Relation3DMOT: Exploiting Deep Affinity for 3D Multi-Object Tracking from View Aggregation.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3D multi-object tracking (MOT) is crucial for safe autonomous navigation.
• Current MOT methods often rely on 2D object detection, limiting 3D localization accuracy.
• Challenges exist in learning discriminative features and object interactions for consistent tracking.

Purpose of the Study:

• To enhance the reliability and accuracy of 3D multi-object tracking for autonomous systems.
• To address limitations in current tracking-by-detection pipelines.
• To improve feature learning and data association in MOT.

Main Methods:

• A joint feature extractor fuses appearance and motion features from 2D RGB images and 3D point clouds.
• A novel convolutional operation, RelationConv, is proposed to model inter-object correlations.
• A deep affinity matrix is learned for improved data association.

Main Results:

• The proposed model achieves state-of-the-art performance on the KITTI tracking benchmark.
• Fusion of 2D and 3D data enhances object localization and tracking reliability.
• RelationConv effectively captures feature interactions between objects for better association.

Conclusions:

• The developed approach significantly advances 3D multi-object tracking capabilities.
• The method provides a more robust solution for perception in autonomous systems.
• Future work can explore further improvements in feature fusion and relation modeling for MOT.