Skip to main content
Bosch Engineering

Fast, efficient, precise: automated 3D labeling of sensor data for automated driving

Im Zeitalter des automatisierten Fahrens ist die präzise Erkennung und Interpretation von Objekten und Verkehrsituationen essentiell. Dafür setzen Fahrzeughersteller eine Vielfalt von Sensoren ein wie Radar, Lidar und Kameras, aus deren Daten wichtige Umgebungsinformationen generiert werden, oft mit KI-Algorithmen. Zum Training dieser KI-Algorithmen müssen große Mengen an Sensordaten annotiert werden. Bei der Datenannotation werden relevante Sensorinformationen besonders markiert, damit die KI der automatisierten Fahrfunktion versteht, was sie verarbeiten soll. Die Annotation erfolgt über Labels, die oft noch manuell in den Datensatz eingefügt werden. Diese Tätigkeiten sind mit erheblichem Kosten- und Zeitaufwand verbunden und werden zu einer immer größeren Herausforderung.

In the age of automated driving, precise recognition and interpretation of objects and traffic situations is essential. For this purpose, vehicle manufacturers use a variety of sensors such as radar, lidar and cameras, from whose data important environmental information is generated, often with AI algorithms. To train these AI algorithms, large amounts of sensor data must be annotated. In data annotation, relevant sensor information is specially marked so that the AI of the automated driving function understands what it is supposed to process. The annotation takes place via labels, which are often inserted manually into the record. These activities are associated with considerable costs and time and are becoming an ever greater challenge.

Neural Automated Labeling

The new Neural Automated Labeling from Bosch Engineering enables the efficient and fast fully automated annotation of sensor data in 3D. Previously, automated labeling was limited to 2D data, and with the help of modern AI methods, the application now generates precise 3D labels for reliable object recognition and tracking through data reconstruction. This creates high-quality ground truth data as the basis for training and validation of AI algorithms for automated driving.

Automated creation of 3D bounding boxes / 3D labels

Bosch Engineering is working on a fully automated 3D labeling solution that revolutionizes the annotation process. The focus is on 3D bounding boxes / 3D labels. 3D bounding boxes are object-enclosing cuboids that can be used to precisely locate objects in images or point clouds. Unlike 2D labels, which only measure height and width, they also provide information about the depth and position of an object in space. Especially in the context of automated driving, this additional spatial dimension is crucial for precise environmental detection, improved decision-making and safe interaction. Depending on customer requirements, Neural Automated Labeling also enables data selection and semantic search and segmentation.

NAL3D

Lower costs, fewer errors, faster results

With Neural Automated Labeling from Bosch Engineering, even large amounts of data can be processed quickly, efficiently and with a low error rate in the cloud. Through a simple API access, it provides seamless integration into the customer’s software toolchain.

The three main advantages:

  • Faster labeling and easier scaling for high data volumes
  • Higher consistency and fewer errors in the label record
  • Considerable cost reduction: For a specific application, the effort was 4 times lower than for manual annotation
Contact

Your contact

Do you have any questions? Feel free to contact our expert.

Guillaume Bernhard, Key Account Manager

Get in touch