Trajectory and behavior prediction for autonomous vehicles in urban traffic

Abstract

Autonomous vehicles should transform the urban transport scenario, increasing its efficiency, making it more accessible and safer, reducing the environmental impact, among other benefits. To do so, they are composed of a set of hardware and algorithms, which allow them to understand the external environment, their own state and interact with other traffic agents. To navigate safely, vehicles have algorithms for detecting, classifying and deflecting obstacles. However, because of the dynamism of urban traffic, just obstacle detection is not sufficient to ensure the safety of passengers and other traffic agents. Thus, the tracking, behavior prediction, and trajectory prediction of these agents are essential sources of information that enable decision-making and planning algorithms to consider likely scenarios regarding the position of these other agents, anticipating possible collisions or dangerous situations. Thus, the objective of this project is to investigate and develop techniques for predicting behavior and trajectories, which considers in addition to physical models of vehicle movement, road geometries, traffic rules and models of interaction between vehicles. Therefore, it is proposed to use dynamic bayesian networks that are able to model conditional relationships between variables and temporal dependence. In addition to Markov hidden models, to model the stage of execution of a maneuver or behavior, and recurrent neural networks, known as LSTM (Long Short-Term Memory), to predict the final trajectories. To validate the proposal, a public database and CARLA simulator will be used. Once validated, a real experiment will be performed using the autonomous vehicle CaRINA II, under development by the Mobile Robotics Laboratory (LRM) and the Intelligent Systems Laboratory (LASI) of the University of São Paulo - São Carlos. (AU)