Parking Space Detection — Project presentation

Project presentation. # Overview **Parking Space Detection** is a vision-based smart-parking system that tells a driver *which* parking spaces are free, not merely how many. It reads an ordinary surveillance-camera feed and classifies every marked space in the frame as **Empty** or **Occupied** in real time, so a free spot is located rather than hunted for. Circling a crowded lot — a mall, a cinema, a university campus — costs time, fuel and temper, and is itself a source of congestion and emissions. Counter-based systems publish only a total. Sensor-based systems need one ground sensor per space plus the installation work that comes with it. A single camera already overlooking the lot is cheaper, covers many spaces at once, and is what this project builds on. ## How it works - **Detector:** YOLOv3 (Darknet) — a 106-layer fully convolutional network that detects at three scales (strides 32 / 16 / 8, giving 13×13, 26×26 and 52×52 grids) with 9 anchor boxes, run through OpenCV's deep-learning module in Python. - **Classes:** two, `Empty` and `Occupied`, so each YOLO layer uses `filters = (classes + 5) × 3 = 21`. - **Transfer learning:** training starts from the COCO pre-trained weights (80 classes, 80,000+ images) instead of from scratch. - **Training configuration:** batch 64, subdivisions 16, learning rate 0.001, saturation 1.5, exposure 1.5, input resized to 608×608. Training is stopped where the average loss flattens, and the checkpoint with the best mAP is the one kept. - **Output:** the detections are drawn back over the video in OpenCV, each box labelled with the state of that space. ## Dataset and annotation Built on **PKLot** — 12,416 parking-lot frames and 695,899 segmented parking spaces captured at UFPR and PUCPR (Curitiba, Brazil) across sunny, cloudy and rainy days. The original rotated-rectangle annotations were converted to axis-aligned bounding boxes, and the images were re-labelled into the project's two classes with **LabelImg**, exported in YOLO format (`class x y w h`, normalised to the image width and height). ## Why a CNN A convolutional network models the spatial correlation between neighbouring pixels, which is what lets it hold up against the things that break simpler parking detectors: cars partially occluding one another, shadows, changing light, non-uniform car sizes and vehicles parked off-centre. Training is expensive; inference afterwards is fast enough for modest hardware such as a Raspberry Pi. ## Limitations and future work Detection quality depends on how close the deployed scene is to the training data — a camera angle or lighting condition absent from the dataset degrades it, and resolution figures shift with the light. The natural extensions are a larger and more varied dataset, and routing the driver to the nearest free space rather than only marking it. ## Team Islam Mohamed Mohamed Soliman · Mohamed Maghawry Ismail · Mahmoud Osama Mohamed · Mahmoud Rezk Saad · Mahmoud Abdulrahman Badr Supervised by Ass. Prof. Gamal Tharwat — Systems and Computer Engineering, Faculty of Engineering (Boys), Al-Azhar University. Defended 7 July 2023.