Novel algorithm for multivariate time series crash risk prediction using CNN-ATT-LSTM model

Abstract

Multivariate Time Series Crash Risk Prediction is essential in the development of Collision Avoidance Systems (CASs), which are vital components of the Intelligent Transportation System. The crash risk prediction performance is degraded with high computational cost and low accuracy. To address this issue, Attention based CNN-LSTM Hybrid model is proposed for Multivariate time series crash risk prediction through the augmentation of Convolutional Neural Network (CNN) with Attention based Long Short Term Memory (ATT-LSTM). Attention mechanism is incorporated with LSTM to learn long-term dependencies of ultra-long sequences. Modified Crash Risk Index (MCRI) is developed to label the crash and non-crash events considering Adaptive Perception Reaction Time (APRT) which enables the enhancement of the accuracy of the Multivariate time series crash risk prediction system. The problem is formulated as multivariate time series prediction and validity of proposed model is evaluated with Next Generation Simulation (NGSIM) dataset. The proposed model outperforms state-of-the-art models where MCRI and CNN-ATT-LSTM enhance accuracy of the crash risk prediction. 98.1% of accuracy has been achieved in the proposed model. The result demonstrates that the proposed model requires less computational cost, high accuracy and minimum preprocessing. The proposed model presents warning to the driver at the time of collision accurately and can be implemented in Collision Avoidance Systems.