Vision-Based Smart Parking Occupancy Detection Using Deep Learning Object Detection Models

Finding parking in urban areas is becoming increasingly challenging as cities continue to grow. Inefficient parking utilization, excessively high infrastructure costs, and the limited ability of traditional sensor-based monitoring systems to scale make finding a solution difficult. The proposed parking occupancy detection framework is vision-based and uses the YOLOv8 object detection model to provide real-time identification of occupied and vacant spaces. The model was trained and tested using a dataset that consisted of a set of about 6,850 annotated images collected from the PKLot benchmark dataset as well as images collected from real-time parking lot monitoring on a college campus in the USA. A complete analysis of YOLOv8 models was performed by examining data augmentation techniques, robustness of YOLOv8 under different environmental conditions, and generalizability of YOLOv8 across different datasets. Results from several experiments suggested that the YOLOv8m model produced the best balance between accurate detection and computational efficiency with a mean Average Precision (mAP@0.5) of 99.4% and was capable of providing real-time inferences. Moreover, YOLOv8 demonstrated robust performance in low-light, blurry, and noisy environments when combined with different synergistic combinations of data augmentation techniques. The proposed parking occupancy detection framework represents an accurate, scalable, and cost-effective alternative to legacy parking monitoring solutions, making it appropriate for implementation in smart city and edge computing applications.