Modeling of Severity in Red-Light-Running Crashes Using Deep Learning Recognition

One of the main factors that contribute to the high number of crashes at signalized intersections is noncompliance with traffic signals or in other words, cutting the red-light signal. Analyzing red-light-running crashes (RLR) and their related factors is of great importance in developing evidence-based platforms for establishing effective traffic regulations and road safety programs. The purpose of this study is to understand the factors affecting the severity of RLR related crashes using a sample from the State of Florida in the United States. All crashes used in the analysis were verified to ensure that the main contributing cause was the red-light violation. The crash severity classification was conducted using deep learning neural networks. The impact of improving lighting conditions, road surface, signal visibility are examples of some factors that the deep learning recognition model can be used to evaluate their improvement in reducing crash severity. The model shows a good recognition ability, where the success rate is higher than 77%. It should be noted that Florida has a tropical climate where annual temperature, sunshine hours, and rainfall do not extremely vary throughout the year. Conducting similar studies using data from other states would be beneficial to validate the results obtained.