AI-based recognition of car brands and years of manufacture using synthetic data

Initial situation

»Vehicle Make and Model Recognition« (VMMR) - the fine-grained recognition of vehicle makes, models, years of manufacture and paint colors in image data - is used to recognize vehicles even in cases where the license plate is not clearly recognizable or potentially falsified.

Using modern AI methods, hundreds of vehicle models can be classified in video data in real time or recognized in specific search queries. However, a particular challenge, especially for critical applications, is the accuracy and timeliness of the data sets used to train the classification process: Teaching vehicles requires high-quality training data in large quantities. The collection of real data, which is annotated by humans, is error-prone, data protection-critical and involves a great deal of effort. Accordingly, current vehicle models are only reliably recognized late in current approaches.

© Fraunhofer IOSB
Result of the vehicle classification.

Expertise

As part of several research and development projects, including the "FeinSyn" project funded by the state of Baden-Württemberg, an approach has been developed to train AI-based VMMRs on simulated, i.e. synthetic, training data. New or specific vehicle models can thus be added "on demand" with just a few mouse clicks.
With »Synset Boulevard« (https://synset.de/) the first large synthetic data set for VMMR worldwide was created, evaluated and made available as an open data set as part of the "FeinSyn" project. The dataset was used to show that the approach achieves equivalent results to training on real data - but with considerably less effort and significantly greater timeliness.

Added value

The approach developed combines state-of-the-art AI methods for "Vehicle Make and Model Recognition" with tools for the synthetic generation of training data. This enables the reliable and flexible recognition of vehicles with full transparency and control over the data sets. New models can be added even before they are used in traffic, as can any country-specific models.
In particular, the approach enables a very high specificity of recognition, right down to specific facelift versions or equipment variants.
The methods were tested in use with the Videmo 360 software for merging and evaluating data from a wide variety of camera data streams. In conjunction with the Ternica VES camera system, the data can be captured flexibly from the emergency vehicle and the information can be merged via the cloud.

Infobox/Key Facts

Recognition of hundreds of vehicle makes, models and years of manufacture in video data
Maximum specificity, flexibility and reliability through the use of synthetic rather than real data
Quality of recognition equivalent to AI methods trained on real data

Video evaluation systems department

Would you like to find out more about our video analysis products? Then visit the page of our department.

To the department page

Interested in a cooperation?

Are you interested in cooperating with our department or do you have questions about our current research topics? Then get in touch with us!

Send e-mail