Abstract

This paper presents NOVA Parking, a transformer-based smart parking monitoring system. Conventional convolutional neural network (CNN) detectors perform poorly when deployed in unfamiliar parking lots with varying viewpoints, markings, and lighting conditions. To address this issue, NOVA Parking employs Grounding DINO, an open-vocabulary vision-language detector that aligns text prompts with image regions for robust detection across different environments. The system integrates closed-circuit television (CCTV) acquisition, intersection-over-union (IoU)-based tracking, region-of-interest (ROI)-based parking violation analysis, a Segment Anything Model (SAM)-assisted automatic labeling pipeline, and web/mobile services with real-time push notifications. On a dataset of 14,908 images from three locations, the proposed approach achieves a mean Average Precision (mAP) of 95.23% at a threshold of 0.5 for trained lots and 94.12% for unseen lots, retaining approximately 99% of the trained-lot mAP. In comparison, YOLOv12 and Faster R-CNN achieve mAPs of 81.67% and 78.34%, respectively. A five-year analysis further indicates that NOVA Parking has a total cost of ownership that is roughly 71% lower than that of magnetometer-based systems, confirming the practicality and economic feasibility of NOVA Parking for intelligent transportation systems (ITS).