Foreground Gating and Background Refining Network for Surveillance Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection in surveillance videos is crucial for traffic control and public security.
• Existing methods struggle with false positives and misalignments, degrading performance.
• Surveillance object detection (SOD) requires robust handling of static and dynamic scenes.

Purpose of the Study:

• To propose a novel framework, FG-BR Net, for enhanced surveillance object detection.
• To address critical challenges in SOD, including false positives in background regions and object misalignments.
• To improve the accuracy and reliability of object detection in complex surveillance environments.

Main Methods:

• Developed a Foreground Gating and Background Refining Network (FG-BR Net).
• Introduced a background subtraction module with a feedback connection to preserve static and moving objects.
• Incorporated a background refining stage using pairwise non-local operations for accurate localization.

Main Results:

• FG-BR Net significantly reduces false positives in background regions.
• The framework achieves improved object localization accuracy.
• Demonstrated competitive performance on real-world traffic surveillance benchmarks, topping UA-DETRAC subsets.

Conclusions:

• FG-BR Net offers a robust solution for surveillance object detection.
• The proposed methods effectively tackle key challenges in SOD, outperforming existing approaches.
• The framework shows state-of-the-art results on challenging datasets without complex additions.