Traffic volume imputation based on multi-source data collection system

• Traffic detectors collect traffic data to understand the flow of traffic on a city scale, which can help build an efficient traffic management system or deliver useful traffic information to road users for better decision-making.
• However, various types of missing data occur depending on the data management system and unexpected situations on the road. In the short term, we found that the gaps were hourly, but in the long term, they were monthly.
• We propose a deep learning methodology based on a graph neural network that can simultaneously interpolate the missing values of a large amount of traffic detector data installed at the city level.
• By applying an attention mechanism in the graph domain to cooperatively consider the spatial and temporal correlation of two different types of detectors, the model recovers the original traffic data with an error of less than 12 cars/5 minutes, regardless of the missing data rate.

→ The overall multi-source data aggregation process and imputation procedure

A deep spatio-temporal approach in maritime accident prediction
– Predicting the risk of maritime accidents is crucial for improving traffic surveillance and marine safety.

• This study aims at investigating the application of deep learning in both short- and long-term predictions of different types of accident risks associated with small vessels by considering multiple influencing factors.
• The results reveal that although the performance of the proposed deep spatiotemporal ocean accident prediction (DSTOAP) model varies according to grid sizes and time intervals, its accuracy (more than 78%) makes it reliable for predicting accidents.
• Furthermore, although all types of accidents are captured with high accuracy, more than 84% of collision accidents can be predicted accurately.

→ The proposed DSTOAP architecture and prediction results for each grid size

Interactive driving behavior in overtaking situation

• Existing studies on overtaking behavior have only examined it from the perspective of the overtaking vehicle, with limited consideration for the risk level of the vehicle approaching from the opposite side.
• The multiagent simulator helped collect a realistic overtaking driving dataset without compromising safety. The collected data was preprocessed using SMOTE_Tomek-based imbalanced data techniques to build a high-quality dataset and model human decision-making based on machine learning

Human Recognition based on ITS technology

• Changes in drivers’ cognitive responses and driving behavior were observed when warning right-turning vehicles to improve pedestrian safety at intersections.
• By considering human factors such as the driver’s pupil size, viewing position, braking, and acceleration, the points where changes in driving behavior were detected and differences in behavior based on the warning method were analyzed.

Pedestrian behavior

• A pedestrian treadmill simulator is connected to the Metaverse simulator platform to calculate parameters for pedestrian height and stride length, ensuring reliable data collection for pedestrians.

Optimized Signal plan development

• Develop an optimized traffic signal plan using the Neighbor Dueling Deep Q-Network (NDDQN) algorithm, which considers neighboring intersections.
• The algorithm controls four neighboring intersections simultaneously by determining signal actions with the goal of reducing emissions as a reward.