Improved Vehicular Congestion Classification using Machine Learning for VANETs

Vehicular Ad-hoc Networks (VANETs) emerge as an inevitable element for autonomous driving, smart cities and intelligent transportation systems. The vehicular traffic density classification plays a crucial role in making important traffic routing and data transfer decisions between vehicles and surrounding infrastructure. However, vehicular density in a given area vastly varies depending on the environment (urban, rural, highway etc.), the day and the specific time of the day. There can also be unpredictable density variations due to traffic incidents or social events. Therefore, accurate classification of traffic density is essential to properly plan data communication in VANET. This paper studies a number of machine learning (ML) algorithms to accurately classify the traffic condition based on the data collected from intelligent sensors. First, the traffic flow and average speed data is collected for each vehicle. In the second step, vehicular density is estimated using speed and flow relationship in a given area. In the third step, traffic state is classified as "Free-Flow", "Dense", and "Congested" based on the congestion cost report by Victoria Transport Policy Institute. Finally, we utilized a range of ML approaches, including Decision Tree (DT), Naive Bayes (NB), K-Nearest Neighbor (KNN), Random Forest (RF), Support Vector Classifier (SVC), Logistic Regression (LR), and Multilayer Perceptron (MLP) to categorize instances of traffic congestion. The results are studied based on classification accuracy, recall and precision metrics. The experimental results indicate that RF and subsequently Ensemble Soft Voting classifiers exhibit the best performance among all other classifiers, including the MLP model.