Real-Time Instance Segmentation for Parking Occupancy Detection Using CNNs

In the context of ongoing urbanization and increasing vehicle density, the importance of accurate monitoring of parking spaces is becoming increasingly evident. The presented work focuses a camera-based system for real-time parking occupancy detection, which is based on single-stage instance segmentation convolutional neural networks (CNN). The objective is to achieve precise detection of vehicles and an efficient assessment of parking occupancy. The innovative approach incorporates a weighted segmentation of parking areas, which facilitates enhanced detection of vehicles parked over the line and a more precise determination of occupancy. Moreover, the models performance was evaluated for real-time applicability on diverse hardware devices, encompassing both high-performance systems and resource-constrained edge devices. The system was evaluated in a simulated model environment to test its robustness against a variety of scenarios. The investigation focused on diffuse and high-contrast lighting conditions, which reflect typical conditions in parking areas. To evaluate the system’s performance, a variety of YOLO-based (You Only Look Once) CNN models were employed, optimized for both vehicle detection and occupancy analysis. The findings indicate that the system displays a high degree of accuracy in assessing vehicles and occupancy in environments with diffuse lighting. This suggests that the system represents a promising solution for scalable, cost-efficient real-time applications in the field of parking space monitoring.