Past Event: Oden Institute Seminar

Data, detection, and prediction for self-driving cars ** Note: Room and Time Change**

Eric Wolff, Principal Research Scientist, nuTonomy (Aptiv)

Abstract

Both 3D object detection and motion prediction are critical components for safe and efficient self-driving cars. This talk details recently published work at nuTonomy (Aptiv), that advances the state-of-the-art in both areas. We show how to create efficient encodings of LIDAR point clouds for fast and accurate 3D object detection (PointPillars). Additionally, we describe a method for sequential fusion of vision and LIDAR, which improves performance of many LIDAR-only object detection methods (PointPainting). Finally, we describe a method for motion prediction of other agents that results in good coverage of corner cases by leveraging dynamic constraints (CoverNet). All of our methods are evaluated on nuScenes, a large multimodal (camera, LIDAR, radar) dataset for autonomous driving, that nuTonomy (Aptiv) released last year. nuScenes contains over 1k driving scenes in Boston and Singapore, and is free for research. Joint work with many of my colleagues at nuTonomy (Aptiv). Bio Eric Wolff is a Principal Research Scientist at nuTonomy (an Aptiv company), where he leads the Prediction and Behavior Modeling team. He completed his PhD in Control and Dynamic Systems at Caltech, and his BS in Mechanical Engineering at Cornell. He is a recipient of both the NSF Graduate Student Research Fellowship and the NDSEG Fellowship.