Digital twin-assisted graph matching multi-task object detection method in complex traffic scenarios.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Traffic data collection and annotation are time-consuming and labor-intensive.
• Current deep learning models face limitations in practical traffic applications.
• Distributional discrepancies exist between virtual and real-world datasets.

Purpose of the Study:

• To propose a cross-domain object detection transfer method using digital twins for traffic scenarios.
• To address the challenges of data collection and model limitations in practical traffic analysis.
• To enhance the robustness and adaptability of object detection in diverse traffic environments.

Main Methods:

• Constructed a digital twin traffic scenario using a simulation platform to generate a virtual traffic dataset.
• Developed a multi-task object detection algorithm utilizing graph matching to align feature distributions between source and target domains.
• Incorporated an attention mechanism for instance segmentation and a multi-level discriminator for adversarial training to enhance representation learning.

Main Results:

• Demonstrated the practical value of the generated virtual dataset through comprehensive comparative experiments.
• Validated the effectiveness of the proposed graph-matching-based transfer method.
• Showcased the dataset's capacity to enhance task performance and the approach's robustness and adaptability.

Conclusions:

• The digital twin approach offers a viable solution for generating large-scale, annotated traffic datasets.
• The proposed cross-domain transfer method significantly improves object detection performance in traffic scenarios.
• The method exhibits robustness and adaptability across diverse traffic conditions, highlighting its practical applicability.