Transformer-Based Parking Slot Detection Using Fixed Anchor Points

Transformer-based architectures have recently gained significant attention in various computer vision tasks. Their ability to capture non-local dependencies and intricate characteristics makes them a promising complement to CNNs. However, their application in parking slot detection tasks is still limited. Thus, this paper proposes an appropriate way to apply transformer-based architectures to parking slot detection tasks. The proposed method adopts the Detection Transformer (DETR) architecture, which employs a standard transformer encoder-decoder framework. Since this approach requires a long training time, this paper suggests utilizing fixed anchor points to replace object queries in the original DETR architecture. Each anchor point is assigned a known location and focuses only on a predefined area of the feature map, resulting in a considerable reduction in training time. In addition, this paper suggests using a more suitable and efficient two-point parking slot representation to improve detection performance. In experiments, the proposed method was evaluated with the public large-scale SNU dataset and showed comparable detection performance to the state-of-the-art CNN-based methods with 96.11% recall and 96.61% precision.