Online object tracking with sparse prototypes

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Online object tracking is complex due to appearance variations from intrinsic and extrinsic factors.
• Existing methods struggle to effectively model these appearance changes.
• Developing robust appearance models is crucial for accurate tracking.

Purpose of the Study:

• To propose a novel online object tracking algorithm utilizing sparse prototypes.
• To enhance appearance modeling by integrating Principal Component Analysis (PCA) with sparse representation.
• To improve tracking accuracy and reduce drift in challenging scenarios.

Main Methods:

• Developed a novel algorithm for object representation using sparse prototypes with L1 regularization in PCA reconstruction.
• Implemented online learning for sparse prototypes to adapt to appearance changes.
• Introduced a method to mitigate tracking drift by accounting for occlusion and motion blur during model updates.

Main Results:

• The proposed algorithm demonstrates effective learning of appearance models.
• Sparse prototypes explicitly account for both data and noise.
• The method successfully reduces tracking drift by considering occlusion and motion blur.

Conclusions:

• The novel online object tracking algorithm with sparse prototypes performs favorably against state-of-the-art methods.
• The integration of L1 regularization into PCA enhances appearance modeling.
• The approach offers a robust solution for tracking in challenging image sequences.