vladyslava-rudas/jrdb-proxemic-risk

JRDB Danger-Level Dataset

Dataset Description

This dataset was created based on the JRDB (JackRabbot Dataset) and contains images annotated according to pedestrian danger levels. Each image is accompanied by metadata describing the video frame id, the sequence id, the danger level, the bounding box of the nearest person, and the annotation method. The dataset is divided into three groups at the sequence level:

• Evaluating – used for model evaluation during development and model selection
• Finetuning – used for model fine-tuning/training
• Validating – used for final validation of model performance

The dataset contains four danger-level classes:

• high
• moderate
• low
• minimum The dataset is intended for research on visual danger-level prediction and pedestrian-aware scene understanding.

---

Annotation Procedure

Images were annotated with a danger level label based on the estimated risk associated with the nearest pedestrian in the frame. For each annotated image, the dataset contains the following information:

• image — image/frame
• frame_id — identifier of the source frame
• sequence_id — identifier of the source sequence
• danger_level — one of four danger level classes
• closest_person_bbox — bounding box of the nearest pedestrian, represented as [x_min, y_min, width, height]
• labeling_technique — the technique or source used to obtain the annotation; two options: manual or from_jrdb The closest_person_bbox parameter identifies the pedestrian considered most relevant for the danger level annotation.

---

Dataset Statistics

Samples per split

Split	Number of samples
Evaluating	952
Finetuning	200
Validating	291
Total	1 443

Samples per class

Danger level	Evaluating	Finetuning	Validating	Total
high	166	50	57	273
moderate	311	50	95	456
low	230	50	69	349
minimum	245	50	70	365
Total	952	200	291	1 443

---

Sequence-Level Split Policy

To prevent data leakage between splits, the dataset is divided at the sequence level. All frames belonging to the same sequence_id are assigned to exactly one split.

---

Limitations and Subjectivity

It is worth noting that the classification of proxemic risk is subjective, as the perception of danger is not universal and depends on many factors, including the annotator’s interpretation of a pedestrian’s proximity and direction of movement. Therefore, these labels should not be regarded as objective or universally applicable indicators, but rather as qualitative assessments of proxemic risk.

---

Source Dataset: JRDB

This dataset is derived from the JackRabbot Dataset (JRDB), a large-scale multimodal dataset collected for research in robotics and autonomous navigation.

The original JRDB dataset should be specified as the source of the source images and sequences. Please cite the original JRDB work when using this dataset.

JRDB Attribution

This dataset does not claim ownership of the original JRDB images. It is a derived dataset containing additional annotations regarding danger levels and metadata. Users should review and comply with the terms and conditions regarding attribution of the original JRDB dataset, in addition to the licence specified for this derived dataset.

Citation

@article{Mart_n_Mart_n_2023,
   title={JRDB: A Dataset and Benchmark of Egocentric Robot Visual Perception of Humans in Built Environments},
   volume={45},
   ISSN={1939-3539},
   url={http://dx.doi.org/10.1109/TPAMI.2021.3070543},
   DOI={10.1109/tpami.2021.3070543},
   number={6},
   journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
   publisher={Institute of Electrical and Electronics Engineers (IEEE)},
   author={Martín-Martín, Roberto and Patel, Mihir and Rezatofighi, Hamid and Shenoi, Abhijeet and Gwak, JunYoung and Frankel, Eric and Sadeghian, Amir and Savarese, Silvio},
   year={2023},
   month=June, pages={6748–6765} }

@misc{rudas2026visionlanguagemodelsassessproxemic,
      title={Can Vision-Language Models Assess Proxemic Risk from Egocentric Robot Images?}, 
      author={Vladyslava Rudas and Dmytro Kuzmenko},
      year={2026},
      eprint={2608.12515},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.12515}, 
}

---

License

This derived dataset is distributed under the:

Creative Commons Attribution-NonCommercial-ShareAlike 3.0 (CC BY-NC-SA 3.0) license.

This license permits sharing and adaptation subject to the conditions of attribution, non-commercial use, and distribution of adaptations under the same license. The licensing terms of the original JRDB dataset and its source images may impose additional requirements. Users are responsible for complying with all applicable terms.

---

Intended Use

This dataset is intended for:

• academic and research use
• computer vision research
• pedestrian-aware perception
• danger-level classification
• robotics and autonomous navigation research
• model fine-tuning and evaluation

It is not intended for commercial use under the stated license.

---

Acknowledgements

We would like to thank the authors of the JRDB project for making the raw data available to the scientific community.