Radius-Aligned Training and Rotated IOU Metrics for Pedestrian Detection in Top-View Fisheye Images

Top-View fisheye cameras offer a broad field of view and reduced occlusion in many surveillances scenarios. However, for the problem of pedestrian detection in such, one main challenge arises from the rotated orientations of people in such images. This leads to the issue of how to specify bounding boxes and, subsequently, how to compute the IOU (intersection-over-union) between the bounding boxes. While this is important during both the training and evaluation of detection models, different researchers have followed different approaches regarding this issue. This paper provides a more comprehensive study comparing the various approaches for specifying bounding boxes and computing IOUs. We experimentally demonstrate their importance in pedestrian detection in top-view fisheye images, including improvements of detection performance over state-of-the-art results on a large dataset.