marcossenecta/thermal-fall-llm-screening

Zero-shot fall screening of thermal video with a multimodal LLM: laboratory data

Data release for Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home (Marcos Junca and Mauro Junca, Senecta Inc., 2026).

All of it concerns the public eHomeSeniors dataset (Riquelme et al., Sensors 2019, doi:10.3390/s19204565, CC BY 4.0): its 90 Melexis MLX90640 recordings (32x24 thermopile array). The dataset labels each recording with a fall type only; we located every fall.

path	what
annotations/ehome_fall_annotations.jsonl	one line per recording: key (e.g. G1-1-f01), falls (landing times in seconds from the recording's first frame), and notes; exclude marks the one recording left out
windows/ehome_windows.jsonl	the 1001 windows the model reviewed: 438 fall windows (3 s before the landing to 6 s after) and 563 non-fall windows (kind: upright, recovery, rest); partition is the frozen subject split (G*-3 held out)
prompts/requests.json	every prompt, answer schema and generation setting, per condition
answers/<condition>.jsonl	every answer the model gave on these windows, per condition: window, model, verdict, ends_on, reason (verbatim), latency
results/lab_results.json	the laboratory numbers reported in the paper

How the annotations were made. An automatic detector proposed collapse times from the height of the largest warm region; each recording was then inspected on strips of frames sampled every 0.5 s and each repetition's landing (the moment the body reaches floor level) recorded, by one annotator (AI-assisted), before any model saw any window. Bed falls (fall type 05) cannot be told from lying down on the bed at 32x24 and are reported apart.

Rebuilding the windows. Download the eHomeSeniors archive, read each Melexis CSV, and cut each window from start_s to end_s on the recording's own clock (seconds from its first frame). The model saw each window rendered as described in the paper (inferno colour map, per-clip 2nd-98th percentile scale, nearest-neighbour upscale to 480x360, H.264 at the frames' own timing, no audio track).

What is not here. The paper's second dataset, eight days of thermal video and sound from an occupied room, is not released: it contains the voices and conversations of the people who live there.

Licence: the annotations, windows, answers and results are released under CC BY 4.0, like the dataset they describe. Cite the paper and the eHomeSeniors dataset.

Paper

Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home. Marcos Junca and Mauro Junca, Senecta Inc., 2026. Summary and abstract · PDF

Citation

The same files are archived on Zenodo with a DOI: 10.5281/zenodo.23541076. Cite the data:

@dataset{junca2026data,
  author    = {Junca, Marcos and Junca, Mauro},
  title     = {Thermal fall annotations and multimodal {LLM} answers for the {eHomeSeniors} dataset},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23541076}
}

and the paper:

@techreport{junca2026thermal,
  title       = {Asking a Multimodal {LLM} Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home},
  author      = {Junca, Marcos and Junca, Mauro},
  institution = {Senecta Inc.},
  year        = {2026},
  month       = oct,
  url         = {https://senecta.ai/research/thermal-fall-detection-llm/}
}

Also cite the dataset the windows come from: Riquelme, F., Espinoza, C., Rodenas, T., Minonzio, J.-G., Taramasco, C. eHomeSeniors Dataset: An Infrared Thermal Sensor Dataset for Automatic Fall Detection Research. Sensors 19(20), 4565 (2019). doi:10.3390/s19204565

About

Released by Senecta, which makes in-home monitoring for home care agencies.