Date of Award
Doctor of Philosophy (PhD)
The development of autonomous vehicle technology has moved to the center of automotive research in recent decades. In the foreseeable future, road vehicles at all levels of automation and connectivity will be required to operate safely in a hybrid traffic where human operated vehicles (HOVs) and fully and semi-autonomous vehicles (AVs) coexist. Having an accurate and reliable perception of the road is an important requirement for achieving this objective. This dissertation addresses some of the associated challenges via developing a human-like social driver model and devising a decentralized cooperative perception framework.
A human-like driver model can aid the development of AVs by building an understanding of interactions among human drivers and AVs in a hybrid traffic, therefore facilitating an efficient and safe integration. The presented social driver model categorizes and defines the driver's psychological decision factors in mathematical representations (target force, object force, and lane force). A model predictive control (MPC) is then employed for the motion planning by evaluating the prevailing social forces and considering the kinematics of the controlled vehicle as well as other operating constraints to ensure a safe maneuver in a way that mimics the predictive nature of the human driver's decision making process. A hierarchical model predictive control structure is also proposed, where an additional upper level controller aggregates the social forces over a longer prediction horizon upon the availability of an extended perception of the upcoming traffic via vehicular networking. Based on the prediction of the upper level controller, a sequence of reference lanes is passed to a lower level controller to track while avoiding local obstacles. This hierarchical scheme helps reduce unnecessary lane changes resulting in smoother maneuvers.
The dynamic vehicular communication environment requires a robust framework that must consistently evaluate and exploit the set of communicated information for the purpose of improving the perception of a participating vehicle beyond the limitations. This dissertation presents a decentralized cooperative perception framework that considers uncertainties in traffic measurements and allows scalability (for various settings of traffic density, participation rate, etc.). The framework utilizes a Bhattacharyya distance filter (BDF) for data association and a fast covariance intersection fusion scheme (FCI) for the data fusion processes. The conservatism of the covariance intersection fusion scheme is investigated in comparison to the traditional Kalman filter (KF), and two different fusion architectures: sensor-to-sensor and sensor-to-system track fusion are evaluated.
The performance of the overall proposed framework is demonstrated via Monte Carlo simulations with a set of empirical communications models and traffic microsimulations where each connected vehicle asynchronously broadcasts its local perception consisting of estimates of the motion states of self and neighboring vehicles along with the corresponding uncertainty measures of the estimates. The evaluated framework includes a vehicle-to-vehicle (V2V) communication model that considers intermittent communications as well as a model that takes into account dynamic changes in an individual vehicle’s sensors’ FoV in accordance with the prevailing traffic conditions. The results show the presence of optimality in participation rate, where increasing participation rate beyond a certain level adversely affects the delay in packet delivery and the computational complexity in data association and fusion processes increase without a significant improvement in the achieved accuracy via the cooperative perception.
In a highly dense traffic environment, the vehicular network can often be congested leading to limited bandwidth availability at high participation rates of the connected vehicles in the cooperative perception scheme. To alleviate the bandwidth utilization issues, an information-value discriminating networking scheme is proposed, where each sender broadcasts selectively chosen perception data based on the novelty-value of information. The potential benefits of these approaches include, but are not limited to, the reduction of bandwidth bottle-necking and the minimization of the computational cost of data association and fusion post processing of the shared perception data at receiving nodes. It is argued that the proposed information-value discriminating communication scheme can alleviate these adverse effects without sacrificing the fidelity of the perception.
Yoon, Daniel, "Cooperative Perception for Social Driving in Connected Vehicle Traffic" (2021). All Dissertations. 2905.