Data Synthesis Methods for Parking-Slot Detection 2023-01-7052
Parking-slot detection plays a critical role in the self-parking system for
autonomous driving. To enhance the complexity of the environmental situations in
parking-slot datasets and reduce the difficulty of manual annotation, we design
several data synthesis methods to generate new parking-slots under different
situations. Methods introduced in this paper include synthesizing parking-slots
in AVM (around view monitor) images, generating parking-slots in fisheye images
and adding 2D symbols inside parking-slots to form special ones. To test the
influence of our synthetic data, we conduct a series of experiments on different
tasks. In the parking-slot detection experiments, we design a novel two-stage
parking-slot detection method. We use YOLOv7 as the object detector and
different from previous methods, we detect the complete parking-slots and
marking points at the same time. Then we match marking points and give them a
certain order in the second stage. We achieve accuracy of 80.28% and recall of
79.94% on our own data which shows the effectiveness of our method. Then we
achieve accuracy of 80.98% and recall of 80.48% with supplementation of
synthetic parking-slots data, slightly better with no extra manual annotation.
Next, we achieve accuracy of 96.78% and recall of 93.68% in another parking-slot
detection experiment for the new type of parking-slots with synthetic symbols
inside them. To test the generalization performance of our synthetic data, we
conduct semantic segmentation experiments on public dataset. MIoU (Mean
Intersection over Union) and the IoU of lane markers decrease by 0.23% and 0.57%
respectively under the interference of synthetic parking-slots when parking-slot
lines set the same as lane markers.