Introducing confidence maps to increase the performance of person detectors

Abstract

This paper deals with the problem of computational performance of person detection using the histogram of oriented gradients feature (HOG). Our approach increases the performance for implementations of person detection using a sliding window by learning the relationship of sizes of search windows and the position within the input image. In an offline training stage, confidence maps are computed at each scale of the search window and analyzed for a reduction of the number of used scales in the detection stage. Confidence maps are also computed during detection in order to make the classification more robust and to further increase the computational performance of the algorithm. Our approach shows a significant improvement of computational performance, while using only one core of the CPU and without using a graphics card in order to allow a low-cost solution of person detection using a sliding window approach. © 2011 Springer-Verlag.