Data Processing for Perception of Autonomous Vehicles In Urban Traffic Environments
This event is free and open to the publicAdd to Google Calendar
Despite recent developments in autonomous driving technologies, the perception of Autonomous Vehicles (AVs) in urban environments remains challenging. A key to advance the field is to have quality data with labels covering diverse natural traffic scenarios. However, in practice, preparing such data is not trivial due to the complexity caused by unpredictable behaviors of agents and costly and laborious hardware setup and labeling process.
This dissertation addresses the challenges of AV perception in urban traffic environments, from data collection and labeling to 3D reconstruction and analysis on intrinsic properties. Focusing on unsignalized urban intersections, we first build a data capture system with a multi-modal sensor suite to simulate the actual AV perception. We then introduce the algorithm to fit parametrized human mesh models to the pedestrians in a scene. The generated models provide free labels such as human pose and trajectories with no cost of manual labeling. Expanding the scope of automatic labeling to the scenes, we propose to perform the entire scene modeling by densely reconstructing the scene. We build a simulator that can generate a rich set of labels using virtual sensors. Finally, we tackle the problem of estimating intrinsic properties and discuss ways to achieve realistic reconstruction.
Chair: Professor Matthew Johnson-Roberson