Mutation detection dual correlation filter with an object-awareness module for real-time target tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Discriminative Correlation Filter (DCF) methods are prevalent in visual tracking.
• Existing trackers like AutoTrack have limitations in filter degradation and boundary effects.
• There's a need for improved DCF methods with broader applications and better performance.

Purpose of the Study:

• To develop an advanced visual tracking method that overcomes limitations of current DCF trackers.
• To introduce a novel tracker, the mutation detection dual correlation filter with an object awareness module (MDDCF-OAM).
• To enhance robustness, appearance modeling, and suppress filter degradation in visual tracking.

Main Methods:

• Proposed an object-mask based context enhancer for a robust appearance model.
• Implemented a dual filter training-learning structure to mitigate filter degradation.
• Utilized a refined joint response map and updated Gaussian label map for mutation detection and attenuation.

Main Results:

• The MDDCF-OAM outperformed nine state-of-the-art trackers on benchmarks like OTB2015, UAV123, TC128, and VOT2019.
• Achieved a real-time processing speed of 32 frames per second.
• Demonstrated effectiveness in diverse tracking scenarios, including UAV and camera tracking.

Conclusions:

• The MDDCF-OAM offers superior performance and robustness compared to existing visual tracking methods.
• The proposed tracker is suitable for real-time applications, particularly in unmanned aerial vehicle and camera tracking.
• The innovations in appearance modeling and filter learning contribute to significant advancements in DCF-based tracking.