Toward Class-Agnostic Tracking Using Feature Decorrelation in Point Clouds.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Single object tracking in 3D point clouds is crucial for autonomous systems.
• Current deep learning trackers struggle with novel object categories not seen during training.

Purpose of the Study:

• To develop a class-agnostic tracking method for LiDAR point clouds that generalizes to unseen object categories.
• To address the performance degradation of existing trackers when encountering new classes.

Main Methods:

• Investigated the class-agnostic performance of state-of-the-art trackers on unseen categories.
• Proposed a feature decorrelation method to constrain fused features and maintain generalization.
• Ensured search region consistency and distinctiveness between foreground and background points.

Main Results:

• The proposed feature decorrelation method significantly improves class-agnostic tracking performance.
• Demonstrated considerable improvements compared to advanced trackers like P2B and BAT on KITTI and NuScenes datasets.
• Effectively handles tracking of unseen objects by reducing spurious feature correlations.

Conclusions:

• Feature decorrelation is key for robust class-agnostic tracking in LiDAR point clouds.
• The developed method offers a more generalizable solution for real-world 3D object tracking applications.
• This work advances the capability of autonomous systems to track diverse and previously unencountered objects.