Pothole Detection using Federated Learning to Enhance Road Safety for Intelligent Transportation Systems

Abstract

One of the biggest challenges in ITS is pothole detection. While traditional image processing has struggled to find potholes in AVs, pothole detection systems have made such an approach attractive. As attractive as such avenues are, most AI re- lies on a centralized approach for their implementation. Unfortunately, this under- mines the approach’s potential through reduced storage availability, single-point- of-failure problems, and privacy concerns. In this paper, we experiment with decentralized AI through Federated Learning. Our FL system uses five local clients and the central server. The five local clients train a global model using their datasets. They send the trained model weights to the central server, aggregating them to update the FL model. Then, the updated weights are sent back to the local clients; thus, the first iteration of Federated Learning is concluded. Such a process, there- fore, improves privacy and reduces the vulnerability of global databases. Given our research, data about potholes were retrieved from the Kaggle tow-along pot- hole dataset. They used the FedAvg aggregation function, the respective way to reach a consensus due to its proven reliability. Ultimately, the global model had up to 84% accuracy and an F1 score of 86%; such results significantly improved potholes identified for the AV software. Additionally, our model allows keeping the data decentralized, which makes the process more private and scalable. We plan to experiment with more complex aggregation functions for future work to make the model more precise and consistent. In addition, we might consider implementing blockchain technologies to the interaction between the local clients and the global model to make the process more secure and private. To summarize, implementing Federated Learning is a major step forward in pothole detection for autonomous vehicles, which has overcome the major constraints of traditional centralized AI methodologies. The novel integration approach increased detection accuracy and guaranteed full data privacy and system reliability, creating smarter and more efficient Intelligent Transportation Systems