Anda di halaman 1dari 5

Search & Detection Theory (OA3602) Lecture Notes Prof.

James Eagle
Text: Search & Detection, 4th Ed., Alan R. Washburn Rev: November 5, 2012

6 Receiver Operating Characteristic (ROC) Curves

6.1 Introduction

In a detection situation, one of four events must occur. Either the target is present or not;
and in each case, either the detector calls “detect” or not. These four cases are associated
with the following four conditional probabilities.
Tgt. Tgt.
present absent Pd = P (“detect” called | target is present)
Detector calls Pd Pf = probability of detection
“detect” Pf = P (“detect” called | target is absent)
Detector does not 1 − Pd 1 − Pf = probability of false alarm
call “detect”
For many sensors there is a user-selectable tradeoff between Pd and Pf . That is, by
increasing the sensitivity of the detector, Pd can be increased towards 1, if an accompanying
increase in Pf is also acceptable. This would produce an “aggressive” sensor, which gives
more true detections, but at the cost of more false detections.
By decreasing detector sensitivity, on the other hand, Pf can be decreased towards 0, if
an accompanying decrease in Pd is also acceptable. This would be a “conservative” sensor,
which gives fewer false detections but also fewer true detections.

Definition: A ROC curve is a graph of


1
(Pf , Pd ) pairs as detector sensitivity is varied
from low (resulting in low Pf and Pd ) to high
better (resulting in high Pf and Pd ).
Pd sensors

For a specified Pf , the best sensor has the


largest Pd .

The sensor with Pd = Pf has no ability to


distinguish between “target” and “no-target”
0 Pf 1 situations.

1
6.2 Derivation of ROC Curve for One Look for a Known Deterministic
Signal in Additive Gaussian Noise

Assumptions:
• When no target is present, the detector output voltage (V ) is normally distributed
noise (N ) with mean µ and variance σ 2 . That is, V = N ∼ N (µ, σ 2 ).
• With a target is present, the detector output voltage is normal noise plus a deter-
ministic signal (s). Now V = N + s ∼ N (µ + s, σ 2 ).
So, if the target is not present, we hear only Gaussian noise. If the target is present, we
hear the same Gaussian noise plus a deterministic signal.
Target present? Detector output voltage (V )
No V = N ∼ N (µ, σ 2 )
Yes V = N + s ∼ N (µ + s, σ 2 )
Now let ν be the detection threshold. If V ≥ ν, “detection” is called; and if V < ν,
“detection” is not called.

Pf Pd

No target present Target present

Figure 1: Target Not Present and Target Present

With ν, µ, and σ specified, and assuming the the target is not present, we can compute
the probability of a false alarm.

Pf = P (“detect” called | no target present)


= P (V ≥ ν | V ∼ N (µ, σ 2 ))
 
ν−µ
= 1−Φ
σ
 
µ−ν
= Φ (1)
σ

2
And assuming that the target is present, we can compute probability of detection.

Pd = P (“detect” called | target present)


= P (V ≥ ν | V ∼ N (µ + s, σ 2 ))
 
ν−µ−s
= 1−Φ
σ
 
µ−ν+s
= Φ (2)
σ

So with µ, σ, and s specified and fixed, we can generate the ROC curve by allowing ν to
vary.     
µ−ν+s µ−ν
(Pd (ν), Pf (ν)) = Φ ,Φ (3)
σ σ

6.3 Derivation of ROC Curve for n Looks for a Known Deterministic


Signal in Additive Gaussian Noise

Now assume that n independent voltage measurements are made, and the sample mean
voltage is calculated. If this mean voltage meets or exceeds the detection threshold ν, the
sensor calls a detection. Now what are Pf and Pd ?
We know that the variance of the mean of an iid sample of size n is σ 2 /n. So if no target
is present,
n
X
V = 1/n Ni , where Ni ∼ N (µ, σ 2 )
i=1
2
∼ N (µ, σ /n).

So,    
ν−µ µ−ν
Pf = P (V ≥ ν) = 1 − Φ √ =Φ √ (4)
σ/ n σ/ n

And, if a target is present,


n
X
V = 1/n (Ni + s), where Ni ∼ N (µ, σ 2 )
i=1
∼ N (µ + s, σ 2 /n).

So,    
ν−µ−s µ−ν+s
Pd = P (V ≥ ν) = 1 − Φ √ =Φ √ . (5)
σ/ n σ/ n

3

The effect is to reduce the standard deviation of the noise in (1) and (2) from σ to σ/ n.
This results in decreasing Pf and increasing Pd .
Example: Let (µ, σ, s) = (0, 3, 4), all in units of dB. And let n ∈ {1, 2, 3}. For threshold
levels ν ∈ [−3, 10] db, the ROC curves are given by
    
−ν + 4 −ν
(Pd (ν), Pf (ν)) = Φ √ ,Φ √ ,
3/ n 3/ n

and are plotted in Figure 2. The ROC curves generated using larger n (i.e., more averaging)
are more discriminating (i.e., for the same ν, Pd is larger and Pf smaller).

Figure 2: ROC Curves when Averaging 1, 2, and 3 Values.

6.4 Optimizing the Detection Threshold ν

Definitions:
• c1 = cost of a missed detection
• c2 = cost of a false alarm
• p = P (target is present)
• Cν = random cost of the next look when the detection threshold is ν

4
Then,

E(Cν ) = E(Cν | target is present)P (target is present) +


E(Cν | target is absent)P (target is absent)
= c1 (1 − Pd (ν))p + c2 Pf (ν)(1 − p).

One-Look Example: Let (µ, σ, s, p, c1 , c2 ) = (0 dB, 5 dB, 10 dB, .3, 10, 3). Assuming the
target is absent,    
µ−ν −ν
Pf (ν) = Φ =Φ .
σ 5

And assuming the target is present,


 
µ − ν + 10
Pd (ν) = Φ .
σ

So,
     
−ν + 10 −ν
E(Cν ) = 10 1 − Φ .3 + 3 Φ .7
5 5
   
ν − 10 −ν
= 3Φ + 2.1 Φ .
5 5

To find the minimum-cost detection threshold, this expression can be minimized numeri-
cally or by analytical differentiation. (Ans: ν ∗ = 4.1 volts and E(Cν ∗ )=.79.)

Figure 3: Determining the Minimum-Cost Detection Threshold ν ∗ .

Anda mungkin juga menyukai