Novel Crash Prevention Framework for C-V2X using Deep Learning

Crash Risk (CR) prediction is essential for Intelligent Transport Systems(ITS), particularly for vehicular users' safety. The rapid development in multivariate deep learning techniques and the emergence of Vehicle to Everything (V2X) communication make it possible to predict CR in smart cities more quickly and precisely. Currently, CRs are predicted using Time-To-Collide, which depends on various interaction data of two conflicting entities. We inspect several factors affecting the CR, like speed, acceleration, Deceleration Rate to Avoid Crashes (DRAC), and Post Encroachment Time (PET). We develop a multivariate LSTM and RNN-ATT model to predict crashes that may occur within the next three seconds based on the past seven seconds of vehicle data. It is simulated on high-density roads of the Ahmedabad city map generated using the Open Street Map. The proposed framework coupling SUMO as traffic simulator and NS-3 as network simulator results in an optimal prediction horizon of 3s with a Root Mean Squared Error of 0.0611. The finding of this paper indicates the promising performance of the proposed framework and LSTM model with an accuracy of 88.20% to deploy in the Indian ITS for real-time crash prevention.