Journal Search Engine
The Journal of The Korea Institute of Intelligent Transport Systems Vol.24 No.6 pp.102-120
멀티모달 인공지능을 이용한 교통사고 유형 예측
Prediction of Traffic Crash Types Using Multimodal Artificial Intelligence
Abstract
The severity of injuries to traffic crash victims varies depending on the type of traffic crash. Therefore, it is necessary to tailor the geometric design of roads and install road features according to the specific type of traffic crash. This study presents a multimodal AI model that can learn structured and image data to classify traffic crash types based on the geometric design of roads. The multimodal model used a custom neural network based on the ReLU function to process structured data, and EfficientNet to process image data. Training results showed that the accuracy of the multimodal model was 60.3% and the F1 Score was 0.604. This is 13.5%p higher than the accuracy of the structured data single-modal model (46.8%) and 12.8%p higher than that of the image data single-modal model (47.5%). It indicates that it is possible to improve the accuracy of the type of traffic crash classification using AI by learning two modalities together. However, three issues are pointed out as limitations of the study. First, Due to the lack of structured data elements input into the model, explanatory power for predicting traffic crash types is insufficient. Second, Despite the theoretical validity of the multimodal model, its performance improvement is not up to par. Third, it is challenging to interpret the functioning of multimodal AI models in the context of transportation studies.
