Machine Learning for VRUs accidents prediction using V2X data

Abstract

Intelligent Transportation Systems (ITS) are systems that consist on an complex set of technologies that are applied to road agents, aiming to provide a more efficient and safe usage of the roads. The aspect of safety is particularly important for Vulnerable Road Users (VRUs), which are entities for whose implementation of automatic safety solutions is challenging for their agility and hard to anticipate behavior. However, the usage of ML techniques on Vehicle to Anything (V2X) data has the potential to implement such systems. This paper proposes a VRUs (motorcycles) accident prediction system by using Long Short-Term Memorys (LSTMs) on top of communication data that is generated using the VEINS simulation framework (pairing SUMO and ns-3). Results show that the proposed system is able to predict 96% of the accidents on Scenario A (with a 4.53s Average Prediction Time and a 41% Correct Decision Percentage (CDP) - 78 False Positives (FP)) and 95% on Scenario B (with a 4.44s Average Prediction Time and a 43% CDP - 68 FP).