Professional Documents
Culture Documents
Matched Filter
Matched Filter
Matched Filter
Dr Usman Zabit
Usman Zabit
but as discussed in Chapter 3, the resulting test
statistic will be identical to that obtained using the
Bayesian risk criterion.
The detection problem is to distinguish between the
hypotheses
2
4.3 MATCHED FILTERS
where the signal 𝑠[𝑛] is assumed known and 𝑤[𝑛] is
WGN with variance 𝜎 2 .
WGN is defined as a zero mean Gaussian noise
process with autocorrelation function (ACF)
Usman Zabit
3
4.3 MATCHED FILTERS
Usman Zabit
where 𝛿[𝑘] is the discrete-time delta function.
The NP detector decides 𝐻1 if the likelihood ratio
exceeds a threshold
4
4.3 MATCHED FILTERS
Usman Zabit
5
4.3 MATCHED FILTERS
Taking the logarithm (a monotonically increasing
transformation) of both sides does not change the
inequality (see Section 3.3) so that
Usman Zabit
6
4.3 MATCHED FILTERS
Since 𝑠[𝑛] is known (and thus not a function of the
data), we may incorporate the energy term into the
threshold to yield
Usman Zabit
7
4.3 MATCHED FILTERS
Calling the right-hand-side of the inequality a new
threshold 𝛾′, we decide 𝐻1 if
Usman Zabit
This is our NP detector and as expected consists of a
test statistic 𝑇(𝒙) (a function of the data) and a
threshold 𝛾′ , which is chosen to satisfy a given
probability of false alarm 𝑃𝐹𝐴 = 𝛼.
8
4.3 MATCHED FILTERS
In general, the test statistic of (4.3) weights the data
samples according to the value of the signal.
More weight is reserved for the larger signal samples.
Usman Zabit
same manner because by multiplying
𝑥[𝑛] by 𝑠[𝑛], negative samples yield a positive
contribution to the sum.
The detector of (4.3) is referred to as a
correlator or replica-correlator since we correlate
the received data with a replica of the signal.
The detector is shown in Figure 4.1a.
9
REPLICA-CORRELATOR
Usman Zabit
10
4.3 MATCHED FILTERS
An alternative interpretation of the test statistic is
based on relating the correlation process to the effect
of a finite impulse response (FIR) filter on the data.
Specifically, if 𝑥[𝑛] is the input to an FIR filter
Usman Zabit
with impulse response ℎ[𝑛], where ℎ[𝑛] is nonzero for
𝑛 = 0,1, … , 𝑁 − 1 , then the output at time n is
11
4.3 MATCHED FILTERS
If we let the impulse response be a “flipped around”
version of the signal or
Usman Zabit
12
4.3 MATCHED FILTERS
which with a change of summation variable is
identical to the replica-correlator.
This implementation of the NP detector is shown and
is known as a matched filter.
Usman Zabit
13
4.3 MATCHED FILTERS
The filter impulse response is matched to the signal.
Figure 4.2 shows some examples.
14
4.3 MATCHED FILTERS
Usman Zabit
signal output is given via the convolution sum in
Figure 4.3.
Note that the signal output attains a maximum at the
sampling time.
This is true in general (see Problem 4.2).
15
MATCHED FILTER OUTPUT
Usman Zabit
16
4.3 MATCHED FILTERS
Usman Zabit
assume that it does and use the corresponding matched
filter, poor detection performance may be obtained.
An example is given in Problem 4.3. Thus, for signals
with unknown arrival times, we cannot use the
matched filter in its present form.
In Chapter 7 we will see how to modify it to circumvent
this problem.
17
ASPECTS OF MATCHED FILTERS
1. Another property of the matched filter is that it maximizes
the SNR at the output of an FIR filter.
2. The use of a matched filter for signal detection leads to the
concept of processing gain.
Usman Zabit
Processing gain can be viewed as the advantage of making a
decision based on a test statistic, which is an optimal
combination of the data, as opposed to a decision based on a
direct examination of the data.
It is defined as the SNR of the optimal test statistic, i.e, the
processed data, divided by the SNR of a single data sample,
i.e., the unprocessed data.
As an example, consider detection of a DC level in WGN.
By using a matched filter to process 𝑁 samples the
performance improves to 18
𝑃𝐺 = 10 log10 𝑁 𝑑𝐵.