Towards Accurate Traffic Accident Detection: Developing Deep Learning Strategies with Distant Past, Recent Past, and Adjacency Features

Abstract

Traffic accidents across the world can cause great loss of human life and create financial burden on institutions. However, early detection of accidents and timely delivery of necessary assistance can greatly reduce the likelihood and period of these negative effects occurring. The aim of this study is to minimize the negative effects of traffic accidents on human life, environmental health and macroeconomics on the long run. The paper introduces a novel input set derived from distant and near past to detect traffic accidents with a delay of 5-30 minutes while informing authorities and citizens. Furthermore, the study presents statistical methods and windowing techniques to pre-process and transform the traffic speed data into various features, which are then fed into the most preferred deep learning architectures for multivariate time-series data such as Multi-Layer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) to detect accidents in traffic. In order to verify our strategy, we ran our tests on the traffic accident dataset obtained from the 100 main road segments in Istanbul during 2018. With a 15-minute detection delay, CNN-based accident detection model outperformed all other candidate models, while generating an F-Score of 0.75 and an AUC of 0.87. The results indicates that, although there is still a gap for an improvement, pre-eliminary results are promising for determining anomaly in traffic, i.e traffic accidents, exploiting only traffic speed info.