Automated Training Data Generation for AI Based on Perspective Transformation and Image Synthesis

Automated parking systems play a vital role in smart cities. Nevertheless, existing image recognition technologies heavily rely on manual annotation, which is costly and struggles to handle parking slot distortions under varying viewpoints. To address this issue, this study proposes an automatic annotation method based on perspective transformation and image synthesis to generate training data that closely simulates real-world scenarios. By computing the transformation between top-down images at a 90° and those at a θ°, we establish a perspective transformation matrix (H) to simulate geometric distortions of parking slots from different angles and positions. This transformation is combined with a variety of background images, and image synthesis technology is used to generate non-rectangular parking slot with diverse backgrounds, thereby enhancing the generalization ability of the model. Our method also enables automatic annotation generation, significantly improving the accuracy of parking slot recognition models while reducing the burden of manual data collection and annotation. This provides an efficient solution for advancing automated parking systems.