Semantic-Aware Real-Time Correlation Tracking Framework for UAV Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Discriminative Correlation Filter (DCF) trackers offer high efficiency but struggle with performance degradation in Unmanned Aerial Vehicle (UAV) tracking scenarios.
• Existing methods often fail to effectively handle background clutter and target-irrelevant areas in complex aerial scenes.

Purpose of the Study:

• To present a novel semantic-aware real-time correlation tracking (SARCT) framework designed to enhance DCF tracker performance for UAV videos.
• To improve tracking accuracy and robustness without a significant increase in computational cost.

Main Methods:

• SARCT integrates a detection module for Region of Interest (ROI) proposal generation and filtering of irrelevant areas.
• A semantic segmentation module, utilizing semantic template generation and coefficient prediction, refines ROI masks to suppress background interference.
• Feature and network layer sharing between object detection and semantic segmentation optimizes the framework for real-time performance.

Main Results:

• SARCT significantly improves the accuracy of conventional DCF-based trackers in UAV tracking tasks.
• The proposed framework demonstrates superior performance compared to state-of-the-art deep trackers on multiple aerial datasets.
• Experiments validate the effectiveness of SARCT in suppressing background interference and maintaining precise ROI masks.

Conclusions:

• The semantic-aware real-time correlation tracking (SARCT) framework offers a robust solution for object tracking in UAV videos.
• SARCT effectively addresses the limitations of traditional DCF trackers in challenging aerial environments.
• The approach achieves real-time performance while enhancing tracking accuracy through semantic information integration.