A Fast Adaptive Multi-Scale Kernel Correlation Filter Tracker for Rigid Object.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Deep learning models offer high tracking accuracy but struggle with real-time performance on limited hardware.
• Kernel Correlation Filter (KCF) trackers are fast but fail to adapt to scale changes and occlusion, leading to template drift.

Purpose of the Study:

• To develop a real-time object tracking algorithm for complex scenes on resource-constrained platforms.
• To improve the robustness of Kernel Correlation Filter trackers against scale variations and occlusions.

Main Methods:

• A fast multi-scale kernel correlation filter tracker incorporating a scale pyramid for scale adaptation.
• An adaptive template updater using Mean of Cumulative Maximum Response Values (MCMRV) to mitigate template drift during occlusion.

Main Results:

• The proposed tracker effectively adapts to target scale changes while maintaining high operational speed.
• The MCMRV-based adaptive template updater significantly reduces template drift in the presence of occlusion.
• Experimental results show superior performance compared to state-of-the-art kernel correlation filter methods on various datasets.

Conclusions:

• The proposed fast multi-scale kernel correlation filter tracker offers an effective solution for real-time object tracking in challenging environments.
• The adaptive template updating mechanism enhances tracker robustness, making it suitable for embedded systems with limited computational resources.