Detector Fire Rate (DFR): Concepts & Applications

• Detector Fire Rate (DFR) is a term with three distinct meanings—segment-level recall in video fire detection, count rate in radiation detectors, and decreasing failure rate in renewal theory—requiring domain-specific qualification.
• In video fire detection, DFR is defined as the segment-level recall (TP/(TP+FN)), with reported values around 82% over 200-frame sections, emphasizing trade-offs between false positives and negatives.
• In radiation detection and reliability theory, DFR corresponds respectively to the expected count rate computed via response matrices and to a property of the inter-renewal time distribution, highlighting the importance of clear contextual usage.

Searching arXiv for the cited papers and topic usage of “DFR”. Detector Fire Rate (DFR) is not a single standardized technical term. In the literature represented by "Towards a solid solution of real-time fire and flame detection" (Jiang et al., 2015), "Detector Response Matrices, Effective Areas, and Flash-Effective Areas for Radiation Detectors" (Bowers et al., 31 Dec 2025), and "Concave Renewal Functions Do Not Imply DFR Inter-Renewal Times" (Yu, 2010), the label maps to three distinct notions: segment-level fire-event recall in video fire detection, detector count rate derived from a Detector Response Function or Detector Response Matrix, and decreasing failure rate for inter-renewal times in renewal and reliability theory. This suggests that any use of DFR requires explicit domain qualification before the quantity can be interpreted.

1. Terminological scope

The three uses of DFR in the cited works are not interchangeable. One source explicitly states, “There is no metric named ‘Detector Fire Rate (DFR)’ in the paper,” and then formalizes the closest quantity as segment-level recall for a real-time video fire detector. A second source treats DFR as count rate, namely the expected number of detector counts per unit time obtained by folding an incident spectrum through a counting DRF or DRM. A third source uses DFR in its standard reliability-theoretic meaning, decreasing failure rate, defined through a nonincreasing hazard rate (Jiang et al., 2015, Bowers et al., 31 Dec 2025, Yu, 2010).

Domain	DFR meaning	Canonical expression
Video fire detection	Segment-level recall	$\dfrac{TP}{TP+FN}$
Radiation detectors	Count rate	$$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$
Renewal theory	Decreasing failure rate	$r(t)$ nonincreasing

This terminological split is central. In video surveillance, DFR refers to whether fire-containing video segments are successfully flagged. In radiation detection, it refers to how often the detector “fires,” meaning registers counts. In renewal theory, it refers not to a detector output rate but to a shape property of the inter-renewal distribution.

2. Segment-level DFR in real-time video fire detection

In the real-time video-based fire detection setting, the measured quantities are standard detection metrics. The benchmark reports True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN) at the level of video sections rather than frames or pixels. From Table 3 of the benchmark, the reported values are $TP=361$, $TN=305$, $FP=27$, and $FN=81$, over 744 sections of 200 frames each. Precision and recall are defined as

$$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$

giving

$$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$

The abstract and summary round these to 82% Recall and 93% Precision (Jiang et al., 2015).

Within this evaluation protocol, the most natural formalization of DFR is the event-level true positive rate computed at the segment level:

$\text{DFR} \equiv \text{Recall}=\frac{TP}{TP+FN}.$

Under the reported experiment, this yields $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$0, rounded to $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$1. The quantity therefore denotes the probability that a randomly chosen 200-frame segment that contains fire will be labeled as “fire” by the detector.

The evaluation unit is explicitly section-level evaluation. Each video is split into segments of 200 frames; each segment is labeled either “contains fire” or “does not contain fire”; and the detector outputs a binary decision per section. Because both the proposed method and the Toreyin & Cetin method make decisions over multiple frames, the paper does not evaluate per frame or per pixel. The reported DFR is therefore an event-level / segment-level detection rate rather than a frame-by-frame measure. The paper also gives derived quantities such as false positive rate and true negative rate:

$$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$2

$$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$3

3. Cascade architecture and determinants of video DFR

The real-time fire detector is a three-stage cascade: candidate region proposal, region classification, and temporal verification. Each stage affects $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$4, $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$5, and $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$6, and hence the effective DFR (Jiang et al., 2015).

The first stage is candidate region proposal based on background modeling and adaptive thresholding. Fire regions tend to be bright, and the system therefore uses a multi-level, adaptive threshold rather than fixed thresholds. In the static-background regime, it applies background subtraction and then thresholding to obtain candidate blobs. In the moving-background regime, it relies solely on adaptive intensity thresholding. The paper states that “Higher background intensity leads to lower threshold, and vice versa.” Too strict a threshold increases FN and lowers DFR; too loose a threshold increases the burden on later stages and can raise FP.

The second stage is region classification using color-texture features, a dictionary of visual words, and SVM. Candidate blobs are represented by dense local descriptors built from SURF (64-dim) and a local LAB color histogram with 24 bins, giving an 88-dimensional local descriptor. Training descriptors are clustered by k-means into 500 cluster centers. For each descriptor $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$7, the method finds its $$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$8 nearest codewords and computes Gaussian-kernel weights

$$R^{(k)}=\sum_j \mathbf{G}_\varphi^{(j,k)}\Phi^{(j)}$$9

with

$r(t)$0

and accumulates a 500-bin histogram

$r(t)$1

A 96-bin global LAB histogram is then concatenated with the 500-dimensional local BoW histogram to form a 596-dim vector per blob. Classification is performed with LIBSVM and an RBF kernel, which the experiments report as better than linear and $r(t)$2 kernels. Dense sampling is emphasized as more robust than keypoint sampling, especially on smooth surfaces; Table 1 reports much higher precision for LAB+SURF with dense sampling, 92.26% vs 84.85% on patch tests. The data states that strong discriminative power of LAB+SURF+RBF improves the TP rate and prevents many FP on skin, yellow clothing, and bright objects.

The third stage is temporal verification over 25 consecutive frames for each blob. The method computes perimeter sequence statistics $r(t)$3, area sequence statistics $r(t)$4, and a spatial distribution statistic

$r(t)$5

where the bounding box is divided into four sub-rectangles and fire-pixel counts are tracked over time. A blob is stable if

$r(t)$6

and unstable if

$r(t)$7

This stage is mainly a false alarm filter. It reduces FP while trying to preserve TP, but aggressive thresholds can also increase FN and lower recall. The paper therefore presents an explicit trade-off: temporal verification can be tuned either toward higher DFR at the cost of more false positives, or toward lower false alarm rate with some recall sacrifice.

The benchmark itself contains 64 video clips, with average length about 2 minutes, covering indoor, urban outdoor, forest, static, and moving backgrounds, and including fire-like distractors such as car lights, people in fire-colored clothes, and illumination changes. Within that setting, the reported DFR is the operating point of a cascade designed for real-time surveillance and event retrieval.

4. DFR as detector count rate in response-matrix formalism

In the radiation-detector setting, DFR is naturally interpreted as the count rate. The formal object is the counting Detector Response Function

$r(t)$8

which relates an incident particle characterization $r(t)$9 to counts as a function of deposited energy. In continuous form,

$TP=361$0

After discretization into incident-energy bins $TP=361$1 and deposited-energy bins $TP=361$2, the Detector Response Matrix $TP=361$3 satisfies

$TP=361$4

If $TP=361$5 is incident flux integrated over bin $TP=361$6 and $TP=361$7 is observation time, the expected counts in channel $TP=361$8 are

$TP=361$9

so the Detector Fire Rate in channel $TN=305$0 is

$TN=305$1

Summing over all channels gives the total Detector Fire Rate,

$TN=305$2

Equivalently, with channel-integrated counting effective area,

$TN=305$3

This is the conventional count-rate formula and is directly interpretable as DFR (Bowers et al., 31 Dec 2025).

The paper also defines the counting effective area

$TN=305$4

which collapses the 2D DRF to a 1D function of incident energy for a chosen deposited-energy interval. Folding this quantity with flux yields the fire rate in a deposit-energy window:

$TN=305$5

A distinct quantity is the flash-effective area

$TN=305$6

which is for total energy deposition from an instantaneous fluence:

$TN=305$7

The data explicitly distinguishes this from event counting. The counting-effective area yields count rate; the flash-effective area yields total deposited energy and is therefore not directly a DFR.

The formalism is linear in the incident spectrum, and the DRF is described as not including pile-up or dead-time effects in the response itself. The data states that high-count-rate effects such as paralyzable or non-paralyzable dead time must be applied afterward as corrections to the DFR if needed:

$TN=305$8

for a non-paralyzable detector, and

$TN=305$9

for a paralyzable detector. This makes the distinction between ideal count-rate DFR and observed count-rate DFR explicit.

5. DFR as decreasing failure rate in renewal theory

In renewal theory, DFR means decreasing failure rate, not detector count rate. For a nonnegative random variable with density $FP=27$0 and survival function $FP=27$1, the hazard rate is

$FP=27$2

and a distribution is DFR if $FP=27$3 is nonincreasing on $FP=27$4. The paper also states the equivalent formulation: a distribution on $FP=27$5 has DFR if its survival function $FP=27$6 is log-convex on $FP=27$7 (Yu, 2010).

The renewal process with i.i.d. inter-renewal times has renewal function

$FP=27$8

which satisfies

$FP=27$9

and, when $FN=81$0 is absolutely continuous, renewal density

$FN=81$1

Brown’s classical implication, restated as Theorem 1, is that if the inter-renewal distribution is DFR, then the renewal function $FN=81$2 is concave. Concavity means the incremental rate of renewals $FN=81$3 is nonincreasing.

The converse, however, is false. Brown conjectured that concavity of the renewal function would imply DFR inter-renewal times, and Yu’s paper refutes this conjecture. The counterexamples are constructed through a piecewise hazard rate on intervals $FN=81$4, $FN=81$5, $FN=81$6, and $FN=81$7 such that the hazard decreases on $FN=81$8, increases on $FN=81$9, then decreases on $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$0 and remains nonincreasing on $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$1. For sufficiently small $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$2 and sufficiently small interval length $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$3, the renewal density $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$4 remains decreasing on $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$5, so the renewal function is concave even though the underlying distribution is not DFR.

An explicit part of the construction uses

$$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$6

followed on $$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$7 by

$$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$8

which produces a hazard of the form

$$\text{Precision}=\frac{TP}{TP+FP}, \qquad \text{Recall}=\frac{TP}{TP+FN},$$9

with $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$0, and this hazard is strictly increasing on $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$1. The paper also gives a numerical illustration with $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$2, $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$3, $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$4, and $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$5.

The same paper also gives a short proof of Shanthikumar’s result that the DFR property is closed under geometric compounding in the discrete setting. Thus, in renewal or reliability language, DFR is a property of the inter-renewal distribution rather than a directly observed event count.

6. Interpretive distinctions and recurrent misconceptions

Several recurrent confusions follow directly from the coexistence of these meanings. First, DFR is not a universal metric. In the video-detection setting, the closest formal quantity to DFR is segment-level recall over 200-frame sections, not a frame-level or pixel-level score. In the radiation-detector setting, DFR is count rate, either channel-specific or total, computed from a counting DRF, DRM, or counting-effective area. In renewal theory, DFR is a hazard-shape property of inter-renewal times, not a rate computed from detector output (Jiang et al., 2015, Bowers et al., 31 Dec 2025, Yu, 2010).

Second, the video-fire interpretation is tied to a specific decision protocol. The paper does not define an explicit “fire must be detected within $$\text{Precision}=\frac{361}{361+27}\approx 93.04\%, \qquad \text{Recall}=\frac{361}{361+81}\approx 81.67\%.$$6 frames after ignition” rule, and its final evaluation is not per frame or per region. A segment labeled “contains fire” is counted as detected when the detector outputs “fire” within the 200-frame block under the system’s section-level decision mechanism.

Third, the response-matrix interpretation distinguishes sharply between event counting and energy deposition. The counting-effective area produces a count-rate DFR; the flash-effective area produces total deposited energy for a brief, high-fluence event. Treating flash-effective area as a count-rate quantity would conflate two different response regimes.

Fourth, in reliability-theoretic usage, concavity of the renewal function does not imply DFR inter-renewal times. Yu’s counterexamples show that a renewal density can be decreasing even when the underlying hazard increases over some interval. A plausible implication is that empirical observation of a decreasing expected firing rate is insufficient, by itself, to certify decreasing failure rate of the underlying mechanism.

Across these domains, the term remains meaningful only when the measured object is specified: probability of correctly flagging fire-containing segments, expected number of counts per unit time, or nonincreasing hazard rate. Without that specification, “Detector Fire Rate” is formally ambiguous.