Aptiv Releases Comprehensive Open-Source Dataset for Autonomous Driving

author: Eric Walz   

Global auto parts supplier Aptiv, formally known as Delphi Automotive, announced today the full release of nuScenes, an open-source autonomous vehicle (AV) dataset. The dataset will help developers improve the safety of autonomous vehicles.

Aptiv is the first company to share such a large, comprehensive dataset with the public. Aptiv says it's solving for a gap in the AV industry, which has limited open source data available for research purposes. Open source means its free for developers to use as needed.

Datasets are used to training machine learning models across different AI fields. They are used by engineers and developers of autonomous vehicles to train autonomous driving systems. However, training machine learning models requires vast amounts of "training data", which is the reason Aptiv made available its nuScenes dataset to developers.

For autonomous driving, these datasets might contain videos of street scenes captured from self-driving vehicle's real-world environment, such as a busy urban intersections filled with pedestrians. The training data is used to "train" machine learning algorithms, so software can better detect each person as well as predicting their intended trajectory, allowing a self-driving car to safely navigate.

"At Aptiv, we believe that we make progress as an industry by sharing—especially when it comes to safety," said Karl Iagnemma, president of Aptiv Autonomous Mobility. "Our team thought carefully about the components of our data that we could open to the public in order to enable safer, smarter systems across the entire autonomous vehicle space."

nuScenes is organized into 1,000 unique "scenes," collected from streets in Boston and Singapore, two cities known for dense traffic and challenging driving environments. Aptiv states It contains some of the most complex driving scenarios in each urban environment.

The nuScenes dataset is composed of 1.4 million images, 390,000 lidar sweeps, and 1.4 million 3D human annotated bounding boxes, representing the largest multimodal 3D AV dataset released to date. Aptiv says that nuScenes has 100 times as many images as the pioneering KITTI dataset.

Each scene is 20 seconds long and fully annotated with 3D bounding boxes for different 23 classes (car, bicycle, person, child) and 8 attributes (moving, parked, stopped).

Robust detection and tracking of objects is crucial for the deployment of autonomous vehicle technology and there is a demand for high-quality datasets. Image-based benchmark datasets have driven the development of computer vision tasks such as object detection, tracking and segmentation of agents (cars, people) in the environment.

Datasets are used to train machine learning models to identify pedestrians. (Photo: IEEE Spectrum)

nuScenes Includes the Full Autonomous Vehicle Hardware Suite

Most autonomous vehicles in development are equipped with a full suite cameras and sensors such as lidar and radar to identify and track objects, such as a pedestrian or other vehicles. The goal of nuScenes is to look at the entire vehicle sensor suite.

As machine learning based methods for detection and tracking become common in the automotive industry, there is a growing need to train and evaluate machine learning models on datasets containing AV sensor data, not just camera images.

Most of the previously released datasets focus on camera-based object detection. Two examples are Cityscapes, Mapillary Vistas. nuScenes is the first dataset to include a full autonomous vehicle sensor suite. The sensor suite includes six cameras, five radars and one lidar, providing a full 360 degree field of view around the vehicle.

Aptiv also defines a new metric for 3D detection which consolidates the multiple aspects of the detection task: classification, localization, size, orientation, velocity and attribute estimation.

Datasets are not just used in autonomous driving development, they are used in the field of AI to train machine learning software learn to identify objects. For example, a dataset of human faces might be used to train AI models for facial recognition. Stanford University even has a dataset of dogs, which can be used to train AI programs to identify dog breeds.

By sharing the critical safety data included in nuScenes with the public, it enables Aptiv to support robust progress and innovation in the industry. Aptiv aims to support research in computer vision and autonomous driving to further advance the mobility industry.

To date, over 1,000 users and over 200 academic institutions have registered to access the nuScenes dataset.

Aptiv is also working with ride-hailing company Lyft on its autonomous robo-taxi service.

Eric Walz

Originally hailing from New Jersey, Eric is a automotive & technology reporter covering the high-tech industry here in Silicon Valley. He has over 15 years of automotive experience and a bachelors degree in computer science. These skills, combined with technical writing and news reporting, allows him to fully understand and identify new and innovative technologies in the auto industry and beyond. He has worked at Uber on self-driving cars and as a technical writer, helping people to understand and work with technology.