An Enhanced YOLOv8‐3D Network for Parking Lot Classification

Parking lot classification is a critical precursor task in autonomous parking systems. However, the accuracy and robustness of this task face significant challenges due to factors such as data imbalance, lighting variations, and environmental complexity. This article proposes an enhanced parking lot classification method based on an improved YOLOv8‐3D network, aiming to achieve higher classification accuracy. First, the 2D convolution layers in YOLOv8 are replaced with 3D convolution layers to enhance the spatial feature extraction capability for multiperspective input images. Second, an edge detection algorithm is introduced to extract edge information, such as parking space lines and building contours, enriching the network’s geometric feature representation. To address low‐light environments, an adaptive brightness enhancement method based on a luminance threshold is proposed, effectively improving classification accuracy under insufficient lighting conditions. Furthermore, a rigorous data partitioning and filtering strategy is designed to mitigate the issue of data imbalance. Experimental results on a newly standardized dataset demonstrate that the improved YOLOv8‐3D model achieves significant performance gains across all classification categories, with average accuracy surpassing other comparative models. These results validate the effectiveness and superiority of the proposed method in complex scenarios.