Machine Learning Techniques for Real-time Traffic Flow Prediction in Autonomous Driving

Abstract

The key insight introduced shows that traffic flow prediction is to predict the numerical values of traffic flow, such as speed, volume, and occupancy in the next several minutes in target road segments. Traffic flow prediction is not just a simple extension of the traditional time-series prediction problem but possesses some distinct features. Firstly, traffic flow data is of high-dimensional, spatiotemporal, and nonlinear. It is a big challenge to accurately model the correlation of traffic flow data in the past for the close to immediate future traffic condition forecasting. Secondly, traffic flow prediction exhibits a strong dependency on environmental conditions. Thus, it is important to quantify the effect of traffic incidents (such as accidents, special events, and weather events) on the traffic flow forecasting. Finally, actions taken by relevant agents can both directly and indirectly influence the traffic outcomes. Therefore, it is important to model the impact of relevant agents’ actions in the short-term traffic flow forecasting.