People detection in top-view fisheye imaging

We address the problem of people detection in top-view fisheye imaging. Even within the same top-view fisheye frame, upright people appear slanted in various directions and are distorted in different ways. Due to this variability, standard people detectors are not directly applicable to top-view fisheye frames, and dedicated people detectors for the top-view fisheye domain are hard to design. We extract features in the fisheye frame, unwrap them to a perspectivelike feature map, and forward them to a people detector designed and trained for common perspective images. Extracting the features before unwrapping prevents harmful smoothing of the gradient information. To facilitate feature unwrapping, we employ dense feature extraction and compute the unwrapping Jacobian. Distortion-free unwrapping is known to be impossible. We optimize the unwrapping method for the explicit goal of people detection performance. Applying a tunable fisheye camera model to project the fisheye image plane onto a unit half sphere, followed by the stereographic map projection, we obtain people detection performance similar to the standard perspective case. We complete the solution by introducing a convenient purposive fisheye-camera calibration process, optimized for subsequent people detection performance.