Virtualizing real-world parking scenarios with DeepScenario’s AI pipeline
Developing autonomous parking functions requires vast amounts of diverse, high-quality data, which forms the basis for a variety of use cases ranging from requirements engineering and virtual testing to model training. To meet this need at scale, DeepScenario enables the reconstruction of parking scenarios from video data through its flexible virtualization pipeline. The company’s software can process footage from any monocular camera, including drones, dashcams, or surveillance cameras installed in parking lots. It virtualizes video streams regardless of camera perspective and extracts the 3D information of all dynamic objects in the scene, resulting in centimeter-accurate trajectory data. In addition, the software also reconstructs the static environment with remarkable accuracy, producing a georeferenced 3D mesh that can serve as the basis for HD map generation.
Drone recordings in particular massively accelerate the collection of parking data. For instance, in DeepScenario’s Fabulous Sindelfingen dataset, approximately 200 parking spaces were recorded simultaneously, resulting in over 450 parking scenarios from only 3 hours of video footage. This and multiple other parking datasets are available for download on DeepScenario’s web app, and customers can also upload their own videos to be processed through the company’s virtualization pipeline.
Mining parking scenarios for simulation
Large volumes of trajectory data make manual extraction of relevant scenarios very time-consuming. DeepScenario addresses this challenge with an intuitive scenario mining tool that automates this process. The tool can filter for specific parking maneuvers based on a set of configurable parameters such as the duration to park or the number of direction switches. Users can further refine their selection by distinguishing between park-in and park-out scenarios to precisely match their testing objective.
For simulation, the desired scenarios can be exported directly into the ASAM OpenSCENARIO standard that enables a seamless integration with major simulation platforms. This allows developers to benchmark their parking algorithms in real-world scenarios.
With DeepScenario’s end-to-end toolchain for autonomous parking development, the company enables its customers to accelerate the entire data loop, from real-world measurement and scenario mining to simulation. The toolchain reduces development time, enhances safety, and ensures robust performance across diverse parking environments.
