Three-Dimensional Multiobject Tracking Based on Voxel Masking Encoder and Deep Hashing Paradigm.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous driving systems require precise 3-D multiobject tracking (MOT) for safety.
• Existing MOT methods struggle with long-distance objects, partial occlusions, and similar object categories.

Purpose of the Study:

• To develop an advanced 3-D MOT framework addressing current limitations.
• To enhance tracking accuracy and reliability in complex driving scenarios.

Main Methods:

• Proposed a 3-D MOT framework integrating a voxel masking encoder (VME) and a deep hashing paradigm (DHP).
• Implemented a near-to-far voxel feature processing strategy for global contextual information.
• Utilized DHP for category discrimination and distance optimization matching (DOM) for precise associations.

Main Results:

• The VME-DHP framework demonstrated superior tracking performance on the KITTI dataset.
• Achieved higher tracking accuracy compared to state-of-the-art methods.
• The DOM method improved efficiency and precision in object association.

Conclusions:

• The proposed 3-D MOT framework effectively handles challenges in autonomous driving.
• The VME-DHP approach offers a significant advancement in 3-D object tracking accuracy.
• This work contributes to safer and more robust autonomous vehicle perception systems.