Enhancing parking space recognition from surveillance cameras

The development of accurate parking spot detection systems is essential for intelligent transportation infrastructure. Conventional approaches, such as those employing convolutional neural networks, often face challenges in addressing complex conditions, including variable lighting, suboptimal camera perspectives, and limited capacity to model global spatial contexts. To address these limitations, this study introduces a novel Transformer-based architecture that utilises the self-attention mechanism to effectively capture long-range dependencies and comprehensive spatial relationships. The proposed model integrates data augmentation, fine-tuning, query optimisation, and feature extraction to enhance detection accuracy. Evaluation on a real-world parking dataset shows that the model achieves a mean Average Precision of 98.1%, which exceeds the results of conventional methods based on convolutional neural networks.