Abstract
Recently, an image feature description calculating a co-occurrence between some pixels which are allocated at local region has been attracted an attention as the method which shows effectiveness in an object detection. However, there are some problems. A calculation cost and the number of feature dimensions of those methods tend to increase exponentially with respect to a feature description method focusing on single pixel. This paper proposed a multiple resolution CoHOG (MRCoHOG) feature description method which is able to suppress these exponential increases into a linear increase without incurring reduction in the classification accuracy. MRCoHOG can reduce the number of a feature dimension by calculating the co-occurrence only between the adjacent pixels, and maintain the classification accuracy by extracting a feature from plural low resolution images. In our classification experiments using the vehicle data set cropped from surveillance images of parking area and “INRIA person data set”, these results showed that the performance of MRCoHOG is equivalent to CoHOG's one.
