Discriminative leaf based Hough Forest for vehicle detection

Abstract

This paper introduces a discriminative framework for the task of vehicle detection based on Hough Forest. The leaf nodes in Hough Forest framework are not discriminative enough, which means that they do not have the ability to classify whether the test patches ended up in each leaf are positive or negative. Hough votes are assigned to all test patches by Hough forest, including negative test patches, which will introduce a large number of noise votes, even lead to false alarms. Furthermore, most of test patches are negative (background), and negative patches do not contribute informative Hough votes. Aggregating voting information from all patches extracted at each pixel location is really time-consuming. Thus we have developed a framework to classify whether the test patches are positive or not, and only positive test patches (object patches) participate in Hough voting, which will not only reduce noise votes but also increase detection efficiency. We demonstrate that discriminative leaf node framework improves the results of Hough Forest framework and achieves desirable performance on UIUC car datasets.