Long short-term memory networks-based Framework for Traffic Crash Detection with Traffic Data

Traffic crash detection explores the interconnected relationship between traffic data and crash risk. It can prevent potential traffic accidents and improve freeway safety. However, some limitations exist in current studies: First, datasets in most studies were simulated or very old, which cannot represent practical traffic patterns; second, few studies applied deep learning methods for crash detection, which had significant performance in other traffic domains. To address the above limitations, this study proposes a deep learning method of long short-term memory (LSTM) networks for crash detection with traffic data on freeways. A dropout technique is adopted to reduce overfitting and improve prediction performance in small datasets. The framework is implemented on datasets extracted from I880-N and I805-N in California, America. 6 LSTM models are constructed for real-time crash detection and validated to have satisfactory prediction performance, with the highest crash accuracy of 68.75% and highest comprehensive indicator of 72.33%. The results also indicate that dropout technique can improve prediction performance, increasing crash accuracy from 59.38% to 68.75%. The detection models established on one freeway can be transferred to another similar freeway, with the highest crash accuracy of 61.29%. Based on the comparisons of LSTM model and other machine learning methods, LSTM model has validated to have better prediction performance than SVM and ANN, especially on the criteria of crash accuracy and comprehensive indicator.