Vehicle detection using TD2DHOG features

Histogram of oriented gradients (HOG) is often used for object detection in images. These HOG features of images can be referred to as 2DHOG when represented in a 2D matrix format instead of a 1D vector. In this paper, we propose a new vehicle detection algorithm by using 2DHOG in the discrete cosine transform (DCT) domain. The proposed technique consists of extracting 2DHOG from the input image and applying on it 2DDCT. This is followed by a low pass filtering in order to obtain novel features called as transform-domain 2DHOG (TD2DHOG). TD2DHOG is used with a classifier pyramid in order to reduce the multi-scale scanning cost. Experimental results show that the proposed algorithm when applied on two public vehicle detection datasets reduces the storage requirement of the classifier pyramid, while providing about the same performance as that provided by the state-of-the-art techniques.