Histogram of Radon Projections: A new descriptor for object detection

One of the important requirements for a good object detector is a set of robust visual features. These features extracted from the reference images containing the desired object instance will be used to identify the objects from the test images. In this paper, we propose a new feature set for object detection, called the Histogram of Radon Projections (HRP). To compute this feature descriptor, the image is first divided into smaller cells and for each cell, the Radon transform values are calculated for different orientations and weighted votes for each transform coefficient are accumulated into bins. These bin values are block-normalized and collected together to get the final descriptor. We use this descriptor for car detection using gray-scale images and pedestrian detection using RGB images. The performance of this descriptor is compared with that of HOG and it is found that the new descriptor performs better for both gray-scale and RGB images.