Automatic Compensation of Radial Distortion by Minimizing Entropy of Histogram of Oriented Gradients

Abstract

A car-mounted camera for driver's assistance has a wide angle view, but at the same time, it also has a serious radial distortion. This paper presents a method which can automatically estimate the distortion parameters without using any specially-made patterns for calibration. Our method uses the fact that we are surrounded by many artificial objects consisted of straight lines, e.g., buildings, signboards, and telephone poles, when we are driving. Although these straight lines become curved lines on the camera image because of the distortion, it is easily expected that the appropriately compensated image has the most straight lines. In order to quantify the amount of straight lines, we introduce the entropy of Histogram of Oriented Gradients (HOG) over the whole image. The entropy of HOG is expected to become minimum when the image has the most straight lines. Using this property, the distortion parameters are estimated. The experimental results show that the estimated distortion parameters generate appropriately undistorted images.