An intelligent parking system based on deep learning an‎d image processing using commercial drones aerial images

1/19/2026 12:00:00 AM

This study presents an integrated vision-based system for intelligent bus parking management at the Barekat terminal (Mehran city), aimed at detecting vacant bays an‎d guiding buses to the nearest available spot. The system input consists of drone-captured videos providing an overhead view of the parking area, which, after refinement an‎d selec‎tion, were employed for training an‎d eva‎luation. The proposed pipeline comprises four main stages: (1) bus detection using the YOLOv8/11/12 families with Oriented Bounding Box (OBB) outputs an‎d the application of SAHI for window-based tiling of large images (640-pixel windows with 20% overlap), (2) scene geometry reconstruction via a homography matrix an‎d mapping of reference-frame labels onto the current frame, (3) occupancy estimation of each bay based on the overlap ratio between the bay polygon an‎d the dilated bus mask (employing morphological dilation) with a decision threshold of 0.10, an‎d (4) cost-aware path planning using the A* algorithm on a cost grid, incorporating turning costs. The training dataset comprised 695 annotated images, which were expan‎ded to 1,359 images through data augmentation techniques including 15° rotations, brightness variations, an‎d the addition of ran‎dom noise. Additionally, four fully annotated images were prepared for eva‎luating geometric alignment. Among the tested models, YOLOv8-Large achieved the best performance with mAP@[50:95] ≈ 0.911; meanwhile, YOLOv11-Small an‎d YOLOv11-Medium reached approximately 0.901 an‎d 0.905, respectively. In terms of inference speed, YOLOv8-Nano demonstrated the lowest runtime on GPU, while YOLOv12-Nano achieved the lowest runtime on CPU. The highest precision was obtained with YOLOv11-Nano. In the alignment stage, the combination of CLAHE preprocessing with KAZE feature extraction achieved the highest mean IoU for label mapping, whereas for the “road” class, the combination of histogram equalisation (HE) with RootSIFT yielded superior performance.

ناوبري بصري , پاركينگ هوشمند , يادگيري عميق , پردازش تصوير , تصاوير هوايي

Visual Navigation , Inteligent Parking , Deep learning , Image Processing , Arial Images

Vahid Rahimi Bafrani

Dr. Ahmad Ayatollahi