Traffic Accident Prediction via Three-Dimensional Convolution Autoencoder and Victim-Party Demographic Data

EasyChair Preprint 14288

Abstract

This paper proposes a new spatial-temporal framework to predict the number of accidents in an urban area, by integrating multiple heterogeneous data sets such as road structure, intersection density, weather, historical collisions, traffic information and victim and party demographics data. Firstly we collect all data sets for the area of Los Angeles, which is our case study. Secondly, we propose a deep learning framework and efficient pipeline based on a three-dimensional convolutional autoencoder that determines the optimal size and sequential combinations of input for predicting the number of accidents to take place in a subarea at a given point in time. The framework is further improved based on victim and demographic party data, which indicates that the age of 23 is common among victims (mostly young men aged between 18-35 years old). Thirdly, we show that our proposed architecture outperforms other baseline models such as stacked denoising autoencoder, multilayer perceptron and convolutional long short-term memory network in terms of root mean square error and symmetrical mean absolute percentage error.

Keyphrases: Accident Prediction, Artificial Intelligence, deep learning, spatial-temporal modelling

@booklet{EasyChair:14288,
  author    = {Walden Ip and Artur Grigorev and Adriana-Simona Mihaita},
  title     = {Traffic Accident Prediction via Three-Dimensional Convolution Autoencoder and Victim-Party Demographic Data},
  howpublished = {EasyChair Preprint 14288},
  year      = {EasyChair, 2024}}