SnailMogur/parking-lot-t10-dataset

Parking-Lot-T10 Dataset

A labeled, real-world parking-lot dataset captured from a single fixed camera overlooking parking lot T10 on the campus of Brno University of Technology (BUT) in Brno, Czech Republic. Frames were collected across many days at 10 minutes period and times of day (varying light, weather and occupancy) and hand-organized into occupancy labels.

It is intended as a compact, PKLot-style benchmark for Deep CNN Machine Learning training datasets to detect and classify parking lot occupancy. Dataset were labeled for both CNN classifier models and YOLO object dection.

The dataset ships in ready-to-train form:

• —Classification — cropped images in PKLOT-style intended for CNN such as MobileNet format with vacant / occupied (t10lot_labeled/crops_classifiers/)
• —Detection — full frames in YOLO format with vacant / occupied boxes (t10lot_labeled/crops_yolo_detect/)

• —Raw captures — the original timestamped frames the labels are derived from (DATA/)

Source & attribution

• —Site: parking lot T10, Brno University of Technology (BUT), Brno, Czech Republic
• —Capture: a single fixed Raspberry Pi camera on the BUT campus, one still every 10 minutes during daylight hours (see `PYTHON/camera.py`)
• —Authors / maintainers: Main: Tomáš Fryža (Brno University of Technology) Labeling: tymfly7 (github)

• —License: MIT — © 2026 Tomáš Fryža

If you use this dataset in academic work, please credit Brno University of Technology and the author, and link back to this repository.

Contents at a glance

Both labeled datasets are derived from the same source captures held in DATA/, so a detection frame and the classification crops taken from it are consistent.

Directory structure

parking-lot-t10-data/
├── DATA/                            # ── raw captures ──
│   └── YYYY-MM-DD/                  #   one folder per capture day
│       └── YYYY-MM-DD_HHMM.jpg      #   ~one still every 10 min
├── PYTHON/                          # ── capture tooling (Raspberry Pi) ──
│   ├── camera.py                    #   picamera2 still-capture loop
│   └── test_time.py
├── t10lot_labeled/                  # ── labeled datasets ──
│   ├── crops_classifiers/           #   classification dataset (per-slot crops)
│   │   ├── occupied/  *.jpg          #     61,057 cropped slots
│   │   └── vacant/    *.jpg          #     36,152 cropped slots
│   └── crops_yolo_detect/           #   detection dataset (YOLO)
│       ├── dataset.yaml             #     Ultralytics data config (nc=2)
│       ├── images/{train,val,test}/ #     3,137 frames, split 2195/470/472
│       └── labels/{train,val,test}/ #     one .txt per image (YOLO format)
├── LICENSE
└── README.md

Getting the data

Clone the repository (it is a standalone data repo, kept out of any application repo):

git clone https://github.com/tymfly7/parking-lot-t10-data ~/parking-lot-t10

Cloning under ~/parking-lot-t10 makes the paths in dataset.yaml (below) work as-is. If your checkout lives elsewhere, edit path in dataset.yaml to match.

Classification dataset (t10lot_labeled/crops_classifiers/)

Each file under crops_classifiers/occupied/ or crops_classifiers/vacant/ is a single parking slot cropped from a full frame using its annotated polygon. The folder name is the label — a PKLot-style layout ready for a torchvision ImageFolder loader and CNN backbones such as MobileNet.

from pathlib import Path
from torchvision import datasets, transforms

tf = transforms.Compose([
    transforms.Resize((96, 96)),
    transforms.ToTensor(),
])
root = Path("~/parking-lot-t10/t10lot_labeled/crops_classifiers").expanduser()
ds = datasets.ImageFolder(root, transform=tf)
# ds.classes -> ['occupied', 'vacant']

Detection dataset (t10lot_labeled/crops_yolo_detect/)

Standard Ultralytics YOLO layout: each image has a matching .txt label file with one row per box: class cx cy w h (normalized 0–1). Classes: 0 = vacant, 1 = occupied.

dataset.yaml:

path: ~/parking-lot-t10/t10lot_labeled/crops_yolo_detect   # ~ expands to your home dir
train: images/train
val:   images/val
test:  images/test
nc: 2
names: {0: vacant, 1: occupied}

`path` is device-agnostic: ~ expands to each user's home directory (Ultralytics calls .expanduser()), so it works on any machine where this repo is cloned under ~/parking-lot-t10. If your checkout lives elsewhere, edit path to match.

Train:

yolo detect train model=yolo11n.pt data=~/parking-lot-t10/t10lot_labeled/crops_yolo_detect/dataset.yaml imgsz=640 epochs=100

Raw captures (DATA/)

The DATA/ tree holds the original frames straight from the camera, grouped into one folder per capture day (YYYY-MM-DD/). Filenames encode the capture timestamp (YYYY-MM-DD_HHMM.jpg). These are the source of truth from which the labeled YOLO frames are derived, so the labeled dataset is reproducible from them.

Capture is performed on the Raspberry Pi by `PYTHON/camera.py`, which takes a full-resolution still roughly every 10 minutes and writes it into the current day's folder.

Annotation format

• —`t10lot_labeled/crops_classifiers/{occupied,vacant}/` — classification labels encoded by folder name (one crop per file).
• —`t10lot_labeled/crops_yolo_detect/labels//.txt`* — YOLO labels, one file per image, one class cx cy w h row per annotated parking space.