Selective Spatial Regularization by Reinforcement Learned Decision Making for Object Tracking.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Spatial regularization (SR) is crucial for correlation filter (CF) based visual object tracking, mitigating boundary effects.
• However, SR increases computational complexity and can lead to target loss during occlusion or when targets are near similar objects.

Purpose of the Study:

• To introduce Selective Spatial Regularization (SSR) for CF tracking, aiming for enhanced accuracy, robustness, and speed.
• To address the limitations of existing SR methods in handling occlusions and complex object scenarios.

Main Methods:

• Extended the CF objective function to learn target-context-regularized filters using target-context-driven weight maps.
• Formulated weight map selection as a Markov Decision Process (MDP) problem, solved using reinforcement learning.
• Incorporated a 'not-updating' state in the MDP to enable skipping erroneous filter updates and accelerate tracking.

Main Results:

• SSR significantly boosts tracking accuracy and robustness compared to traditional SR methods.
• The proposed method achieves considerably faster online tracking speeds.
• Empirically validated the effectiveness of SSR by integrating it with three popular SR-CF trackers across five benchmark datasets.

Conclusions:

• SSR offers a superior alternative to conventional SR in CF-based visual object tracking.
• The reinforcement learning-based approach effectively manages filter updates for improved performance and efficiency.
• SSR demonstrates broad applicability and significant performance gains in challenging tracking scenarios.