Data Synthesis and Validation Methods for Parking-Slots

Parking-slot detection plays a crucial role in autonomous driving self-parking systems. To simulate more complex parking scenarios in datasets and alleviate the challenges of manual annotation, we have developed several methods for synthesizing parking-slots under various conditions. This paper introduces three such methods: synthesizing parking-slots in AVM (around view monitor) images, generating parking-slots in fisheye images, and incorporating 2D symbols within parking-slots to create special types. To address the interference caused by lane markers and original parking-slots and to validate the effectiveness of synthetic data on public datasets, we propose a method to remove them through semantic segmentation and mask inpainting. In parking-slot detection experiments, the accuracy rate and recall rate increase by 0.44% and 0.91% respectively when we combine synthetic data with original training data. Furthermore, in another experiment involving the detection of new parking-slots with synthetic symbols, we achieve an accuracy of 96.82% and a recall of 93.83%. In order to evaluate the interference caused by synthetic data, we conduct semantic segmentation experiments on another benchmark dataset. Under the presence of synthetic parking-slots with the same markings as lane markers, the MIoU (Mean Intersection over Union) and the IoU of lane markers decrease by 0.23 and 0.57 respectively.