Parking Slot Detection Based on Wavelet Transform and Subgraph Aggregation

Parking slot detection plays a crucial role in autonomous driving systems. Most existing approaches adopt vision-based schemes, where parking slots are inferred from parking slot marking points. However, these methods often suffer from unreliable recognition due to the high overlap between marking points and complex background textures. Furthermore, many existing approaches heavily rely on hand-crafted rules and post-processing heuristics, which restrict their ability to fully capture the relative spatial relationships among marking points. To address these limitations, we propose WTSA-Net, a novel parking slot detection network based on wavelet transform and subgraph aggregation. In the marking point detection stage, the process is divided into two parts: marking point feature extraction and marking point identification. In the marking point extraction part, potential marking point features are searched through multi-scale receptive fields. In the marking point identification part, learnable features are decomposed into different frequency bands via wavelet transform to capture subtle cues that distinguish true marking points from background noise. In the parking slot inference stage, we introduce a subgraph-aggregation-based parking slot inference method. This method exploits the geometric relationships among parking slots to sample candidate points that may constitute a slot and form subgraphs. A graph neural network is then employed to aggregate both intra-subgraph and inter-subgraph neighborhood information of labeled points, followed by an artificial neural network for output decoding to obtain the final parking slot information. Extensive experiments conducted on the public PS2.0 dataset and the private WL dataset validate the effectiveness of WTSA-Net, demonstrating its superior performance in parking slot detection.