DATaR: Depth Augmented Target Redetection using Kernelized Correlation Filter.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Correlation filter trackers, such as Kernelized Correlation Filter (KCF), offer real-time performance by leveraging image properties like circulant structure, avoiding large datasets.
• However, these trackers struggle with occlusions and out-of-view scenarios, limiting their practical application.
• Existing methods often lack robustness in dynamic or complex environments.

Purpose of the Study:

• To develop a novel RGB-D Kernel Correlation tracker for enhanced target re-detection.
• To address the limitations of existing trackers in handling occlusions and out-of-view situations.
• To create a more robust tracking system that adapts intelligently to avoid boundary issues.

Main Methods:

• A novel RGB-D Kernel Correlation tracker was designed, integrating depth information for improved target representation.
• The framework incorporates a target re-detection mechanism specifically for challenging scenarios.
• Experimental validation was performed using a standard dataset and real-time evaluation with a Microsoft Kinect V2 sensor.

Main Results:

• The proposed tracker successfully re-detects targets in challenging scenarios, demonstrating improved robustness.
• The system effectively adapts to avoid boundary issues during tracking.
• Experimental results validate the tracker's performance on both benchmark datasets and real-world sensor data.

Conclusions:

• The novel RGB-D Kernel Correlation tracker significantly enhances target re-detection capabilities.
• This approach offers a more robust solution compared to traditional Kernelized Correlation Filter (KCF) trackers, particularly in adverse conditions.
• The work lays the foundation for developing more effective and resilient kernel-based correlation filter trackers.