Safe intention-aware maneuvering of autonomous vehicles
Author(s)Huang, Xin(Xin Cyrus)
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics.
Brian C. Williams.
MetadataShow full item record
In order to improve driving performance, while achieving safety in a dynamic environment, it is crucial for a vehicle motion planner to be aware of the intentions of the surrounding agent vehicles. Many existing approaches that ignore the intentions of surrounding vehicles would produce risky or over-conservative plans. In this thesis, we describe a maneuver motion planning system that achieves both safety and efficiency by estimating the types of surrounding drivers and the vehicle motions being executed by them over a finite horizon in the future. Our claim is that a vehicle is able to efficiently and safely navigate in dynamic traffic situations by estimating the possible types of drivers in its immediate vicinity, such as aggressive or careful, and by predicting their likely maneuvers and motions as probability distributions.To perform these predictions, we first employ a vehicle model that incorporates different driving styles, possible types of maneuvers, the likely trajectories that these maneuvers produce, and the likelihood of transitioning between successive maneuvers. The vehicle models are combined and encoded as hybrid partially observable Markov decision processes (POMDPs) whose discrete elements represent driving styles and maneuvers for each style and whose continuous parts represent vehicle motions. We then frame the problem of recognizing a vehicle's current driving style and maneuver as a belief state update on the hybrid POMDP. The driving style is assessed using multinomial logistic regression classification, while the maneuver is estimated using Bayesian filtering over a variant of probabilistic hybrid automata and a library of pre-learned motion primitive models.Multinomial logistic regression classification allows us to predict driving styles probabilistically using multiple driving features, and Bayesian filtering provides robust estimation results based on the prior information. Given the recognition results and the learned motion primitive models, we provide probabilistically sound predictions of the future maneuver and trajectory sequence of each agent vehicle. Finally, we compute safe motion plans of the ego vehicle in light of recognized agent vehicle driver styles, intended maneuvers, and future vehicle trajectories, by performing risk-bounded planning on the hybrid POMDP model. We demonstrate our system in a number of challenging simulated environments, including unprotected intersection left turns and lane changes with multiple dynamic vehicles.The demonstration shows that our intent recognition algorithms achieve an average driving style estimation accuracy of 89.89%, an average maneuver estimation accuracy of 98.9%, and an average trajectory prediction error of 2.12 meters. Furthermore, our maneuver planning system guarantees safety with respect to the safety constraint, while arriving at the goal 13.71% faster compared to a state-of-the-art planner without the intent recognition capability.
Thesis: S.M., Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, 2019Cataloged from PDF version of thesis.Includes bibliographical references (pages 119-123).
DepartmentMassachusetts Institute of Technology. Department of Aeronautics and Astronautics
Massachusetts Institute of Technology
Aeronautics and Astronautics.