license: cc-by-4.0
task_categories:
- object-detection
- video-classification
pretty_name: >-
PMOF: A Dataset and Benchmark for Passenger Monitoring Using Overhead Fisheye
Cameras
size_categories:
- 10K<n<100K
extra_gated_prompt: >-
You agree to not use the dataset to conduct experiments that cause harm to
human subjects.
extra_gated_fields:
Company: text
Country: country
I want to use this dataset for:
type: select
options:
- Research
- Education
- label: Other
value: other
I have read and understood the dataset license: checkbox
Dataset Card for PMOF:
Dataset Description
PMOF (Passenger Monitoring using Overhead Fisheye Cameras) is a public dataset for passenger monitoring from top-view fisheye cameras mounted inside a moving vehicle. It contains more than 19,000 manually annotated frames with rotated bounding boxes, tracking identifiers, and action labels, supporting research in object detection, multi-object tracking, and action recognition. The dataset supports both frame-based and video-based tasks, enabling research on individual-frame perception as well as temporal analysis across video sequences. Unlike existing overhead fisheye datasets captured in static environments, PMOF captures the domain shift caused by vehicle motion, including changing backgrounds, illumination, occlusion, and constrained viewpoints. The dataset is intended for developing and evaluating robust fisheye-based perception methods for public transport and related monitoring applications.
Published by: Stella Katharina Wermuth, Qazi Arbab Ahmed, Klaus Neumann, and Thorsten Jungeblut
License: CC-By 4.0
Landing Page for Paper, Dataset and Code: PMOF Landing Page
Code: Github Repo
Funded by: This work was funded by the projects enableATO (German Federal Ministry of Transport, German Center for Future Mobility DZM, grant: 19DZ23002D), FH-Personal (German Federal Ministry of Research, Technology and Space (BMFTR), grant: 03FHP106) and KI-Akademie OWL (BMFTR, supported by VDI/VDE Innovation+Technik GmbH, grant: 16IS24057C).
Dataset Creation
Acquisition
Data were collected in May 2025 during demonstration runs of the autonomous rail vehicle Monocab. Participation was voluntary, and all participants provided written informed consent. No behavioral instructions were given, resulting in natural behavior, although some participants deliberately performed actions such as falling, covering themselves, or fighting to broaden the range of captured activities.
Recordings were captured using a Vivotek FE9180-H-v2 fisheye camera mounted on the vehicle ceiling at 1.5m, providing a 180◦ top-down view. The camera operated in WDR-Pro mode and recorded at a resolution of 1920×1920 pixels with a variable frame rate of 8–15 fps. Each recording corresponds to a single ride with a unique passenger group, at speeds of up to 20 km/h. Additional empty-cabin frames were collected during vehicle shunting to provide background-only samples.
All frames were manually reviewed, and those showing individuals outside the vehicle or visible phone screens were removed. Temporal continuity is largely preserved, with only minor gaps due to camera dropouts and privacy-related removals in 5 of the 30 passenger recordings. These discontinuities can be identified via incrementing frame IDs, which preserves suitability for video-based tasks.
Annotations
- Two annotators manually labeled the data using a local instance of the Computer Vision Annotation Tool, with review by a third annotator. The annotations follow the MS COCO format and use human-aligned rotated bounding boxes. Each box is assigned one of three classes (person, clothing, bag) and is linked to a consistent tracking identifier. Person instances additionally include an action attribute (seated, seated on the ground, standing, lying).
Dataset Statistics
- The PMOF dataset comprises 19,696 frames across 31 recordings, including background-only frames (rec0) and 30 passenger recordings. These passenger recordings contribute 19,345 annotated frames with 44,718 person instances from 67 distinct participants. Recordings range from 294 to 1,061 frames and contain between one and four passengers, with an average of 2.3 passengers per recording. Most passengers remain seated throughout the ride.
Citation
BibTeX:
@article{wermuth_pmof_2026,
title = {PMOF: A Dataset and Benchmark for Passenger Monitoring Using Overhead Fisheye Cameras},
author = {Wermuth, Stella Katharina and Ahmed, Qazi Arbab and Neumann, Klaus and Jungeblut, Thorsten},
journal = {Proc. 2026 IEEE International Conference on Advanced Visual and Signal-Based Systems (AVSS)},
year = {2026},
month = {Sep.},
note = {accepted, to appear}
}