Large-Scale Omnidirectional Person Positioning.

Area of Science:

• Computer Vision
• Robotics
• Surveillance Systems

Background:

• Existing person localization methods primarily use perspective cameras, limiting their field of view and practicality.
• Overhead fisheye cameras offer a wider field of view, lower cost, and anti-occlusion benefits for person positioning.
• Research in overhead fisheye person localization is hindered by a lack of large-scale, diverse datasets.

Purpose of the Study:

• To introduce LOAF, the first large-scale dataset for overhead fisheye person detection and localization.
• To develop and validate a novel fisheye person detection network for accurate overhead positioning.
• To stimulate further research in fisheye camera-based person localization.

Main Methods:

• Creation of the LOAF dataset, featuring 457K annotated pedestrian bounding boxes with ground-truth orientation and location.
• Development of a fisheye person detection network leveraging rotation-equivariance and distortion-awareness.
• Implementation of training strategies for accurate, radius-aligned human box and angle predictions.
• Calculation of actual person locations using a numerical solution based on the fisheye model and camera altitude.

Main Results:

• The LOAF dataset provides extensive diversity in scenes, human poses, densities, and locations.
• The proposed fisheye detector significantly outperforms previous methods on the LOAF dataset.
• The complete fisheye positioning solution accurately locates all persons within 0.5m in under 0.1 seconds.

Conclusions:

• The developed fisheye person detection and localization system is highly effective for overhead camera applications.
• The LOAF dataset is a valuable resource for advancing research in fisheye-based person positioning.
• This work demonstrates the potential of fisheye cameras for efficient and accurate real-world person localization.