ADA-Track++: End-to-End Multi-Camera 3D Multi-Object Tracking With Alternating Detection and Association.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Query-based 3D Multi-Object Tracking (MOT) methods often use tracking-by-attention or tracking-by-detection paradigms.
• Tracking-by-attention entangles detection and tracking queries, while tracking-by-detection lacks synergy between detection and association.
• Existing methods present sub-optimal performance due to limitations in query entanglement or task synergy.

Purpose of the Study:

• To introduce ADA-Track++, a novel end-to-end framework for 3D MOT from multi-view cameras.
• To combine the strengths of tracking-by-attention and tracking-by-detection paradigms.
• To improve the efficiency and accuracy of 3D Multi-Object Tracking.

Main Methods:

• Developed a learnable data association module with edge-augmented cross-attention, utilizing appearance and geometric features.
• Introduced an auxiliary token in the association module to mitigate attention normalization issues.
• Integrated the association module into a DETR-based 3D detector's decoder layers for simultaneous detection and association.

Main Results:

• The proposed framework enables alternating query refinement for detection and association tasks, harnessing task dependencies.
• Evaluated on the nuScenes dataset, demonstrating superior performance compared to previous paradigms.
• The edge-augmented cross-attention and auxiliary token effectively improved tracking accuracy.

Conclusions:

• ADA-Track++ offers a significant advancement in 3D Multi-Object Tracking by effectively integrating detection and association.
• The novel framework demonstrates the benefits of synergistic task learning in query-based tracking.
• Future work can explore further optimizations and applications of this integrated approach.