Siamese Networks-Based People Tracking Using Template Update for 360-Degree Videos Using EAC Format.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• 360-degree videos offer rich information but suffer from geometric deformation and large data volumes, hindering real-time tracking.
• Existing trackers struggle with non-uniform geometric deformation from sphere-to-plane projection and data size challenges.

Purpose of the Study:

• To develop a robust and efficient people tracking system for 360-degree equi-angular cubemap (EAC) videos.
• To address challenges of geometric deformation, content discontinuity, and partial occlusions in EAC video tracking.

Main Methods:

• A Siamese networks-based people tracker utilizing template update for EAC videos.
• Face stitching to resolve EAC content discontinuity and avoid new geometric distortions.
• A Bayes classifier-based template update timing detector using linear discriminant features and score map statistics.

Main Results:

• The proposed tracker significantly enhances tracking accuracy for fully convolutional Siamese networks (SiamFC) on EAC format videos.
• The system achieves real-time performance, operating beyond the frame acquisition rate.
• The novel score map-based template update detector outperforms existing methods.

Conclusions:

• The developed Siamese networks tracker with template update offers a robust solution for people tracking in 360-degree EAC videos.
• The proposed Bayes classifier-based update detector effectively handles geometric deformation and occlusions, improving tracking reliability.