Explainable AI for Real-Time Threat Analysis in Autonomous Vehicle Networks

Abstract

The paper will present a ML-based approach that aims to improve the robustness and reliability of AVs through on-device real-time traffic threat assessment. In particular, we will investigate how real-time traffic analysis and prediction can estimate near-future traffic situations in an unstructured and non-instrumented urban environment. We will delve into how a feedforward neural network is deemed as efficient to achieve the aforementioned objective. As the essential novelty, our approach is mostly based on real-world data, which might mitigate overfitting and data-scraping profoundly, if training data is not accurately selected, processed and probabilistically analyzed. We summarize the contributions as follows: (a) implementing a methodology to track vehicle trajectory data on urban networks; (b) real-time traffic origin-destination identification as well as time delay traffic map generation which represent the baseline ground-truth problems in our study; (c) local and fast intersection-based traffic light prediction; (d) global real-time parking lot prediction; (e) local and real-time traffic speed prediction; and (f) local and fast traffic density prediction for high and low traffic intensity areas with a prioritization for high-delay scenarios, unlike prior works.