Robust Target Association Method with Weighted Bipartite Graph Optimal Matching in Multi-Sensor Fusion.

Area of Science:

• Sensor Fusion
• Data Association
• Machine Learning

Background:

• Multi-sensor multi-target tracking faces challenges in heterogeneous radar systems due to biases, asynchronous data, and dense formations.
• Existing association methods struggle with measurement distortions and sensor-specific deviations, limiting their robustness.

Purpose of the Study:

• To develop a robust framework for group target association in challenging multi-sensor environments.
• To improve accuracy and reliability in heterogeneous radar systems with systematic errors and non-uniform target densities.

Main Methods:

• Integration of deep feature embedding, density-adaptive clustering (autoencoder-HDBSCAN), and global graph-theoretic matching.
• Utilizing a weighted bipartite graph and optimal matching to handle systematic errors and maintain structural consistency.
• Employing a mutual-support verification mechanism for enhanced robustness against random disturbances.

Main Results:

• Achieved over 90% association accuracy in dense scenarios (1.4 km target spacing).
• Outperformed baseline methods (Deep Association, JPDA) by over 20% under various systematic bias conditions.
• Demonstrated robustness, adaptability, and suitability for practical multi-radar applications.

Conclusions:

• The proposed framework offers a reliable, training-free, and deployable solution for group target association in real-world multi-sensor fusion.
• The method effectively addresses nonlinear distortions, non-uniform densities, and heterogeneous systematic errors.
• Significant improvements in association accuracy highlight its potential for advanced multi-radar tracking systems.