End-to-end Style Parking Algorithm Development Pipeline Based on Procedural Parking Scenario Synthetic Data Generation

Less costs and higher efficiency may be the constant technological pursuit of engineers. Despite the great success, data-driven AI development still requires multiple stages such as data collection, curation, annotation, training, and deployment to work together. We expect an end-to-end style development process that can integrate these processes, achieving the an automatic data production and algorithm development process that can complete with just one click of the mouse. For this purpose, we explore an end-to-end style parking lot algorithm development pipeline based on procedural parking scenario synthetic data generation. Our approach allows for the automated generation of parking scenarios according to input parameters, such as scene construction, static and dynamic obstacles arrangement, material textures modification, and background changes. It then combines with the ego-vehicle trajectories into the scenarios to render high-quality images and corresponding label data based on blender render engine. We conducted experiments based on our own vehicles, automatically generated 200 parking scenes, combined expert controllers on the SIL platform and offline optimization algorithms to generate 1K parking trajectories, and render 200K frames end-to-end parking data. Finally, the generated data was feed into the follow up end-to-end parking algorithm and applied in real vehicles to park the vehicle into specific parking lots. The experimental results demonstrate that our goal has been basically achieved, with only mouse clicks to develop a parking function.