IET Radar, Sonar & Navigation

Research Article

Comparison of supersonic bullet ballistic ISSN 1751-8784

Received on 2nd March 2016
Revised 29th May 2016
models for accurate localisation of small arms Accepted on 1st June 2016
doi: 10.1049/iet-rsn.2016.0098
fire www.ietdl.org

Kam W. Lo1 , Brian G. Ferguson1

1Maritime Operations Division, Defence Science and Technology Group, NICTA Building, 13 Garden Street, Eveleigh, NSW 2015, Australia
E-mail: [email protected]

Abstract: This study evaluates and compares the accuracies of five different ballistic models in predicting the speed profile of a
supersonic bullet using radar measured speed data for 36 types of bullets. Each of the five ballistic models is parameterised by
a ballistic constant and the muzzle speed of the bullet. In practice, each of these two parameters spans a wide range of values
because of the large number of different bullet types that are available for use in a variety of applications. For a given bullet
type, the bullet's ballistic constant and muzzle speed are determined by fitting the ballistic model to the data in a least-squares
(LS) sense, and the root-mean-square (RMS) and maximum absolute (MA) deviations of the LS fit of the model from the data
are computed. The best ballistic model (out of the five) is the one that has the smallest RMS and MA errors for the maximum
number of bullet types. This ballistic model has been applied to accurate ranging of small arms fire using a single sensor node,
with and without a priori knowledge of the model parameters, and its effectiveness for both cases is demonstrated using real
data recorded from a field experiment.

1Introduction supersonic bullet using Doppler radar measured speed data for 36
types of bullets which are manufactured for a variety of activities
The firing of a supersonic bullet produces two acoustic impulses including sports shooting, hunting, national defence, and law
[1, 2]: the muzzle blast (MB) generated during the launch of the enforcement. The best ballistic model is then used for ranging
bullet from the rifle and the ballistic shock wave (SW) generated small arms fire using a single sensor node, with and without a
by the supersonic flight of the bullet. There has been considerable priori knowledge of the ballistic model parameters.
research in locating the shooter (or point of fire) using both MB
and ballistic SW arrivals at a single sensor node (which consists of
a small array of sensors) [35], a network of sensor nodes [6, 7], or 2Ballistic models
distributed sensors [811]. For accurate localisation results, A supersonic bullet is subject to a gravitational force and a drag
especially for long range shooting, the deceleration of the bullet force as it travels along its trajectory. For flat fire, the gravitational
must be taken into account, and this is accomplished by force is much smaller than the drag force and can be ignored [14].
incorporating an external ballistics model of the bullet into the With this assumption, the trajectory of a supersonic bullet is
localisation algorithm [36, 10]. Experimental results obtained approximately linear. Let v(x) denote the supersonic speed of the
with a single sensor node at long firing ranges have shown that bullet at a range x from the point of fire. The drag force acting on
using an accurate ballistic (or speed) model for the bullet results in the supersonic bullet can be expressed as [12, 13]
much smaller range estimation bias errors than using a constant
speed model [4, 5]. The employment of a ballistic model for the d 1
bullet also enables the point of fire to be located using only the SW D = = D2 (1)
d 2
arrivals at a network of sensor nodes or distributed sensors [5, 12].
There are two different but equivalent approaches to this where m and A are the mass and cross-sectional area of the bullet,
localisation problem. The first approach [5] is to locate directly the respectively, is the air density, and CD is the drag coefficient
point of fire, while the second approach [12] is to first estimate the (dimensionless) . In many practical applications, if the elevation
bullet trajectory, which is then traced back to the point where the
angle of fire is <10, (1) can be used to derive an accurate
speed of the bullet equals its muzzle speed. In all cases, whether
expression for the speed of the bullet [14].
using both MB and SW or only SW, the ballistic model parameters:
The drag coefficient is given by [12, 13]
the muzzle speed of the bullet and its ballistic constant need to be
known a priori; any uncertainty in these parameters will degrade
the localisation accuracy.
D = (2)
Different ballistic models have been used to improve the
accuracy of localisation for small arms fire [35, 10, 12]. The
degree of improvement depends on how accurately the adopted where c is the speed of sound in air, is a factor that depends on
ballistic model predicts the speed profile of the bullet (which is the the actual shape of the bullet, and the exponent 0. Substituting
variation with distance travelled of the bullet's speed). Using (2) into (1) and rearranging the result yields
simulated data, Kozick et al. [13] studied the effect of mismatch
between a given ballistic model and the actual speed profile of the d
= 21
b
2
(3)
bullet on the localisation accuracy for the point of fire. The d
simulated data used in their study were generated using an accurate
ballistic model, assuming that this model predicted exactly the where Cb, referred to as the ballistic constant, is given by
speed profile of the bullet. In practice, Doppler radars are used to
provide accurate measurements of a bullet's speed along its 4
b = . (4)
trajectory. This paper evaluates and compares the accuracies of five
different ballistic models in predicting the speed profile of a
IET Radar Sonar Navig. 1
Commonwealth Crown Copyright 2016
Writing the left hand side of (3) as vdv/dx, then integrating both 3Model evaluation
sides of (3) with respect to x and imposing the initial condition: v=
V0 at x=0 gives an expression for the speed of the bullet as a The accuracies of the five different models in predicting bullet
speed profiles are evaluated using real data for 36 different types of
function of x (i.e. the speed profile) [4, 5]
bullets which are manufactured for a variety of activities including
sports shooting, hunting, national defence, and law enforcement.
1/ 1
() = 0 21
b , 0 ( ) (5) The speed profiles for these 36 types of bullets were obtained from
2 b 0 the manufacturer's data tables [15]. The data table for each type of
bullet (e.g. 0.338 Lapua Mag.) has columns for the calibre (8.6
where V0 is the muzzle (or initial) speed of the bullet. In (5), the mm70mm); product no. (N318020); type/weight/code
upper limit on x defines the maximum distance beyond which the (Naturalis/15.0g/N508); velocity at 0m (920m/s), 100m (835
bullet speed becomes subsonic. As Cb, v(x)V0 for all x m/s), 200m (753m/s), 300m (674m/s), 600m/s (473m/s), 800m
which, in the limit, is equivalent to the bullet travelling at a (375m/s); test barrel length (680mm). The bullet was fired in a
constant speed V0 [4, 5]. horizontal direction and its speed profile was measured using
The five different ballistic models [35, 10, 12, 13] considered Doppler radar, which provides the most accurate measurement of a
in this paper correspond to the five different values of : 0, 0.5, bullet's flight characteristics of any method yet devised [16]. With
0.514, 1, and 2. The five models are referred to as the exponential Doppler radar technology, the muzzle speed for each shot can be
speed (=0) [3, 13], quadratic speed (=0.5) [4, 5, 13], determined to be typically within 0.1m/s [16]. Multiple shots were
Duckworth (=0.514) [12], linear speed (=1) [10, 13], and used to measure the speed profile for each type of bullet. There are
constant drag force (=2) [13] models, respectively. a total of 36 data sets, each representing the radar measured speed
profile for a particular type of bullet. The speed data were available
2.1 Quadratic speed model (=0.5) at ranges of 0, 100, 200, 300, 600, 800, and for some bullet types,
900 and 1000m from the point of fire. Speed data falling below
When =0.5, (5) becomes 330m/s were discarded. The calibres of the 36 bullet types ranged
from 5.56 to 12.95mm, and the barrel lengths of the rifles ranged
2
() = 1/2
0 1
b , 0 b(1/2
0 1/2) . (6) from 510 to 960mm.
A least-squares (LS) fit of the ballistic model (5) (for a specific
The decreasing speed of the bullet is a quadratic function of the value of ) to the radar measured speed data for a given bullet type
distance travelled. Kozick et al. [13] referred to this model as the is given by
drag force model and used it to generate simulated data for their
^ 1 1/
study of the effect of model mismatch on the localisation accuracy ^
^ LS() = 0 2b , 1 (10)
for the point of fire.
^ ^
2.2 Duckworth model (=0.514) where K is the number of data points, and (0, b) are the values of
(V0, Cb) that provide the LS fit, or more precisely, that minimise
By postulating that the sound radiation (SW) dominates the drag the sum of the squared deviations of the observed bullet speeds
for supersonic bullets, Duckworth et al. [12] were able to find that from their predicted values
the exponent in the expression for the drag coefficient CD has a
numerical value of 0.514. Substituting =0.514 into (5) gives the
1/ 2

^ ^
Duckworth model. The decreasing speed of the bullet is a near- (0, b) = argmin ^ () 0 21
b (11)
(0, b) = 1
quadratic function of the distance travelled.
where ^ () is the observed bullet speed at range xk from the point
2.3 Exponential speed model (=0)
of fire. The root-mean-square (RMS) deviation rms and the
This model can be derived from (5) by taking the limit 0 maximum absolute (MA) deviation max of the LS fit of the
ballistic model from the speed data are defined, respectively, as
21
b
1 follows:
() = 0e , 0 (ln 0 ln ) . (7)
2 b
1 ^

2
The speed of the bullet decreases exponentially with the distance rms = () ^ LS() (12)
travelled. When =0, the drag coefficient CD as given by (2) is =1
independent of the bullet speed (i.e. CD=constant). Therefore, this
model is also called the constant drag coefficient model. max = max ^ () ^ LS() . (13)

2.4 Linear speed model (=1) The 36 data sets were processed in turn using (10)(13) for five
different values of : 0.5, 0.514, 0, 1, and 2, which correspond to
When =1, (5) becomes
the five specific ballistic models considered in this paper. The non-
1 linear LS minimisation (11) was implemented in MATLAB using
() = 0 21
b , 0 ( ) . (8) the optimisation toolbox function lsqnonlin. Figs. 1a and b show,
2 b 0
as examples, the LS fits of the respective quadratic speed and
The speed of the bullet decreases linearly with the distance constant drag force models to the radar measured speed data for
travelled (i.e. dv/dx=constant). bullet type 28, with the estimated values of (V0, Cb) shown on the
top of the associated figure. Fig. 2 shows for each of the 36 bullet
types, the values of (V0, Cb) that provide the LS fit of the quadratic
2.5 Constant drag force model (=2)
speed model to the corresponding set of radar measured speed data,
When =2, it can be shown using (1) and (2) that the drag force and Fig. 3 shows the resulting RMS error rms and the MA error
FD is independent of the bullet speed (i.e. FD=constant). max. Define the difference in RMS (MA) error between model k
Substituting =2 into (5) gives the constant drag force model and model 1 as the value of rms (max) for model k minus the value
of rms (max) for model 1, where 2k5 and models 15
1/2 1
() = 20 41
b , 0 (2 2) . (9) represent the quadratic speed, Duckworth, exponential speed,
4 b 0
linear speed, and constant drag force models, respectively. Fig. 4
shows (a) the difference in RMS error and (b) the difference in MA

2 IET Radar Sonar Navig.

Commonwealth Crown Copyright 2016
 The quadratic speed model is more accurate than the exponential
speed model for 33 of the 36 bullet types, and is more accurate
than the linear speed and constant drag force models for all 36
bullet types. The linear speed model has a similar accuracy to the
exponential speed model for about 75% of the total number of
bullet types. The constant drag force model is the least accurate,
followed by the linear speed model and then the exponential speed
model.
The accuracy of each ballistic model can be described
quantitatively by the mean values and standard deviations of rms
and max over all 36 bullet types. Table 1 shows these accuracy
measures for all five models. It is recommended that either the
quadratic speed model or the Duckworth model be used to describe
the variation with distance of a supersonic bullet's speed of travel
along its trajectory. However, as compared with the Duckworth
model, the quadratic speed model has the advantage that the detach
point of SW can be computed efficiently by finding the relevant
root of a sixth-order polynomial equation (rather than solving
numerically a non-linear equation) [5].

4Localisation using the quadratic speed model

and a single sensor node
4.1 Known ballistic model parameters
The quadratic speed model has been used to locate points of fire
using both MB and SW [4, 5] or only SW [5] assuming that the
two ballistic model parameters are known. While the MBSW
approach [4, 5] needs only a single sensor node (which consists of
an acoustic array), the SW approach [5] requires a network of
sensor nodes or distributed sensors. With the MBSW approach [4,
5], the range of the point of fire is estimated by measuring the
differential time of arrival (DTOA) and the differential angle of
arrival (DAOA) of the MB and SW at the sensor node, while the
direction of the point of fire is estimated by measuring the DTOA
of the MB at each sensor pair of the node (and so the direction
estimate is not affected by the ballistic model adopted). Fig. 5
shows the schematic diagram of the trajectory of a supersonic
bullet, the point of fire, and a sensor node. In Fig. 5, P is the detach
point at which the SW arriving at the sensor node was generated;
xp is the distance to P from the point of fire; is the Mach cone
angle at P; is the DAOA of the MB and SW at the node; and d
Fig. 1 LS fits of are, respectively, the miss angle and the miss distance for the node.
(a) Quadratic speed model, (b) Constant drag force model to radar measured speed Let denote the DTOA of the MB and SW at the sensor node.
data for bullet type 28 Using the quadratic speed model for the bullet, the range to the
point of fire from the sensor node is given by [4, 5]
4
1/2
0 1
b 2
= (14)
sin 1/2
0
1

b
2

where xp can be obtained by finding the relevant root of a six-order

polynomial equation whose coefficients are functions of , , V0,
and Cb (see [5] for details). If the constant speed model is used for
the bullet (which is equivalent to assuming Cb), the range to
the point of fire from the sensor node is given by [4, 5]

= . (15)
1 cos

By processing real data recorded from a field experiment, it has

been shown that the ranging accuracy of the MBSW localisation
method is significantly improved when the quadratic speed model
is used instead of the constant speed model [4, 5]. As an illustration
Fig. 2 Values of (V0, Cb) that provide the LS fit of the quadratic speed of the experimental results, Fig. 6 compares the cumulative
model to the corresponding set of speed data for each of the 36 bullet types probability distributions (CPDs) of the relative range errors that
result from using (15) (constant speed model) and (14) (quadratic
error between model k and model 1 for each bullet type (136) and speed model), respectively, for the following scenario. The point of
for each model number k (25). The following observations can be fire was located 476.9m away from the sensor node that consists
made from Fig. 4. of a three-element array, the miss angle was 1.2, and over 100
The small differences in both RMS error and MA error between rounds of a type A 7.62mm calibre bullet [4, 5] (V0=833.69m/s,
model 2 and model 1 for all 36 bullet types indicate that the Cb=90.82(m.s)1/2) were fired. The wavelet-based edge detection
quadratic speed and Duckworth models have a similar accuracy.

IET Radar Sonar Navig. 3

Commonwealth Crown Copyright 2016
 and MB at the node, which were then substituted into (14) or (15)
to obtain the range estimate for the point of fire. More details on
the experimental setup, sensor configuration, and data processing
can be found in [5]. The superiority of the quadratic speed model
(labelled 1) over the constant speed model (labelled 4) for ranging
the point of fire is clearly observed from Fig. 6. However, the
quadratic speed model requires a priori knowledge of the muzzle
speed and ballistic constant of the bullet; any uncertainty in these
two parameters will degrade the ranging accuracy. For example,
using the muzzle speed and ballistic constant for a type B 12.95
mm calibre bullet (V0=870.94m/s, Cb=183.05(m.s)1/2) in the
quadratic speed model when bullet A is fired will produce a large
range bias error. The dashed-dotted curve (labelled 3) in Fig. 6
denotes the CPD of the relative range error when bullet A is
mistaken as bullet B.

4.2 Unknown ballistic model parameters

Fig. 2 shows the wide ranges of muzzle speeds and ballistic
constants that characterise the 36 different types of bullets for the
Fig. 3 RMS error rms and MA error max that result from fitting the
quadratic speed model (similar distributions of these two
quadratic speed model to the corresponding set of speed data for each of parameters were observed for the other four models). It can be seen
the 36 bullet types that the ballistic constant of one bullet type can be very different
from that of another bullet type, even if the bullets have the same
calibre. Thus, there is no one-to-one relationship between the
ballistic constant of the bullet and its calibre. This observation
agrees with (4), which shows the dependence of the ballistic
constant on the calibre (through the cross-sectional area) and other
physical properties of the bullet. On the other hand, the muzzle
speed of a bullet depends not only on the bullet type but also on the
length of the rifle's barrel. Fig. 7 shows the effect of the barrel
length on the muzzle speed for a particular type of bullet [18]; the
muzzle speed increases with the barrel length. Therefore, a priori
knowledge of the muzzle speed and ballistic constant of the bullet
requires a priori knowledge of the weapon type.
In the practical case where the weapon type or the quadratic
speed model parameters of the bullet are unknown, a network of
sensor nodes can be used to estimate simultaneously in situ the
muzzle speed and ballistic constant of the bullet along with the
range of the point of fire [6]. Recently, a method [19] has been
proposed for ranging small arms fire using DTOA and DAOA
measurements of the MB and SW at a single sensor node
collocated with the target, when the quadratic speed model
parameters are only known to belong to a two-dimensional (2D)
parameter space. The proposed method [19] was applied to the
DTOA and DAOA measurements from the previous example,
assuming that the 2D parameter space was defined by 650V0
1000m/s and 30Cb210(m.s)1/2. The lower and upper bounds
of the miss angle, which are required by the method, were set equal
to 1 and 3, respectively. Fig. 6 shows, as a dotted curve
(labelled 2), the resulting CPD of the relative range error. The
proposed method was originally developed for a single target with
a collocated sensor node, assuming (reasonably) a small miss
angle. However, in practical situations, there may be two or more
potential targets (or protected assets) in the same area. In order to
apply the proposed method to these situations, it is necessary that
each potential target is equipped with a collocated sensor node. For
example, a potential target can be a vehicle with a sensor node
mounted on the roof. The actual target under attack can be
identified by determining which sensor node/potential target has
the shortest miss distance, or equivalently, the shortest duration of
the received SW waveform [7]. Once the actual target is identified,
the proposed method can be applied to the DTOA and DAOA
measurements of the MB and SW at the sensor node of that target.
Fig. 4 Difference in RMS error and MA error
(a) Difference in RMS error, (b) Difference in MA error between model k and model 1
for each bullet type (136) and for each model number k (25). Models 15 represent 5Conclusions
the quadratic speed, Duckworth, exponential speed, linear speed, and constant drag The accuracies of the five different ballistic models in predicting
force models, respectively bullet speed profiles have been evaluated and compared using radar
measured speed data for 36 types of bullets. The quadratic speed
algorithm [17] was employed to estimate the arrival times of the model has a similar accuracy to the Duckworth model and is much
SW and MB at each sensor element, and the results were used to more accurate than the other three models. When compared with
compute the estimates of the DTOA and DAOA of the SW the Duckworth model, the quadratic speed model has the advantage

4 IET Radar Sonar Navig.

Commonwealth Crown Copyright 2016
Table 1Mean values and standard deviations of rms and
max over all 36 bullet types for the five models
Model rms, m/s max, m/s
Mean Standard Mean Standard
deviation deviation
quadratic speed 1.70 1.92 2.73 3.21
Duckworth 1.82 1.94 2.91 3.25
exponential speed 7.61 3.09 11.20 4.84
linear speed 9.09 4.10 12.69 5.84
constant drag force 25.20 8.86 36.96 13.52

Fig. 7 Effect of barrel length on muzzle speed for a particular type of

bullet [18]

[2] Donzier, A., Millet, J.: Gunshot acoustic signature specific features and false
alarms reduction. Proc. of SPIE 5778, 2005, pp. 254263
Fig. 5 Schematic diagram of the trajectory of a supersonic bullet, point of [3] Bedard, J., Pare, S.: Ferret, a small arms fire detection system: localization
concepts. Proc. of SPIE 5071, 2003, pp. 497509
fire, and sensor node
[4] Lo, K.W., Ferguson, B.G.: A ballistic model-based method for ranging direct
fire weapons using the acoustic muzzle blast and shock wave. Proc. of
ISSNIP, 2008, pp. 453458
[5] Lo, K.W., Ferguson, B.G.: Localization of small arms fire using acoustic
measurements of muzzle blast and/or ballistic shock wave arrivals, J. Acoust.
Soc. Am., 2012, 132, pp. 29973017
[6] Lo, K.W., Ferguson, B.G.: Simultaneous classification and ranging of direct
fire weapons using an asynchronous acoustic sensor network. Proc. of
ISSNIP, 2011, pp. 425430
[7] Volgyesi, P., Balogh, G., Nadas, A., et al.: Shooter localization and weapon
classification with soldier wearable networked sensors. Technical Report,
TR# ISIS-07-802, Institute for Software Integrated Systems, Vanderbilt
University, 2007
[8] Ledeczi, A., Nadas, A., Volgyesi, P., et al.: Countersniper system for urban
warfare, ACM Trans. Sensor Netw., 2005, 1, pp. 153177
[9] Balogh, G., Ledeczi, A., Maroti, M.: Wireless sensor network-based
projectile trajectory estimation. Technical Report, TR# ISIS-05-601, Institute
for Software Integrated Systems, Vanderbilt University, 2005
[10] Lindgren, D., Wilsson, O., Gustafsson, F., et al.: Shooter localization in
wireless microphone networks, EURASIP J. Adv. Signal Process., 2010, pp.
690732/1690732/11
[11] Damarla, T., Kaplan, L.M., Whipps, G.T.: Sniper localization using acoustic
asynchronous sensors, IEEE Sens. J., 2010, 10, pp. 14691478
[12] Duckworth, G.L., Gilbert, D.C., Barger, J.E.: Acoustic counter-sniper
system. Proc. of SPIE 2938, 1997, pp. 262275
[13] Kozick, R.J., Whipps, G.T., Ash, J.N.: Supersonic projectile models for
Fig. 6 CPDs of relative range errors that result from using the constant asynchronous shooter localization. Proc. of SPIE 8046, 2011, pp. 804604-1
speed model (labelled 4) and the quadratic speed model when (i) the 804604-15
ballistic model parameters are exactly known (labelled 1); (ii) a wrong set [14] Weinacht, P., Cooper, G.R., Newill, J.F.: Analytical prediction of trajectories
for high velocity direct-fire munitions, ARL-TR-3567, U.S. Army Research
of ballistic model parameters: V0=870.94m/s and Cb=183.05(m.s)1/2 is Laboratory, Aberdeen Proving Ground, MD 21005-5066, August 2005
used (labelled 3); (iii) the ballistic model parameters are only known to [15] The data used for model evaluation were obtained from http://
www.lapua.com/en/home.html, accessed 6 May 2013
belong to the 2D parameter space [19]: 650V01000m/s and 30Cb [16] Arrow Tech Associates: Uses of radar 2000 during projectile development &
210(m.s)1/2 (labelled 2). The point of fire was 476.9m away from the production (Arrow Tech Associates, 2009). Available at http://
www.prodas.com/XQ/ASP/P.603/QX/Documents/Uses%20of%20Radar
sensor node and the miss angle was 1.2. Over 100 rounds of ammunition
%202000%20during%20Projectile%20Development.pdf, accessed 29 April
with ballistic model parameters V0=833.69m/s and Cb=90.82(m.s)1/2 2016
were fired [17] Sadler, B.M., Pham, T., Sadler, L.C.: Optimal and wavelet-based shock wave
detection and estimation, J. Acoust. Soc. Am. 1998, 104, Pt. 1, pp. 955963
[18] Arvidsson, P.G.: Is there a problem with the lethality of the 5.56 NATO
that the detach point of SW can be computed more efficiently. The calibre?. Available at http://www.dtic.mil/ndia/2010armament/
quadratic speed model has been applied to accurate ranging of WednesdayLandmarkBPerArvidsson.pdf, accessed 29 April 2016
small arms fire using a single sensor node, with and without a [19] Lo, K.W., Ferguson, B.G.: Acoustic ranging of small arms fire using a single
sensor node collocated with the target, J. Acoust. Soc. Am., 2015, 137, pp.
priori knowledge of the ballistic model parameters, and its EL422EL428
effectiveness for both cases has been demonstrated using real data
recorded from a field experiment.

6References
[1] Donzier, A., Cadavid, S.: Small arm fire acoustic detection and localization
systems: gunfire detection system. Proc. of SPIE 5778, 2005, pp. 245253

IET Radar Sonar Navig. 5

Commonwealth Crown Copyright 2016