Efficient Online Object Tracking Scheme for Challenging Scenarios.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visual object tracking (VOT) is crucial for surveillance, UAVs, and medical imaging.
• Existing VOT methods face challenges like scale variation, occlusions, motion blur, and illumination changes.
• Spatiotemporal context (STC) frameworks offer a basis for robust tracking but have limitations.

Purpose of the Study:

• To propose an improved visual object tracking algorithm within a spatiotemporal context (STC) framework.
• To address limitations of STC, specifically scale variation and target model drift due to occlusions and motion blur.
• To enhance tracking accuracy and robustness in challenging visual environments.

Main Methods:

• A max-pooling-based scale scheme is integrated to overcome scale variation issues in the STC framework.
• An occlusion handling mechanism is developed, utilizing the average peak to correlation energy (APCE) to detect occlusions.
• A fractional-gain Kalman filter is employed for occlusion handling, and APCE criteria are used to adapt the target model against motion blur.

Main Results:

• The proposed algorithm demonstrates significant improvements in tracking performance.
• The method effectively handles scale variations, occlusions, and motion blur.
• Extensive evaluations show superior results compared to various existing tracking methods.

Conclusions:

• The developed STC-based tracking algorithm offers a robust solution for visual object tracking.
• The integrated scale scheme, occlusion handling, and APCE criteria contribute to enhanced tracking accuracy.
• The algorithm shows promise for real-world applications requiring reliable visual object tracking.