Unified detection and tracking of instruments during retinal microsurgery

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Traditional object tracking methods are divided into detection-based and local optimization-based approaches.
• Detection-based tracking excels at handling target appearance/disappearance but neglects motion continuity.
• Local optimization is efficient and accurate but struggles with target loss and re-initialization.

Purpose of the Study:

• To bridge the gap between detection-based and local optimization-based tracking methods.
• To propose a unified framework for simultaneous object detection and tracking.
• To develop a robust solution for scenarios with frequent target appearance and disappearance.

Main Methods:

• A novel framework for unified detection and tracking is proposed, conceptualized as a time-series Bayesian estimation problem.
• The approach treats detection and tracking as a sequential entropy minimization problem.
• Integration of the Active Testing (AT) paradigm with Bayesian filtering enables robust performance.

Main Results:

• The developed framework effectively integrates detection and tracking capabilities.
• The method demonstrates robustness in handling objects that regularly enter and leave the field of view.
• Experimental validation on a retinal tool tracking problem confirms efficient and robust tracking.

Conclusions:

• The proposed unified framework offers an efficient and robust solution for object detection and tracking.
• This approach successfully addresses limitations of existing methods, particularly in dynamic environments.
• The integration of Active Testing and Bayesian filtering provides a powerful paradigm for sequential estimation problems.