GridReconNet: a grid-structured light-based model for 3D obstacle detection in automated parking systems

Abstract Accurate obstacle detection is crucial for automated parking systems (APS), yet existing methods often face limitations such as low accuracy in ultrasonic sensors, the inability to capture three-dimensional (3D) information with vision-based techniques, and the high costs and complexity associated with multi-sensor fusion. This study introduces GridReconNet, a novel 3D obstacle detection model for APS, which combines grid-structured light projection with deep learning techniques. The model integrates recurrent residual convolutional units, deformable convolutions, and attention mechanisms, significantly enhancing feature extraction, reducing noise, and enabling precise 3D depth reconstruction from two-dimensional images. Experimental results demonstrate that GridReconNet achieves a 122.0% improvement in structural similarity index measure and a 15.6% improvement in peak signal-to-noise ratio compared to the baseline UNet, outperforming R2U-Net and DUNet models. While introducing a moderate increase in computational complexity, the approach offers significant improvements in accuracy and robustness, providing an efficient and practical solution for 3D obstacle detection in automated parking environments. Future research will focus on optimizing the model for lightweight applications and enhancing its generalization in dynamic and complex parking scenarios.