Abstract
Smart traffic monitoring systems rely on a distributed network of RoadSide Units (RSUs) to collect and transmit visual data for training Deep Learning (DL) models. However, limited wireless bandwidth and varying channel conditions often create a data bottleneck. Traditional scheduling policies focus on maximizing throughput or minimizing delay, often ignoring the semantic value of the transmitted data. This paper presents a diversity-maximized scheduling policy that prioritizes "semantic information flow." By combining coalition game theory for RSU selection with uncertainty-based sampling for data prioritization, the proposed framework ensures that the most informative data reaches the central server. Experimental results demonstrate that this approach achieves a classification accuracy of 87%, outperforming traditional network-centric methods by 5% and showing high resilience to packet errors.
Introduction
In the evolution of smart cities, RoadSide Units (RSUs) serve as the primary sensory organs for traffic management, monitoring everything from vehicle flow to pedestrian safety. These units generate massive volumes of high-resolution visual data that must be offloaded to a central server to update or train Deep Learning (DL) models.
Despite the capabilities of modern wireless networks, they remain a significant bottleneck. Traditional network engineering typically treats all data packets equally, aiming to maximize raw throughput or minimize latency. This "content-blind" approach is problematic for AI applications: if a scheduling policy favors an RSU with a perfect connection that only sees repetitive, redundant traffic, the resulting dataset will be imbalanced. Consequently, the central AI model may fail to recognize rare but critical events, such as traffic accidents or specific signal changes. This paper addresses the need for a scheduling policy that prioritizes data diversity and semantic quality over sheer volume.
Methodology
The researchers propose a two-layered optimization strategy designed to maximize the "semantic information flow" from the edge to the cloud.
1. Coalition-Based Greedy Scheduling
Selecting the optimal subset of RSUs for transmission is a non-convex, NP-hard problem. To manage this complexity, the authors model the RSUs as players in a Coalition Game. The system employs a greedy algorithm to select RSUs based on their marginal contribution to the collective dataset's quality.
Key to this selection is Jain’s fairness index, which ensures that data classes (e.g., different traffic light states) are represented equitably. By approximating Shapley values, the scheduler identifies which RSU provides the best balance of favorable channel conditions and unique data contributions, ensuring that the network doesn't just listen to the "loudest" (best-connected) units, but the most "informative" ones.
2. Uncertainty-Based Sample Selection
Once an RSU is granted a transmission slot, it must decide which specific images to send. Rather than random sampling, the system utilizes an uncertainty-based selection method using a min-margin criterion. The RSU identifies samples where the current global ML model is most uncertain—for instance, an image where the model cannot clearly distinguish between a yellow and red light. By prioritizing these "difficult" samples, the RSU ensures that the central server receives data that directly addresses the model's current weaknesses.
3. Network Constraint Satisfaction
The framework translates these high-level transmission probabilities into a concrete binary schedule. This ensures that the system adheres to physical constraints, such as the maximum number of available wireless channels ($M$), while optimizing the timing of each packet to minimize interference and loss.
Experimental Setup
The proposed policy was evaluated using two distinct datasets to test its robustness:
- Traffic Light Dataset: A real-world collection of 44,000 images from five different RSU cameras. This dataset focuses on the classification of red, yellow, and green lights under varying environmental conditions.
- CIFAR-10 Benchmark: Used to simulate extreme "non-IID" (Independent and Identically Distributed) scenarios. In this setup, each RSU was restricted to only two out of ten possible image classes to test the scheduler's ability to correct for severe data imbalance.
Results and Discussion
The study revealed several critical insights into the relationship between network performance and AI accuracy:
- The Throughput-Diversity Trade-off: Interestingly, the results show that maximizing raw throughput can actually degrade model accuracy. When the network favors RSUs with the best signal strength, it often ignores RSUs with poorer connections that may hold more diverse or rare data. The proposed policy intentionally sacrifices some throughput to include these unique data points, resulting in a more robust final model.
- Superior Accuracy: The diversity-maximized approach achieved a classification accuracy of 87%, a 5% improvement over standard methods like FedCS or uniform random selection. This improvement was most pronounced in scenarios where data was highly imbalanced across different RSUs.
- Resilience to Network Noise: As the Packet Error Rate (PER) increased, the performance of traditional models dropped sharply. In contrast, the fairness-optimized model maintained higher accuracy, proving that when bandwidth is scarce, what you send is more important than how much you send.
Conclusion
This research highlights a shift toward "application-aware" networking. By treating the wireless network not just as a pipe, but as a filtered conduit for semantic information, smart cities can train more accurate AI models with less data. While challenges remain regarding the computational overhead of calculating marginal contributions in real-time, the use of coalition games provides a scalable path forward for distributed learning in constrained environments. Future work will explore how these techniques can be applied to detect "unknown unknowns"—traffic events that the model has never encountered before.
Summary updated