A Dataset of Annotated Omnidirectional Videos for Distancing Applications.

Area of Science:

• Computer Vision
• Robotics
• Surveillance Technology

Background:

• Omnidirectional (360°) cameras capture spherical views, offering comprehensive environmental data.
• These cameras have significant potential in video surveillance, research, and industry.
• Accurate distance estimation from 360° imagery is crucial for many applications.

Purpose of the Study:

• To introduce the CVIP360 dataset, an annotated collection of 360° videos for distance estimation.
• To present a new, uncalibrated method for estimating object distances from single 360° images.
• To validate the proposed method's efficacy for distancing applications.

Main Methods:

• Development of the CVIP360 dataset with 16 indoor/outdoor videos, including pedestrian bounding boxes and ground truth distances.
• Formulation of a geometry-based distance estimation algorithm utilizing omnidirectional camera acquisition principles.
• The algorithm requires only camera height as input, making it practically uncalibrated.

Main Results:

• The CVIP360 dataset provides valuable annotated data for 360° vision research.
• The proposed distance estimation algorithm demonstrated negligible error when tested on the CVIP360 dataset.
• Empirical results confirm the algorithm's suitability for real-world distancing applications.

Conclusions:

• The CVIP360 dataset is a significant resource for advancing 360° video analysis.
• The novel uncalibrated distance estimation method is accurate and practical for surveillance and other applications.
• This work contributes to the effective utilization of omnidirectional cameras in various fields.