Introduction
Autonomous driving systems rely heavily on Street Scene Semantic Understanding (TriSU) to navigate complex urban environments safely. A persistent challenge in this field is "domain shift," where models trained in specific geographic or weather conditions (e.g., sunny California) fail to generalize when deployed in different environments (e.g., rainy London).
Hierarchical Federated Learning (HFL) has emerged as a privacy-preserving solution, allowing vehicles to train a global model collaboratively via intermediate edge servers. However, HFL is often hampered by two major hurdles: data heterogeneity across different regions and significant communication overhead caused by transmitting high-dimensional model weights.
Proposed Methodology: FedGau and AdapRS
This research introduces a two-pronged approach to optimize HFL for autonomous driving: FedGau for handling data heterogeneity and AdapRS for alleviating communication bottlenecks.
FedGau: Gaussian-Weighted Heterogeneous HFL
Standard HFL typically weights model updates based solely on the volume of local data. In contrast, FedGau utilizes statistical similarity. By modeling the pixel intensity of local datasets as Gaussian distributions, the system can calculate the Bhattacharyya distance between a vehicle's data and the aggregate distribution at the edge or cloud level. Updates from vehicles that are more statistically representative of the target domain are assigned higher weights, which ensures the global model prioritizes representative features and accelerates convergence.
AdapRS: Performance-Aware Adaptive Resource Scheduling
To address the high cost of transmitting model weights, AdapRS introduces dynamic communication scheduling. It monitors the "Quality of Communication" (QoC), which measures the ratio of performance gain (e.g., increase in mIoU) to communication cost. By scheduling more frequent updates during early training phases when the model learns quickly and reducing frequency as the model stabilizes, AdapRS minimizes bandwidth waste without sacrificing final accuracy.
Empirical Analysis and Results
A core insight of this work is that pixel intensities in street scenes naturally follow a "bell curve" or Gaussian distribution. This observation allows for a computationally efficient proxy for complex image data, enabling statistical weighting without the need for raw data sharing.
Experimental results on the Cityscapes and CamVid datasets demonstrate substantial improvements over state-of-the-art methods:
- Convergence Speed: FedGau achieved convergence 35.5% to 40.6% faster than standard FedAvg across metrics like mIoU, Precision, and Recall.
- Communication Efficiency: AdapRS reduced communication overhead by approximately 29.65% compared to static scheduling policies.
- Model Accuracy: On the Cityscapes benchmark using a DeepLabv3+ architecture, the framework achieved a 55.44% mIoU, significantly outperforming existing federated learning algorithms.
Conclusion and Future Directions
The integration of FedGau and AdapRS provides a scalable and efficient framework for hierarchical federated learning in autonomous driving. By prioritizing statistically relevant data and optimizing communication resources, the system achieves faster convergence and lower overhead. Future work will focus on modeling RGB channels independently and extending the framework to handle multi-modal data sources such as LiDAR and Radar.
Summary updated