Toll plazas with both Electronic Toll Collection (ETC) lane(s) and Manual Toll Collection (MTC) lane(s) could increase crash risks especially at upstream diverging areas because of frequency lane-change behaviors. This study develops the logistic regression (LR) model and five typical non-parametric models including, K-Nearest Neighbor (KNN), Artificial Neural Networks (ANN), Support Vector Machines (SVM), Decision Trees (DT), and Random Forest (RF) to examine the relationship between influencing factors and vehicle collision risk. Based on the vehicle trajectory data extracted from unmanned aerial vehicle (UAV) videos using an automated video analysis system, the unconstrained vehicle motion’s collision risk can be evaluated by the extended time to collision (ETTC). Results of model performance comparison indicate that not all non-parametric models have a better prediction performance than the LR model. Specifically, the KNN, SVM, DT and RF models have better model performance than LR model in model training, while the ANN model has the worst model performance. In model prediction, the accuracy of LR model is higher than that of other five non-parametric models under various ETTC thresholds conditions. The LR model implies a pretty good performance and its results also indicate that vehicle yields the higher collision risk when it drives on the left side of toll plaza diverging area and more dangerous situations could be found for an ETC vehicle. Moreover, the vehicle collision risks are positively associated with the speed of the following vehicle and the angle between the leading vehicle speed vector and X axis. Furthermore, the results of DT model show that three factors play important roles in classifying vehicle collision risk and the effects of them on collision risk are consistent with the results of LR model. These findings provide valuable information for accurate assessment of collision risk, which is a key step toward improving safety performance of the toll plaza diverging area.