Dynamic High-accuracy Parking Occupancy Detection Using Lightweight CNN-Based Framework and Advanced Image Processing for Smart City Integration

The increasing vehicle density is making it more difficult to manage urban parking efficiently, which leads to wasted time, traffic jams, and needless fuel consumption. Uneven lighting, shadows, occlusions, and other environmental factors frequently make conventional parking monitoring techniques less effective in dynamic real-world situations. In this study, a lightweight convolutional neural network (CNN) based on the MobileNetV2 architecture is used to develop a real-time smart parking management system that uses Closed-Circuit Tele-Vision (CCTV) video streams to detect parking space occupancy. Using a meticulously prepared dataset of extracted and labeled parking space images, the system applies sophisticated image preprocessing techniques, such as color conversion, Gaussian blur filtering, adaptive thresholding, median blur filtering, white pixel counting, and parking space coloring. Based on experimental results, the suggested model demonstrated robustness and balanced performance with an overall accuracy of 98.02%, precision of 98.94%, recall of 98.25%, F1-score of 98.60%, specificity of 97.48%, and AUC-ROC and AUC-PR values above 0.99. High predictive certainty, few misclassifications, steady training convergence, and peak performance were all displayed by the system at decision thresholds ranging from 0.3 to 0.5. The method is easily adaptable for integration into smart city infrastructures and has been validated through dynamic monitoring of multiple parking lots. It allows for scalable, reliable, and real-time parking occupancy/availability detection, which can improve the driving experience in urban areas, minimize traffic, and maximize space utilization.