The Accid3nD dataset is the first high-quality real-world accident dataset for the 3D object detection, segmentation, tracking, trajectory prediction and accident detection task in autonomous driving. It features 112k labeled frames from 5 highway mounted sensors and captures shocking real-world footage of about 10 accidents. See https://accident-dataset.github.io/.
In the LeAP paper we use foundation models to label datasets in domains that do not currently have labeled datasets. We leverage foundation models and devise a novel Bayesian voxel aggregation method to aggregate labels over time, leading to state-of-the-art results. This work was done in collaboration with Perciv AI. See https://arxiv.org/abs/2502.03901.
Neuro NCAP is a novel photorealistic crash test simulator that leverages Neural Radiance Fields (NeRF) to manipulate recorded scenes and test state-of-the-art end-to-end planners whether they can navigate these crash tests safely. Leading planners fail in over 50% of the cases, pointing out a huge shortcoming of these approaches. See https://research.zenseact.com/publications/neuro-ncap/.
The TU Delft SenseBike is a sensor-equipped bicycle used to capture information about the rider of the bicycle, as well as its environment. It is equipped with lidars, cameras, GNSS and IMU. We tackle numerous problems of interest, like object detection, lidar segmentation, motion planning and eye tracking of the rider. This bike is featured prominently in the media, as well as at TU Delft's Dies Natalis (birthday) 2024 with the theme transportation and mobility.
The Intelligent Vehicles group at TU Delft aims to address challenges across the entire autonomous vehicle stack, from perception, mapping, planning and control to human factors. We perform regular real-world demos and create yearly retrospective videos of our activities.
nuScenes is the world's first autonomous vehicle dataset created from a **real** autonomous vehicle tested on public roads. It has a full 360 sensor suite, radar sensors, data from two continents. nuPlan is the world's first dataset for ML-based planning and end-to-end planning. Both datasets were developed and lead by Holger during his time at Motional (formerly nuTonomy). See http://nuplan.org and https://nuscenes.org.