A New Diverse Dataset and Comparison of Algorithms for Pedestrian Detection in Top-View Fisheye Images

Abstract

Pedestrian detection in images has been extensively researched, however existing detectors designed for perspective images usually fail to detect people on top-view fisheye images due to the various appearances of people. In this paper, we establish a new diverse fisheye dataset which consist of indoor and outdoor fisheye images from public and private datasets. We also adapt three types of spatial transformation to make the visual look of the pedestrians as upright as possible and four commonly used algorithms for pedestrian detection without retraining of the detector models. In addition, we analyze the pedestrian detection results with different conditions to figure out the reason of the results.