evaluates untrained images with the appropriate uncertainty.