Date of Award
Doctor of Philosophy (PhD)
Matthias J Schmid
Venkat N Krovi
This research aims at developing new methods that predict the behaviors of the human driven traffic participants to enable safe operation of autonomous vehicles in complex traffic environments. Autonomous vehicles are expected to operate amongst human driven conventional vehicles in the traffic at least for the next few decades. For safe navigation they will need to infer the intents as well as the behaviors of the human traffic participants using extrinsically observable information, so that their trajectories can be predicted for a time horizon long enough to do a predictive risk analysis and gracefully avert any risky situation. This research approaches this challenge by recognizing that any maneuver performed by a human driver can be divided into four stages that depend on the surrounding context: intent determination, maneuver preparation, gap acceptance and maneuver execution. It builds on the hypothesis that for a given driver, the behavior not only spans across these four maneuver stages, but across multiple maneuvers. As a result, identifying the driver behavior in any of these stages can help characterize the nature of all the subsequent maneuvers that the driver is likely to perform, thus resulting in a more accurate prediction for a longer time horizon. To enable this, a novel probabilistic framework is proposed that couples the different maneuver stages of the observed traffic participant together and associates them to a driving style. To realize this framework two candidate Multiple Model Adaptive Estimation approaches were compared: Autonomous Multiple Model (AMM) and Interacting Multiple Model(IMM) filtering approach. The IMM approach proved superior to the AMM approach and was eventually validated using a trajectory extracted from a real world dataset for efficacy. The proposed framework was then implemented by extending the validated IMM approach with contextual information of the observed traffic participant. The classification of the driving style of the traffic participant (behavior characterization) was then demonstrated for two use case scenarios. The proposed contextual IMM (CIMM) framework also showed improvements in the performance of the behavior classification of the traffic participants compared to the IMM for the identified use case scenarios. This outcome warrants further exploration of this framework for different traffic scenarios. Further, it contributes towards the ongoing endeavors for safe deployment of autonomous vehicles on public roads.
Gill, Jasprit Singh, "Probabilistic Framework for Behavior Characterization of Traffic Participants Enabling Long Term Prediction" (2019). All Dissertations. 2509.