Unsupervised Domain Adaptive Vehicle Re-Identification: A Federated Learning Scheme

While traditional domain adaptation methods have yielded satisfactory outcomes in vehicle re-identification on established datasets, they typically operate within a centralized framework. This approach falls short in addressing the challenges of large-scale cross-domain adaptation encountered in real-world traffic networks, and often overlooks the computational constraints of roadside units. In response, we introduce a federated learning-based distributed framework for unsupervised domain adaptive vehicle re-identification, structured in three stages: local model processing, dynamic model aggregation, and aggregated model downloading. Initially, diverse local re-identification models are trained independently within edge computing units. For unlabeled target data, we employ a novel clustering technique that averages features from multiple local models to generate pseudo labels. The framework’s second stage incorporates a historical accumulation strategy to retain and utilize the knowledge from local models, ensuring effective adaptation. Additionally, we propose a weight regularization approach for dynamic model aggregation, significantly boosting the framework’s discriminative capabilities. Our methodology, tested on the VeRi-776 and VehicleID datasets, showcases marked improvements in the discriminative performance of domain adaptive vehicle re-identification models, affirming the efficacy of our approach.