Incorporating contextual information in pedestrian recognition

Local classifiers are often used in automotive pedestrian detection systems. The disadvantage of such systems is that they only regard local image cutouts to discriminate pedestrian class from its background. In those cases where false alarms bear a great resemblance to true positives it is difficult to solve the classification task in that way. As a possible solution this paper presents a general and mathematically founded model which incorporates the pedestrian contextual information in the classification task. Our approach allows the use of any relevant contextual information to improve the detection results. This contribution shows how to define possible contextual hints and how to combine them into a contextual classifier.