Learning spatio-temporal representation for cooperative 3D object detection and tracking.

Area of Science:

• Intelligent transportation systems
• Computer vision
• Robotics

Background:

• Multi-agent collaborative perception is crucial for intelligent driving but faces challenges like localization errors, data sparsity, and bandwidth limits.
• Existing methods struggle to balance perception accuracy with communication efficiency.

Purpose of the Study:

• To propose CoTrack, a novel collaborative detection and tracking method for intelligent driving.
• To enhance perception effectiveness while optimizing communication efficiency.

Main Methods:

• Developed a spatio-temporal aggregation module with spatial cross-agent collaboration and temporal ego-agent enhancement.
• Implemented an unsupervised feature compressor to reduce communication volume.
• Designed a two-stage online association strategy for improved detection-track matching.

Main Results:

• CoTrack effectively mitigates feature misalignment from localization errors.
• Compensates for sparse data by leveraging historical ego-agent information.
• Achieved state-of-the-art performance in collaborative 3D object detection and tracking on simulated and real datasets.
• Demonstrated robustness in challenging, noisy environments.

Conclusions:

• CoTrack offers a robust solution for multi-agent collaborative perception in intelligent driving.
• The method successfully balances perception accuracy and communication efficiency.
• Paves the way for more reliable autonomous systems in complex scenarios.