Warping Approach for Rearview Pedestrian Detection with Fish Eye Cameras

Abstract

The trade-off between distortion-handling and keeping a wide field of view (FOV) is a problem affecting pedestrian detection systems of automobile rearview monitors that use on-vehicle fisheye cameras. We propose an image warping approach, called horizontal panorama mosaicing (HPM), as a solution to the problem. HPM is designed to (1) recover collinearity in the vertical direction and (2) create a sweeping panorama in the horizontal direction, such that it effectively normalize the appearances of vertically long objects like pedestrians in the full horizontal FOV. Using HPM warping as a preprocessing of the image input, conventional detectors (e.g. Faster r-cnn (FRCN) and DPM) can be used for rearview pedestrian detection without losing a wide horizontal FOV. In our evaluations, log-average miss rate of pretrained FRCN were reduced from 80.1 to 24.6 percents in the marginal view angle range using HPM. Our method is simple enough to be easily translated into a look-up-table representation, so it also has an advantage when it comes to its implementation in on on-board real-time systems.