MaskUKF: An Instance Segmentation Aided Unscented Kalman Filter for 6D Object Pose and Velocity Tracking

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Accurate 6D object pose and velocity tracking is crucial for robotic manipulation.
• Existing methods often require expensive ground truth annotations or lack real-time performance.

Purpose of the Study:

• To propose a real-time 6D object pose and velocity tracking algorithm named MaskUKF.
• To evaluate MaskUKF's performance against state-of-the-art methods on a benchmark dataset.
• To demonstrate the benefits of joint pose and velocity tracking for closed-loop control.

Main Methods:

• Combines deep object segmentation networks with depth information.
• Utilizes a serial Unscented Kalman Filter (UKF) for tracking.
• Evaluated on the YCB-Video pose estimation benchmark and iCub humanoid platform in simulation.

Main Results:

• MaskUKF achieves state-of-the-art performance on the YCB-Video benchmark, often surpassing existing methods.
• The algorithm does not require expensive ground truth pose annotations during training.
• Closed-loop control experiments show improved precision and reliability compared to one-shot pose estimation.

Conclusions:

• MaskUKF offers an efficient and accurate solution for real-time 6D object pose and velocity tracking.
• The method reduces the need for extensive training data, making it more practical.
• Joint pose and velocity tracking enhances robotic manipulation task performance.