The short answer
Fewer pixels, fewer falls caught. On the same staged falls, an AI reviewer caught 69% at the sensor's 32 by 24 pixels, 48% at 16 by 12 and 34% at 8 by 6, the size of the cheapest thermal arrays, mostly because at the coarsest size it read a fallen person as someone leaving the picture. In a lived-in room, the same AI at 8 by 6 caught 0 of the 9 staged falls the camera could see, against 4 at the camera's full 160 by 120. Coarser sensors reveal less about a person; that privacy is paid for in missed falls.
Why resolution is a privacy question
Low-resolution thermal sensing is how many fall detection products avoid cameras: a sensor that sees heat in a grid of a few dozen pixels cannot show a face, a screen or a letter on the table. The coarser the grid, the stronger that promise. The question this finding answers is what the promise costs in falls caught.
What we did
We took the staged falls of a public laboratory dataset, recorded by a 32 by 24 pixel thermal sensor, and averaged blocks of pixels to make the same clips at 16 by 12 and at 8 by 6, keeping the field of view and the size of the picture the AI saw. The same instructions went to two Google models, Gemini 3.8 Flash and 3.7 Flash, with no training on falls. We did the same with clips from our research prototype's 160 by 120 thermal camera in a lived-in room, reduced to 32 by 24 and to 8 by 6.
What we found
| Resolution | Falls caught, Gemini 3.8 Flash | Falls caught, Gemini 3.7 Flash | Non-falls called a fall (3.8 / 3.7) |
|---|---|---|---|
| 32 by 24 (the sensor) | 69% | 80% | 0 / 0 of 182 |
| 16 by 12 | 48% | 61% | 0 / 0 of 182 |
| 8 by 6 | 34% | 51% | 2 / 5 of 182 |
Laboratory falls of the two held-out subjects, 140 falls and 182 non-fall clips.
In the lived-in room, Gemini 3.7 Flash caught 6 of the 9 visible staged falls at 160 by 120, 7 at 32 by 24 and 3 at 8 by 6, with 0, 1 and 0 false calls on 73 sampled clips of ordinary living. The deployed 3.8 Flash at 8 by 6 caught 0 of 9.
Why a coarse image fools the AI
At 8 by 6 pixels a person lying near the bottom of the picture is a few warm blocks at its edge, and that looks much like someone walking out of view. The AI's instructions say a person who goes out of view has not fallen, so it declined more falls on those grounds as the image got coarser: 13 of 140 at 32 by 24, 24 at 16 by 12 and 49 at 8 by 6. Fewer pixels did not make it more trigger-happy: only 2 of the 182 non-fall clips were called a fall at 8 by 6. It made it blind.
What it means if you are choosing a sensor
- Ask the resolution. “Thermal” covers everything from an 8 by 8 array to a 160 by 120 camera, and the difference is large.
- Ask for detection measured at that resolution, in homes, not a figure from a higher-resolution test.
- Privacy is more than pixels. What is kept, for how long and who can see it matters as much as how coarse the image is; see what Senecta keeps.
Questions this answers
Is an 8 by 8 thermal sensor enough for fall detection?
In our tests an AI reviewer working from 8 by 6 pixel thermal images caught 34% of staged laboratory falls, against 69% at 32 by 24. Arrays that coarse lose most of the evidence a fall leaves.
Does a thermal camera show faces?
At the resolutions in this study, 32 by 24 and 160 by 120 pixels, thermal video shows a warm body and its posture, not a face. This paper does not measure privacy; our next study compares thermal and normal video of the same moments.
Does lower resolution cause more false alarms?
Not in our tests: at 8 by 6, 2 of 182 laboratory non-fall clips were called a fall. The cost of coarse images was missed falls.
The paper
This finding comes from Can an AI tell from thermal video whether someone fell?, which describes the methods, the data and every limitation we know of.
Honest scope
- The falls were staged, by younger adults, in one laboratory and one room. Real falls of older people are different.
- The room recordings come from a research prototype, not the product's final sensor, mounting or software.
- These are research results, not a promise of how often a product will catch a fall. No system catches every fall.
Senecta makes the pods this research is for, so this is our own study, not an independent evaluation.