Enhancing Multi-Camera People Detection by Online Automatic Parametrization Using Detection Transfer and...

Area of Science:

• Computer Vision
• Machine Learning
• Object Detection

Background:

• Optimizing people detector parameters is challenging due to numerous variables and diverse application scenarios.
• Existing methods often require extensive manual labeling or struggle with multi-camera viewpoint variations.

Purpose of the Study:

• To propose an automated framework for adapting and enhancing people detectors in multi-camera environments.
• To identify optimal detection thresholds online for each detector-viewpoint pair without additional ground truth data.

Main Methods:

• Developing a framework to transfer detector results accurately between camera viewpoints.
• Implementing a self-correlation mechanism for transferred results to identify optimal detection thresholds.
• Evaluating the framework's performance across four state-of-the-art detectors: DPM, ACF, Faster R-CNN, and YOLO9000.

Main Results:

• The proposed framework successfully improves the performance of multiple state-of-the-art people detectors.
• Online identification of optimal detection thresholds enhances detector accuracy compared to fixed thresholds determined during offline training.
• The method effectively handles variations in viewpoints from multi-camera setups.

Conclusions:

• The framework offers an effective online solution for optimizing people detector confidence thresholds in multi-camera systems.
• This approach reduces the need for manual annotation in dynamic, multi-view environments.
• The method demonstrates significant performance gains for diverse object detection models.