An improved algorithm on Viola-Jones object detector

Abstract

In image processing, Viola-Jones object detector [1] is one of the most successful and widely used object detectors. A popular implementation used by the community is the one in OpenCV. The detector shows its strong power in detecting faces, but we found it hard to be extended to other kinds of objects. The convergence of the training phase of this algorithm depends a lot on the training data. And the prediction precision stays low. In this paper, we have come up with new ideas to improve its performance for diverse object categories. We incorporated six different types of feature images into the Viola and Jones' framework. The integral image [1] used by the Viola-Jones detector is then computed on these feature images respectively instead of only on the gray image. The stage classifier in Viola-Jones detector is now trained on one of these feature images. We also present a new stopping criterion for the stage training. In addition, we integrate a key points based SVM [2] predictor into the prediction phase to improve the confidence of the detection result.