Three-Dimensional Extended Target Tracking and Shape Learning Based on Double Fourier Series and Expectation...

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Signal Processing

Background:

• Target tracking with unknown shapes is challenging due to model uncertainties.
• Existing methods often rely on predefined, potentially inaccurate, prior models for shape evolution.
• Accurate estimation of 3D kinematics, extent, and orientation is crucial for many applications.

Purpose of the Study:

• To develop a robust algorithm for tracking targets with unknown, fixed 3D star-convex shapes.
• To jointly estimate target kinematics, extent, and orientation from point cloud measurements.
• To circumvent limitations of prior models by directly optimizing shape parameters.

Main Methods:

• Utilizing the Expectation Conditional Maximization (ECM) framework for joint parameter estimation.
• Modeling 3D shape with a radial function via double Fourier series (DFS) expansion.
• Representing orientation using the singularity-free axis-angle method.
• Employing an Unscented Kalman Smoother for kinematic inference (E-step) and regularized cost minimization for shape and orientation estimation (M-step).

Main Results:

• The proposed ECM-based algorithm effectively performs joint estimation of kinematics, extent, and orientation.
• The method demonstrates robustness and smoothness in shape and orientation estimation.
• Experimental evaluations validate the algorithm's effectiveness in tracking unknown 3D star-convex shapes.

Conclusions:

• The developed ECM approach offers a significant advancement in tracking targets with unknown shapes.
• Direct parameter optimization provides a more practical solution compared to methods relying on prior shape evolution models.
• The algorithm shows promise for real-world applications requiring precise 3D target tracking.