Robust object tracking with reacquisition ability using online learned detector.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Long-term visual tracking is complex due to appearance variations like illumination, motion, and occlusions.
• Existing tracking methods often struggle with robust reacquisition after target loss.

Purpose of the Study:

• To develop an adaptive visual tracking approach capable of handling appearance changes.
• To enable robust reacquisition of a target after tracking failure or drift.

Main Methods:

• Utilized a condensation-based method integrated with an online support vector machine (SVM) for adaptive visual tracking.
• Implemented a cascade detector using random ferns for real-time target re-detection.
• Introduced a refinement strategy involving template matching to remove incorrect support vectors and improve accuracy.

Main Results:

• The proposed tracking approach demonstrated adaptability to significant appearance changes.
• The cascade detector provided robust and real-time re-detection capabilities.
• The refinement strategy effectively improved tracker performance post-reacquisition.

Conclusions:

• The novel tracking method offers robust and adaptive long-term visual tracking.
• The combination of online SVM, random ferns detector, and refinement strategy significantly enhances tracking performance and reacquisition.
• The approach shows encouraging results on challenging benchmark datasets.