Fusing transportation rules and diverse motion behaviors for trajectory prediction in traffic intersections

Predicting future trajectories of traffic participants (also called agents) plays an important role in realizing autonomous driving. However, accurately forecasting future trajectories at traffic intersections poses significant challenges. Drivers pay varying attention levels to vehicles in different lanes, and their driving behaviors are influenced by traffic signals. Existing models struggle to effectively establish interaction relationships between vehicles in different lanes to reflect drivers' attention, and often overlook the impacts of traffic signals. Moreover, different types of agents display diverse motion patterns at traffic intersections, rendering a single trajectory prediction method difficult to model such diversity. This paper proposes a traffic intersection trajectory prediction model to address these challenges. Firstly, a directed graph is designed to characterize the high attention levels to vehicles in relevant lanes, while minimizing the impacts of other vehicles. Secondly, an action prediction method is proposed to estimate the future motion actions of vehicles based on traffic signal information. Finally, multiple t-distributions are used to separately forecast the future paths of vehicles with different motion actions and pedestrians. The experiments demonstrate the superiority of our algorithm over twenty-four trajectory prediction models on two datasets with traffic intersection scenes.