Learning multi-regularized mutation-aware correlation filter for object tracking via an adaptive hybrid model.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Discriminative Correlation Filters (DCF) are effective for object tracking but struggle with appearance mutations, filter degradation, and target distortion.
• Existing DCF trackers face performance decreases due to these challenges, necessitating improved robustness.

Purpose of the Study:

• To introduce a novel Multi-Regularized Mutation-Aware Correlation Filter (MRMACF) to address limitations in DCF-based object tracking.
• To enhance tracker resilience against appearance changes, filter degradation, and target distortion.

Main Methods:

• Developed a mutation-aware strategy with an adaptive hybrid model and mutation threat mechanism to handle appearance mutations and filter degradation.
• Implemented an improved sparse spatial feature selection incorporating row/column methods to address target distortion.
• Introduced a surrounding-aware approach to leverage context information and prevent filter deviation.

Main Results:

• The MRMACF approach demonstrated superior performance compared to modern trackers on benchmark datasets.
• Achieved the highest performance on the OTB-2015 dataset with a DP score of 93.2% and an AUC score of 69.8%.

Conclusions:

• The proposed MRMACF approach significantly improves object tracking performance by effectively mitigating appearance mutations, filter degradation, and target distortion.
• MRMACF offers a robust and efficient solution for challenging object tracking tasks.