Research · 2 min read
Thermal Cameras Do Not Win Everywhere. They Win When the Light Disappears.
The practical decision is where thermal should supplement visible coverage, not whether it should replace visible cameras everywhere.
Thermal beats visible imaging — that's the pitch, anyway. In 720 paired trials, the pitch holds up exactly where you'd hope: when the light disappears. In daylight the two are nearly tied. Near darkness is a different story, and it has a price.
Our analysis of 720 paired positive trials; 48 negative hours per modality, 720 paired positive trials plus 48 negative hours per modality.
Is thermal always better than visible imaging?
No. Across 720 paired positive trials and 48 negative hours per modality, thermal's sensitivity advantage was 0.6 points in daylight and 24.4 points near darkness, while its observed false-alert rate was 1.76 times the visible rate. The decision is condition-specific.
The numbers behind the answer
Selected measures only. Denominators and interpretation stay attached so the headline cannot stand alone.
+24.4 pts
Near-dark sensitivity advantage
Thermal sensitivity was 75.0% versus 50.6% for visible imaging in the near-dark band.
+0.6 pts
Daylight sensitivity advantage
Thermal and visible imaging were nearly tied in daylight: 80.0% versus 79.4%.
1.76×
Thermal/visible false-alert rate ratio
Seven versus four alerts over 48 negative hours per modality; the small counts make the estimate imprecise.
Thermal's observed advantage grows as visible light falls
Line the results up by light level and the pattern is clean. In daylight, thermal and visible are nearly tied — 80.0% of positive trials detected versus 79.4%, a 0.6-point gap. Twilight widens it to 8.3 points. Low light: 13.4. Near darkness: 24.4 points, 75.0% against 50.6%.
So the honest headline isn't "thermal wins." It's "thermal wins when visible can't see." That argues for supplementing coverage where light fails, not ripping out visible cameras everywhere. And no — thermal still doesn't see through walls.
Deployment question
Ask where light loss creates a meaningful detection gap—not which modality wins in the abstract.
Daylight is nearly tied; near darkness separates the modalities
The overall trial count hides the point if you read it alone. Across 720 paired positive trials per modality, thermal detected 574 and visible 490 — 84 fewer misses for thermal. Respectable, but that figure blends a near-tie in daylight with a blowout in the dark.
Which is why the buying question is geographic, not technological. Your perimeter, your loading area, and your operating hours have a light profile. Match the tested condition to the actual spot, or the averages will lie to you.
Evidence visual
Detection sensitivity by light condition
Every value is a share of positive trials detected within that modality and light band. The comparison is paired at the trial level.
Faster detected events came with a higher observed false-alert rate
Speed comes first. Among detected events, median time-to-alert was 1,498 milliseconds for thermal versus 1,751 for visible — thermal was faster. But latency only counts events that were detected at all, so read it beside the miss counts, never alone. For the record: the 90th-percentile values were 2,574 and 2,928 milliseconds.
Then the cost. Over 48 negative hours per modality, thermal produced 7 false alerts to visible's 4 — a 1.76× rate. Seven and four are small numbers with wide intervals; treat that ratio as a warning, not a verdict. Detection gain, latency, and review burden belong on the same scorecard, or you'll buy the wrong thing.
Pilot the difficult location before choosing system-wide replacement
The defensible purchase is a pilot at the difficult location, not a system-wide swap. Define the target, the light band, the distance, the obstructions, the acceptable latency, and the false-alert burden you can live with. Then measure both modalities there, paired, and write down the hardware and detector versions so the test can be repeated.
What this aggregate can't tell you: whether thermal-plus-visible fusion beats either alone — it wasn't tested — and whether these exact results transfer to different hardware or sites. Where light loss is real, a selective thermal supplement has evidence behind it. Everywhere else is still a guess.
- Match the pilot to the real light and distance.
- Use paired positive events and negative observation hours.
- Report misses and false alerts together.
- Keep latency conditional-on-detection clear.
- Record hardware and detector configuration.
Questions property teams ask
Can thermal cameras see through walls or opaque objects?
No. Thermal imaging detects emitted heat patterns; ordinary walls and opaque objects still block the view.
Is thermal always better than visible imaging?
No. The daylight sensitivity difference here is only 0.6 points, while the much larger difference appears in near-dark conditions.
Why is sensitivity different from accuracy?
Sensitivity measures how many positive trials were detected. Accuracy would also require a defined negative-trial set and a combined treatment of correct positives and negatives.
Why is latency reported only among detections?
A missed event has no detection timestamp. That means latency comparisons describe detected subsets and must be read beside miss counts.
Does this justify replacing visible cameras?
No. It supports testing thermal as a supplement in matching low-light conditions while measuring false alerts and review workload.
Our methods, limits, and sources
How we calculated this
This is original 911 Sentinel research — we gathered the records, ran every calculation below, and published the aggregate dataset.
We ran paired thermal and visible positive trials by light band and separately counted false alerts during equal negative-observation hours.
- We validated one thermal and one visible observation for each paired positive trial.
- We calculated detected, missed, sensitivity, and latency summaries by modality and light band.
- We built the paired outcome table: both, thermal-only, visible-only, and neither.
- We calculated false-alert rates and approximate intervals from negative-observation hours.
What this analysis cannot establish
- Generalization depends on the exact hardware, configuration, detector, environment, and annotation protocol.
- Latency is conditional on detection, so missed events do not enter the latency samples.
- Only 48 negative hours per modality support the false-alert comparison, making the rate ratio imprecise.
- Condition splits need per-cell sample sizes and paired intervals.
- Partial obstruction requires a reproducible material and obscured-area definition.
- The result does not test thermal-visible fusion.
Sources
This study is 911 Sentinel's own work, built from our operational records.
Related questions and practical guides
New coverage — intake open
Test the difficult view in its real conditions
A camera assessment can identify where daylight performance is already adequate and where a paired low-light pilot is worth measuring.