The progress of the jbDATA Smart Data Loop and the latest research results from the project were presented by the project partners on 12/13 June 2024.
At the 3rd consortium meeting of just better DATA, all project partners gathered at the FZI in Karlsruhe to exchange the latest research results, plan future steps, and discuss challenges on the project roadmap. The focus was on the jbDATA Smart Data Loop, innovative approaches to integrating sensors, improving data quality by combining real and synthetic datasets and developing intelligent triggers for recording relevant road scenarios.
The meeting provided an update on the current status of the jbDATA project, among other things. With great commitmentand interdisciplinary collaboration, the project has achieved impressive initial research results that have the potential to fundamentally change the future of intelligent transportation systems.
The jbDATA project aims to develop an advanced demonstration environment that integrates both synthetic and real data from all partners. Three vehicles from the partners collect real data, while another team combines this data with synthetic inputs to complement the jbDATA Smart Data Loop. Intensive research is being conducted to identify triggers for recording relevant road scenarios, such as corner cases and specific weather conditions.
The integration of sensors in the recording vehicles requires careful planning regarding hardware positioning, mounting, and cable routing. How LiDAR, radar, and cameras can be integrated will be discussed later in the project. It will also be discussed how real-time recording with a unified timestamp across various systems can ensure the coherent operation of the vehicle. Recording can be triggered by the detection and prediction of uncertainties, such as critical image situations and sensor anomalies. The jbDATA vehicles from AVL, b-plus, the FZI, THD, and Valeo will be equipped to apply these methods.
In jbDATA, a uniform containerized environment based on ROS2 (Robot Operating System) is being established, which forms the basis for data processing within the vehicle. This environment allows the utilization of hardware acceleration on various systems, enabling the developed algorithms to be executed as efficiently as possible. These algorithms are designed to recognize selected environmental conditions (e.g., weather) based on current sensor data and to trigger the recording of relevant data. Precise calibration of the vehicle’s sensors is crucial for the quality of the generated data. Therefore, a mobile setup is being developed within the project to enable quick and efficient sensor calibration.
Image augmentations expand the training set with modified images, thereby improving model performance. Comparing many sequences under different weather conditions aims to create comprehensive datasets that cover all scenarios, including varying rain intensities and corresponding road conditions. These data are crucial for closing domain gaps in model training.
jbDATA focuses on developing advanced methods for generating synthetic data tailored specifically to underrepresented scenarios. By closing these gaps, we increase the robustness and generalisation ability of AI models, ensuring they work effectively in diverse real-world conditions. These synthetic data improve the efficiency and completeness of training datasets, leading to more accurate and reliable AI predictions. Ultimately, the jbDATA project work bridges data deficiencies and promotes more inclusive and adaptable AI systems.
As part of the Smart Data Loop, jbDATA’s metadata storage supports the enrichment of labels, annotations, and attributes for machine learning scenarios. APIs allow raw and synthetic data to be combined to enrich images, while a user interface enables data filtering by tags such as date or vehicle type.
The jbDATA project continues to set new standards in the integration of vehicle data, combining synthetic and real data for smarter and more responsive traffic systems. Through collaborative efforts and cutting-edge technology, jbDATA is ready to make a significant impact on the future of intelligent transport.
Images: eict