The advancement of autonomous driving perception frequently necessitates the
aggregation of data, its subsequent annotation, the implementation of training
procedures, and other related activities. In contrast, the utilisation of
synthetic data obviates the necessity for data collection, annotation, and the
generation of accurate and reliable labels. Its incorporation into the
development process is anticipated to streamline the entire algorithmic
development process. In this study, we propose a novel approach utilising the
Blender software to create a virtual representation of an underground car park
and develop an automated parking dataset. The utilisation of virtual simulation
technology enables the generation of diverse and high-quality training data,
thereby addressing the challenge of acquiring data in the actual scene. The
experimental results demonstrate that the model trained based on the synthetic
dataset exhibits superior performance in the automatic parking task, thereby
substantiating the efficacy and practicality of the proposed approach.
Furthermore, previous research on synthetic data has often concentrated on the
creation of the final perceptual algorithm dataset, without providing access to
the material and method used to generate the synthetic data. This study
therefore makes the underground car park scene library files available under an
open-source licence, in the hope that subsequent developers will be able to
generate the required datasets based on these materials with greater freedom.The
open-source scene library, code and dataset provided in this paper are as
follows:
https://github.com/Snyard/parkinglot-generator.