Lateral Control of an Autonomous and Connected Following Vehicle With Limited Preview Information
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
Lateral control of an autonomous and connected vehicle (ACV), especially in emergency situations, is important from the safety viewpoint. In these situations, the trajectory to be followed by an ACV must either be planned in real-time (e.g., for a possible evasion maneuver if the obstacle to be avoided is detected) or be communicated from its preceding vehicle. Typically, the trajectory information is available to the following ACV in the form of GPS time samples. From the viewpoint of lateral control, the lateral velocity information is not readily available, and the feedback structure must reflect this reality. In this work, we develop a methodology to synthesize a lateral control algorithm for a following ACV in a two-vehicle platoon in two steps: (1) From the limited preview information of the trajectory to be tracked via samples of GPS waypoints, and we estimate the radius of curvature of the trajectory using ``least-square'' estimation and (2) develop a fixed-structure feedback control scheme for following the predecessor by synthesizing the set of stabilizing gains corresponding to lateral position error, heading error and heading rate error. Numerical simulation and experimental results corroborate the effectiveness of the proposed feedback-feedforward schemes. Based on this proposed feedback-feedforward controller, we investigated Emergency Lane Change (ELC) control problem for a convoy of autonomous and connected vehicles. Typically, an ELC maneuver is triggered by emergency cues from the front or the end of convoy as a response to either avoiding an obstacle or making way for other vehicles to pass. From a safety viewpoint, connectivity of ACVs is essential as it entails obtaining or exchanging information about other ACVs in the convoy. This thesis assumes that ACVs have reliable connectivity and that every following ACV has the information about GPS position traces of the lead and immediately preceding vehicles in the convoy. This information provides a ``discretized'' preview of the trajectory to be tracked. Based on the available information, this thesis focuses on two schemes for synthesizing lateral control of ACVs based on (a) a single composite ELC trajectory that fuses lead and preceding vehicle's GPS traces and (b) separate ELC trajectories based on preview data of preceding and lead vehicles. The former case entails the construction of a single composite ELC trajectory, determination of the cross track error, heading and yaw rate errors with respect to this trajectory and synthesis of a lateral control action. The latter case entails the construction of two separate trajectories corresponding to the lead vehicle's and preceding vehicle's data separately and the subsequent computation of two sets of associated errors and lateral control actions and combining them to provide a steering command. Numerical and experimental results corroborate the effectiveness of these two schemes. For multiple vehicle control in a convoy/platooning, identifying vehicles and objects in the vicinity of platooning vehicles is critical for convoy/platooning safety. A neighboring vehicle may change lanes and enter a lane with platooning vehicles. Any deployable platooning system must be able to first identify the presence of a cut-in vehicle before performing any control actions on the vehicles. In this work, we present a sensor fusion algorithm that combines radar and vision data obtained onboard a moving truck to identify the states of the cut-in vehicle. We then present experimental results to corroborate the performance of the proposed algorithms. First, I investigate temporal patterns of user trust and reliance in XAI systems (Objective 3). My study results show that model explanations not only affected user final trust but also shape how user trust evolves over time; indicating the importance of user behavior for evaluating XAI systems. Lastly, I propose an open-sourced human-attention evaluation baseline for direct evaluation of saliency map explanations (Objective 4). I demonstrate my human-attention benchmark's utility for quantitative evaluation of model explanations by comparing it with single-layer feature masks baseline. My experiments also show the advantage of my evaluation baseline by revealing different user biases in the subjective rating evaluation of model saliency explanations.