Predictive Risk Analysis using Deep Learning in Indian Traffic

Abstract

The Indian traffic is chaotic comprising of complex interactions between heterogeneous entities and a lack of display of traffic rules and etiquette. The traffic hence poses a risk that has not been appreciably analyzed in the literature. We propose a novel weighted-elliptical-model based approach to asses the risks in the Indian traffic. Then using the capabilities of deep learning, we also model the behavior of road agents for trajectory prediction. Finally, combining trajectory prediction with our risk-analysis model, we present Predictive Risk Analysis, capable of forecasting risky scenarios about 3 seconds before they happen, accurately. Using our own collected and labelled data, we evaluate the performance of our approach and show how it beats other baseline approaches for the same.