Selecting and Comparing Temporal-Spatial Features for Traffic Incident Detection

Abstract

To better understand the results from traffic incident detection models, this study tried to have an in-depth comparison of the temporal-spatial variables used for traffic incident detection. Specifically, the traffic incidents occurred on the I-80 expressway in 2019 were collected, and the typical incidents were selected according to the incident occurrence time, duration, and traffic state. A total number of 144 feature variables was constructed with 108 temporal and 36 spatial ones. The Spearman correlation analysis was first used to explore the correlation between the collected traffic incidents and the feature variables. Moreover, the random forest (RF) model as the best-performing one among four different machine learning and statistical models was selected as the basic model for incident detection, and the effectiveness of the selected temporal-spatial feature variables was further compared. The results showed that the incident detection model built in this study performed well and the incident detection rate reached 90% by applying RF. The correlation analysis and the RF feature importance ranking both indicated that the temporal features have a more significant impact on incident detection than the spatial features.