Object detection based on several samples with trained Hough spaces

Abstract

A new method is proposed to handle several samples object detection case. Firstly, a new feature at each pixel, named Locally Adaptive Steering (LAS), is derived which contains the local contour and gradient variation information. Then, the cell sizes of Hough space are trained based on several samples, which represent the tolerance of appearance at each pixel location, where the Hough space is constructed by the image coordinates and the ranges of the feature values at the corresponding positions. Next, one sample image is randomly chosen as the query image and the patches with the same size of the query image are taken from the target image by sliding window. The similarities of these matching are based on the posterior decision rule in Hough space and the histogram distance. In the end, the mean shift method is applied to the similarity map to localize the instances of the object. © 2012 Springer-Verlag.