Learning Rich Feature Representation and State Change Monitoring for Accurate Animal Target Tracking.

Area of Science:

• Computer Vision
• Animal Behavior Analysis
• Machine Learning

Background:

• Animal tracking is vital for ecological research but faces challenges due to video data acquisition difficulties and unpredictable animal movements.
• Existing methods struggle with maintaining accuracy when animal appearance or motion changes significantly.

Purpose of the Study:

• To develop a novel and robust animal tracking method that overcomes limitations of current approaches.
• To enhance the accuracy and reliability of animal tracking in complex, real-world scenarios.

Main Methods:

• A new animal tracking method based on correlation filters is proposed.
• The approach integrates hand-crafted (e.g., Histogram of Oriented Gradients) and deep features, combined with temporal context information for a rich target representation.
• Adaptive model updating using temporal context and robust features addresses significant changes in animal state.

Main Results:

• The proposed method demonstrates improved tracking performance compared to mainstream algorithms on self-constructed animal datasets.
• Fusion features, selected based on oscillation degree, effectively capture appearance and motion characteristics.
• Adaptive updating significantly enhances re-tracking accuracy after target state changes.

Conclusions:

• The novel correlation filter-based animal tracking method offers a promising solution for accurate and effective animal behavior research.
• The integration of diverse features and adaptive model updating provides robustness against challenges in animal tracking.
• This approach facilitates more reliable monitoring of migration, habitat selection, and behavior patterns.