Parking–Occupancy Detection Through Adaptive Multisensor Camera-CNN Fusion

Abstract

A robust multicamera sensor system for parking–occupancy detection is introduced, combining convolutional neural networks with an adaptive fusion mechanism that leverages angular diversity. The proposed pipeline integrates viewpoint-specific bounding-box components and a distortion–reduction module that compensates for perspective-induced deformations. Under different azimuth angles and illumination conditions, including overcast, sunny, and nighttime scenarios, the fusion approach consistently outperformed single-camera systems. Notably, fusing cameras at 0$^\circ$ and 90$^\circ$ yielded an intersection-over-union (IoU) of 0.898 without correction, while the distortion–reduction module improved IoU from 0.734 to 0.856 in geometrically challenging cases. The method also maintained robust performance in low-light environments, where individual camera views degraded. Designed for scalability and minimal calibration effort, the architecture supports geometry-consistent localization across multiple sensor perspectives. These results demonstrate that combining angular fusion with correction-aware processing offers substantial gains in precision and robustness. The system is particularly suited for real-world deployment in smart parking applications under complex environmental conditions.