IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO.

3, MARCH 2009 1119

Using Hidden Markov Models

in Vehicular Crash Detection
Gautam B. Singh, Senior Member, IEEE, and Haiping Song

Abstract—This paper presents a system for automotive crash rithms and are generally tuned to detect specific types of crash
detection based on hidden Markov models (HMMs). The crash events. These enhancements over the performance of the first-
pulse library used for training comprises a number of head- generation products were achieved through the incorporation
on and oblique angular crash events involving rigid and off-
set deformable barriers. Stochastic distribution characteristics of complex processing logic [2]. The fundamental paradigm
of crash signals are validated to ensure conformity with the of analyzing the signal parameters in the time domain con-
modeling assumptions. This step is achieved by analyzing the tinues to provide the necessary underpinnings to algorithmic
quantile–quantile (Q–Q) plot of actual pulses against the as- enhancements.
sumed bivariate Gaussian distribution. HMM parameters are next Existing crash detection algorithms can be divided into two
induced by utilizing the expectation–maximization (EM) proce-
dure. The search for an optimal crash pulse model proceeds categories, namely, 1) those that utilize changes in vehicle
using the “leave-one-out” technique with the exploration encom- speed and 2) those that utilize the vehicular crush information.
passing both fully connected and left–right HMM topologies. The speed-based sensors rely on variables such as speed change
The optimal crash pulse architecture is identified as a seven- ΔV , jerk, speed, displacement, energy, etc. The crush-based
state left–right HMM with its parameters computed using real algorithms utilize data from one or more sensors mounted in
and computer-aided engineering (CAE)-generated data. The
system described in the paper has the following advantages. the crush zone, which help predict crash severity. In both cases,
First, it is fast and can accurately detect crashes within 6 ms. crash-sensing algorithms are developed using a parameter tun-
Second, its implementation is simple and uses only two sensors, ing process where engineers analyze a library of crash pulses
which makes it less vulnerable to failures, considering the overall and develop rules and thresholds for airbag deployment. The
simplicity of interconnects. Finally, it represents a general and resulting algorithms are not universal.
modularized algorithm that can be adapted to any vehicle line and
readily extended to use additional sensors. In recent years, crash-detection algorithms have begun uti-
lizing machine learning and pattern-recognition techniques. For
Index Terms—Automotive crash detection, computer-aided example, Barnard and Riesner’s algorithm uses momentum and
engineering (CAE), continuous-value emission hidden Markov
models (HMMs), crash pulse, discrete-value emission HMM, energy in semimetric spaces [3]. Singh and Song’s earlier al-
finite-element analysis (FEA). gorithm utilized a discrete Markov model to detect crashes [4].
Yin et al. proposed an algorithm based on artificial neural net-
works [5]. The goal has been to develop smart airbag systems
I. I NTRODUCTION
that not only can predict crash severity but can also consider the

W ITH THE vast majority of vehicle consumers citing

safety features as an important factor in their purchasing
decision, crashworthiness of vehicles plays an important role
occupant’s size, weight, position, and velocity in a multistage
bag deployment regime. Such a task is fairly complex but
achievable through a trainable algorithm that utilizes inductive
in establishing quality and customer acceptance. In addition learning procedures such as that described in this paper. We
to meeting customer expectations, tangible benefits realized by demonstrate that a pattern recognition system trained on a large
a reduction in fatal and nonfatal injuries are estimated to be variety of crash acceleration classes constitutes an effective
around $US 2 billion [1]. solution for designing smart airbag systems.
From a technological standpoint, an early detection of crash Hidden Markov models (HMMs) are being applied in many
event allows safety engineers to better tailor airbag inflation fields [6] such as control [7]–[9], communications [10], [11],
rates, pressures, and belt-pretensioning systems. The automo- handwriting and text recognition [12], [13], image processing,
bile industry continues to make advancements by providing computer vision [11], [15], [16], and bioinformatics [17]–[19].
smarter restraint systems. The current crash-detection tech- Most state-of-the-art speech systems today are based on HMMs
niques are based on multistage sequential signal analysis algo- [20], [21]. The reason that HMMs succeed in accurately mod-
eling the real-world phenomenon is attributable to their flexible
Manuscript received July 5, 2005; revised November 18, 2005, architecture and stochastic learning bases, rendering them gen-
November 27, 2006, July 28, 2007, November 6, 2007, February 5, 2008, erally immune to overtraining when an appropriate number of
April 10, 2008, and May 5, 2008. First published July 22, 2008; current version hidden states are used, and sufficient observation distributions
published March 17, 2009. The review of this paper was coordinated by
Dr. M. S. Ahmed. and training data are used for learning.
The authors are with the Department of Computer Science and Engineering, An HMM models a doubly stochastic process: First, there
Oakland University, Rochester, MI 48309 USA (e-mail: [email protected]). is an underlying stochastic process that is not observable, and
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. second, this hidden process influences another process that pro-
Digital Object Identifier 10.1109/TVT.2008.928904 duces the sequence of observable symbols. When applied to the

0018-9545/$25.00 © 2008 IEEE

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 1120 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 3, MARCH 2009

crash-detection problem, the observed symbols are the ob- either. However, the data set usually contains at least
served accelerometer readings. The hidden states correspond one crash condition above the deployment threshold [2].
to the underlying crushing, recoil, elastic, and plastic deforma- Typically, the pulses in this group are utilized to set up
tions caused by crash dynamics. Since these hidden changes thresholds.
should be similar for similar types of vehicles and crash sit- 3) The third group includes higher speed crashes of various
uations, it should be possible to reliably correlate symbols types, including head-on and angular (usually 30◦ ) im-
observed and recorded by the accelerometer sensors in an pacts with a barrier, crashes with a center pole, as well as
aggregate sense. crashes with offset poles.
This paper extends our earlier work on discrete HMMs where Sensor Locations: Crash pulse libraries are a collection of
the underlying crash-detection model utilized quantized crash crash pulses measured by accelerometers mounted at specific
pulse data [4]. In this paper, the crash pulse data is not quantized locations in the vehicle. Typical crash sensor locations include
and used for inducing a continuous-mode HMM. Our results the radiator bracket, inside fender, frame rails, steering wheel,
with the continuous-mode HMM are far more superior than pillar rockers, etc. In general, two crash pulses or acceleration
those obtained using a discrete HMM. The continuous-mode signatures are recorded for each crash experiment [22]. One
HMM detects a large variety of crash events within 6 ms of of the crash pulses recorded is for the crush zone located
onset. in the engine compartment. The crush zone in a vehicle is
designed to absorb a significant portion of the crash energy as it
II. C RASH P ULSE L IBRARY crushes during impact. The other crash pulse is recorded from
the passenger compartment, where the goal is to minimize the
Crash libraries are generally developed by the OEM airbag mechanical deformations and maximize its integrity during a
module suppliers in close collaboration with automobile man- crash.
ufacturers. The data set used in this paper is from a complete In our case, the crash pulse library, whether actual or sim-
crash library that was actually used by an automobile manu- ulated, consists of acceleration values in the X-direction only
facturer. A crash library contains a collection of acceleration for two sensor locations: one each in the crush zone and the
data measured by accelerometers. A crash library contains ac- passenger compartment.
celerometer values measured in real-world experiments where
a specific vehicle is test-crashed under a set of controlled
variables, including the vehicle speed, barrier type, and angle A. Crash Pulses: Real Crashes
of impact. A crash library may also contain 3-D accelerometer While learning inductive models, the quality and quantity
samples obtained by simulating a crash using computer-aided of the observations govern the overall performance of a clas-
engineering (CAE) models. sification system. Since the dimensionality of the problem is
Acceleration Data: The vehicle acceleration values used generally high, and the training data are finite and nonuniformly
for crash sensing are measured by an electronic sensor called sampled, the modeling problem is almost always ill-posed [23].
the accelerometer. It measures instantaneous values for vehicle Within the context of crash classification, engineers construct
acceleration as it undergoes crash. The accelerometer produces a crash pulse library composed of accelerometer readings for
a 3-D analog signal proportional to the acceleration of the a carefully selected set of crash events that would hopefully
vehicle in three dimensions: X—forward, rear to front of the capture the variability in pulse signatures.
vehicle; Y —right, driver to passenger side; and Z—down, roof Crash pulse data are captured by crashing vehicles by col-
to floor. liding them against each other, against rigid and deformable
Fig. 1 shows the acceleration signatures in X- and barriers, and against stationary poles and moving barriers [22].
Y -directions, or the crash pulse, for two cases. In all cases, Impact configurations are varied as well with front, rear,
the crash event occurs at time t = 0. Fig. 1(a) shows crash bumper override, and overlap collisions. The acceleration val-
pulses for an event that does not require airbag deployment. The ues are recorded and collectively constitute a crash pulse library
magnitude of acceleration sensed in these cases is bounded by for a vehicle line that the manufacturer uses to calibrate its
about 10 g but can go up as high as 50 g. Crash pulses requiring sensors for that line.
airbag deployment are shown in Fig. 1(b). The magnitude of An analog-to-digital converter transforms the analog ac-
instantaneous accelerations along the X-axis in this case can celeration signal values into a uniformly sampled digital
reach up to 300–500 g. time series. The accelerometer crash pulse is prefiltered at
Crash Pulse Groups: Generally, the crash pulses in a library a rolloff frequency of 4 kHz and then sampled at a rate of
may broadly be classified into three groups. 12 500 samples/s. A digital filter SAE J211 is utilized to obtain
1) The first group consists of the nondeployment cases, the useful frequency components of the acceleration signal [22].
including low-speed and rough-road driving conditions. As the acceleration experienced by the vehicle is dependent on
A vehicle-to-barrier crash of 6–9 mi/h is considered to be the physical characteristics of the vehicle and the barrier, the
a low-speed crash. crash pulses in a library are a unique signature of a specific
2) The second group consists of medium-speed crashes of vehicle.
12–25 mi/h. Crash types can be diverse and include head- Fig. 2 compares the characteristics of a “nondeploy” and a
on collisions and pole-impact signals. Generally, crashes “deploy” crash pulse. The waveform characteristics before and
in this group do not require the deployment of the airbag after it has been filtered are illustrated. The Δv profile of the
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 SINGH AND SONG: USING HMMs IN VEHICULAR CRASH DETECTION 1121

Fig. 1. (a) Acceleration in X- and Y -directions during a noncrash scenario. (b) Acceleration in X- and Y -directions for a crash that requires airbag deployment.

two pulses is also shown. The slightly convex shape of the Δv compares the real crash data and the simulated data generated
profile in Fig. 2(b) is indicative of the rapid drop in velocity by FEA for signals observed in the noncrush and crush sensors1
to bring the vehicle to a stop in about the same time as the in a 35-mi/h frontal barrier impact crash test.
nondeploy case in Fig. 2(a). The FEA methods are based on mathematical techniques for
analyzing stresses encountered in the car crash by modeling the
physical structure into substructures called “finite elements.”
B. Crash Pulses: Simulated Crashes The finite elements and their stress interrelationships are con-
Due to the expense of crashing a vehicle for adding a crash verted into mathematical equations that are solved using com-
pulse to the library, the number of pulses in the library tends putational techniques. In general, the crush zones are designed
to be limited due to economical concerns. The advancement to absorb most of the impact energy. The pulses produced by
of CAE provides indispensable tools in today’s vehicle design. FEA are filtered by a three-stage 4-kHz, low-pass digital filter.
Today, in addition to the actual measurements of acceleration
data, a crash library contains data obtained from running finite- 1 These locations are physically identified on a vehicle’s body or chassis and
element analysis (FEA). Simulated pulses from FEA can pro- labeled as such based on the extent to which a specific zone is crushed or not
vide accurate lower frequency range information [24]. Fig. 3 crushed in the event of a typical crash.
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 1122 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 3, MARCH 2009

Fig. 2. Crash pulse characteristics. (a) “Nondeploy” 14-mi/h crash pulse after filtering and its Δv profile. (b) “Deploy” 40-mi/h crash pulse after filtering and its
Δv profile. Note that Δv at the conclusion of the crash event is equal to the speed of the vehicle at the time of the crash.

Fig. 3. Actual and filtered crash pulses. (a) Before filtering. (b) After filtering with a 4-kHz low-pass filter. Although the original crash pulses look distinct,
filtered signals show substantial correspondence.

Fig. 3(a) and (b) depicts the waveforms before and after filtering are used for training our HMM. The positive direction of the
for both real and CAE-generated pulses, respectively. A Δv X-axes is from rear to front of the vehicle. Overall, the crash
analysis (Δv = a dt) and the power spectrum for the two pulse library is composed of 21 actual crash pulses and 15 crash
types of pulses are shown in Fig. 4(a) and (b), respectively. pulses generated by simulation of CAE models.
The comparison of the power spectrum indicates that the
CAE-generated pulses exhibit good correspondence to the ac-
III. HMM OF V EHICULAR C RASH P ULSES
tual pulses. This is also evident in the filtered signal and Δv
analysis. HMMs are a class of models for representing the probability
For training our HMM, we utilize accelerometer data ob- distribution of an observed time series. They are essentially
tained from sensors in both the “crush” and “noncrush” zones. stochastic finite state machines that output a symbol or a real
Since all crashes in the library are either head-on or slightly value each time they leave a state. By specifying the state
oblique, only the data along X-axes (longitudinal/forward) transition probabilities between states and the output generation
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 SINGH AND SONG: USING HMMs IN VEHICULAR CRASH DETECTION 1123

Fig. 4. Δv analysis and power spectrum of real and CAE-generated crash pulses. Again, the real and simulated pulses show close correspondence.

probabilities for each state, we can attempt to capture the data, we construct a chi-square plot for each sample, as shown
underlying structure of the system that generates the output. in the list that follows.
1) Di2 are ordered from smallest to largest as D(1) 2
,
A. Multivariate Normality Test 2
D(2) 2
, . . . , D(n) .
The assumption that signals belong to a single Gaussian 2) The pair (D(j) 2
, χ2p ((j − 0.5)/n)) is plotted, where χ2 is
distribution must be validated as a prerequisite for using an the 100(j − 0.5)/n percentile of the chi-square distribu-
HMM for representing crash pulses. The method for assessing tion with df = p degrees of freedom.
multivariate normality utilizes the Mahalanobis distance and is
Fig. 5(a) is one of the Q–Q plots for a crash pulse. The plot
obtained by plotting this distance against chi-square percentiles
for each group produces a straight line with a norm of residual
[25] in a quantile–quantile (Q–Q) plot. The Mahalanobis dis-
of 5.9 and, thus, conclusively establishes that the crash pulses
tance is utilized for determining the similarity of an unknown
approximately follow a multivariate normal distribution. The
sample set to a known set [14], [16]. This distance metric
norm of residual, which is the largest singular value of the
takes into account the correlations of the data set and is scale
residual, can be used as a measure of the goodness of a linear
invariant.
regression. Fig. 5(b) shows a table of the norm of residuals for
A Q–Q plot is a comparison of two distributions, either or
21 real crash pulses, 15 of which are representative of crash
both of which may be empirical or theoretical. When we com-
events requiring airbag deployment and six being nondeploy
pare the probability of distributions of two random variables,
crash events. Similar results were also obtained with CAE
the Q–Q plot will result in a straight line if the two variables
pulses and, thus, also established their correspondence with a
come from a common distribution. The advantages of the
multivariate normal distribution.
Q–Q plot are that the sample sizes do not need to be equal
Crash detection is a real-time process, and thus, a simple
and that many distribution aspects, such as shifts in location
model is preferred to enable us in making crash classifica-
and changes in symmetry, can simultaneously be tested. The
tion within hard real-time deadlines. The fully connected and
Q–Q plot is particularly good for discriminating in the tail
left–right HMMs are the most commonly used models. A
areas of the distributions since quantiles more rapidly change
left–right HMM has only forward transitions, implying that the
in those ranges, in turn requiring coverage between consecutive
state transitions may never occur that take the HMM back to
quantiles.
a state that has left in the past. The left–right HMMs have
Assume that there are n independent observations
successfully been used in many time series applications. The
X1 , . . . , Xn , with each observation Xi = [xi1 , xi2 , . . . , xip ]
fully connected HMMs are not subject to any such restriction,
being a p-dimensional random vector sampled from a
and it is possible for the HMM to transition from any state to
multivariate normal distribution with mean vector μ and
another state during any time epoch.
covariance matrix Σ. The squared Mahalanobis distance
As will be discussed in Section III-D, we use seven states to
between Xi and μ is then given by
model the acceleration readings from the onset of the crash until
Di2 = (Xi − μ)T Σ−1 (Xi − μ). (1) 6 ms after the crash. The seven states represent the movement
of the vehicle within 6 ms after the crash, i.e., moving forward,
When both n and (n − p) are greater than 25, the Di2 are bouncing backward, moving forward again, bouncing backward
distributed as independent chi-squared variate with p degrees again, and so on. At any time index t, it can either move forward
of freedom df [26]. Because this is the case for the crash pulse to the next state or stay at the current state. A comprehensive
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 1124 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 3, MARCH 2009

Fig. 5. (a) Test for multivariate normality. (b) Norm of residuals for linear regression of the Q–Q plot.

review of HMMs is provided in [10]. In this section, we will where bj (ot ) is the probability of observation at time t given
briefly review the learning evaluation of HMMs with real- the hidden state is Sj , i.e., bj (ot ) = P (ot |qt = Sj ), and Σj is
valued emission. the covariance matrix for state Sj , i.e.,

P (qt = Sj )(Ot − μj )(Ot − μj )
B. Components in an HMM Σj = t  . (2)
t P (qt = Sj )
In this paper, the sequence of observations is denoted as
O = o1 , o2 , . . . , oT . The corresponding hidden state sequence For a multivariate normal distribution, the emission probability
is denoted as Q = q1 , q2 , . . . , qT . Supposing that the obser- can be written as
vation has p dimensions, an HMM with N states has four  
1 (x − μj ) ∗ Σ−1
j ∗ (x − μj )
p exp −
components. bj (x) = .
1) Initial state distribution Π = {πi }(N ×1) , which is an N × (2π) 2 |Σj | 2
1 vector. πi is the probability of the event that the first (3)
hidden state is Si . It can be written as πi = P (q1 = Si ). The expectation probability is recursively computed using the
2) Transition matrix A = {aij }(N ×N ) , which is an (N × following variables.
N ) matrix. aij is given by aij = P (qt+1 = Sj |qt = Si ),
1) αt (i): The forward variable is defined as the probability
where qt is the hidden state at time t, and Si is the ith
that the state after t observations is Si , i.e.,
state.
3) Mean matrix μ = {μi }p×N , which is a (p × N ) matrix. αt (i) = P (o1 , o2 , . . . , ot , qt = Si |λ) (4)
μi is the mean value of the ith state.
4) Covariance matrix Σ(p×p) , which is a (p × p) matrix. and can recursively be calculated using
In this paper, we denote an HMM as λ = {Π, A, μ, Σ}. ⎧
⎨π b (o ), t = 1; 1 ≤ j ≤ N
j Nj 1
αt+1 (j) = α (i)aij bj (ot+1 ), 1 ≤ t ≤ T − 1;
C. Training HMMs With Real-Valued Emission ⎩ i=1 t
1 ≤ j ≤ N.
The Baum–Welch algorithm is used to train an HMM. The (5)
Baum–Welch algorithm can be divided into two steps, i.e., 2) βt (i): The backward variable is defined as the probability
the E-step, where the likelihood P (O|λ) is calculated, and that the sequence of observations from (t + 1) to T is
the M-step, where the likelihood by updating the model λ is observed after the system has been in state Si at t, i.e.,
maximized.
1) E-Step: The likelihood of a crash pulse P (O|λ) is βt (i) = P (ot+1 , ot+2 , . . . , oT |qt = Si , λ). (6)
defined as follows:
 Moreover, it can recursively be calculated using
P (O|λ) = P (O|Q, λ)P (Q|λ)
1,
N t = T; 1 ≤ i ≤ N
all Q
 βt (i) = j=1 aij bj (ot+1 )βt+1 (j), T − 1 ≥ t ≥ 1;
= πq1 bq1 (o1 ), aq1 q2 bq2 (o2 ), . . . , aqT −1 qT bqT (oT ) 1 ≤ i ≤ N.
all Q (7)
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 SINGH AND SONG: USING HMMs IN VEHICULAR CRASH DETECTION 1125

Evaluation of an HMM is much simpler than training an by the method of mixtures [28]. Here, we use an estimator
HMM. To calculate the likelihood P (O|λ) by substituting based on the likelihood of crash pulses. The rationale for using
(4) into (2), we have this estimator is that the likelihood is a by-product of HMM
 parameter estimation, which will ultimately be used to discrim-
P (O|λ) = P (O|Q, λ)P (Q|λ) inate crash pulses from noncrash pulses. From the statistical
all Q learning theory, the likelihood can be written as P (O|λ), where

= πq1 bq1(O1 ), aq1 q2 bq2(O2 ), . . . , aqT −1 qT bqT(OT ) O is the observation (data) under the model λ. The likelihood
is defined as the probability of the data (i.e., observation),

N 
N
= αT (i) = πi bi (o1 )β1 (i). (8) given the underlying model. The probability P (O|λ) can be
i=1 i=1 calculated using (8).
Inductive modeling procedures typically begin by dividing
2) M-Step: In the M-Step, we update the model λ = the available data into two sets, namely, a training set and a
{Π, A, μ, Σ} in such a way that Baum et al. have proved would testing/evaluation set. The general rule of thumb is to use two
guarantee the increase of likelihood, i.e., thirds of the data for training and leave the remaining one third
for evaluation and follow the procedure outlined here to induct
π j = γ1 (j), j = 1, . . . , N (9) and validate the model.
T −1
ξt (i, j)
aij = t=1 T −1
, i = 1, . . . , N 1) Divide the available data into a training set and a test-
t=1 γt (i) ing/evaluation set.
j = 1, . . . , N (10) 2) Split the training data set into n folds; each fold contains
T
t=1 γt (j) · ot approximately the same number of crash pulses.
μj =  T
, j = 1, . . . , N (11) 3) Set the parameters for the model being validated (in our
t=1 γt (j)
T case, this is the HMM topology, number of states, and
t=1 γt (j) · (Ot − μj )(Ot − μj ) accelerometer buffer size).
Σj = T (12)
t=1 γt (j)
4) Use (n − 1) folds to train a model and the remaining
folds for testing.
where ξt (i, j) is the probability of being in state Si at time t 5) Repeat steps 2–4 for each of the architectures—this is
and state Sj at time t + 1, given the model and the observation analogous to searching the architecture space for an
sequence, i.e., optimal model.
6) Select the best model and train it using the entire
ξt (i, j) = P (qt = SI , qt+1 = Sj |O, λ)
training set.
αt (i)aij bj (Ot+1 )βt+1 (j) 7) Benchmark the final model using the evaluation set.
=
P (O|λ) These are the data that the model has never seen before,
αt (i)aij bj (Ot+1 )βt+1 (j) which provide a more accurate estimate of the perfor-
= N N (13) mance on the testing/evaluation data selected in step 1.
i=1 j=1 αt (i)aij bj (Ot+1 )βt+1 (j)
Steps 2–4 are devised to establish the optimal HMM config-
and γt (i) is the probability of being in state Si at time t, given uration. This technique is also referred to as the leave-one-out
the observation sequence O and the model λ, i.e., strategy. This strategy is applied by keeping the topology fixed
αt (i)βt (i) and optimizing the detection system architecture with different
γt (i) = P (qt = Si |O, λ) = parameter settings for a number of states and pulse storage
P (O|λ)
buffer sizes. In the same manner, the HMM topology is varied
αt (i)βt (i)  N
to establish the optimal number of states and interconnectivity
= N = ξt (i, j). (14)
i=1 αt (i)βt (i)
(fully connected or left–right connected). Subsequently, after
j=1
selecting the best topology and training parameters, all the
training data are utilized to train the final model.
The process of establishing and validating the crash model in
D. Inducing the Optimal Crash Pulse Model
our study is somewhat modified to facilitate learning a compre-
The process of crash detection is based on accumulation of hensive set of crash characteristics from a complete set of crash
readings sampled from the onboard acceleration sensors and pulses. Accordingly, two optimal HMMs are constructed. In the
stored in a buffer. Since crash sensing is a real-time applica- first instance, an optimal HMM is induced using the entire set
tion, simpler algorithms are preferred if the detection accuracy of “actual” crash pulses with the entire set of “simulated” pulses
requirements can be satisfied. We have developed experiments serving as the testing set. In the second instance, the entire
to optimize the HMM topology, the number of states, and the set of “simulated” pulses is used for training with the “actual”
size of the buffer used to store the past accelerometer readings. crash pulses serving as the testing set. Thus, two optimal crash
Consistent estimators based on real crash data are used to HMMs, i.e., one with real pulses and the other with CAE pulses,
optimize HMM parameters. Ziv and Merhav proposed an esti- are induced and validated using the other data set.
mator of the number of states of an HMM based on entropy and Due to the lack of information on the noncrash events, our
data compression [27]. Liu and Narayan proposed an estimator training process is geared toward training crash HMMs only;
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 1126 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 3, MARCH 2009

Fig. 6. “Leave-one-out” cross validation for (I) fully connected HMMs and (II) left–right HMMs. The Z-axis represents the normalized log likelihood of the ten
pulses, and the Y -axis represents the fold index. (a) X-axis represents the number of states from 1 to 20. (b) X-axis represents the sequence length in milliseconds.

Fig. 7. Seven-state left–right HMM is used to model crash events.

we do not train any noncrash HMMs, as would sometimes be pulses yield similar LL(N, L) characteristics. This further val-
useful when measuring the log likelihood of a crash versus idates our earlier observations that HMMs are noise tolerant.
a noncrash. Thus, although the entire crash pulse library is For models with similar likelihoods, the simplest model is
composed of both the real and simulation data, the use of preferred. Thus, a seven-state left–right HMM with a 6-ms
one type of data for induction in one case ensures that all buffer size is chosen to implement a crash-detection sys-
representatives pulses in a set are utilized in the computation tem. This induced HMM is shown in Fig. 7. The μ =
of model parameters as both sets are generally developed to be [μnoncrush , μcrush ]T vector associated with each state is the
comprehensive to encompass all major crash events. average acceleration for the noncrush and crush zones. Class-
Fig. 6(I) shows the LL(N, F ) and LL(L, F ) for crash pulses ification of crash pulses using this model is presented next.
for the fully connected HMMs, where LL is the normalized log
likelihood, N is the number of states of the HMM, F is the
IV. R ESULTS
fold index, and L is the pulse length considered. The likelihood
is normalized with respect to a random sequence of the same Left–right HMMs with seven states are trained by two sets
length. Thus, we can conclude that when N ≥ 7 and L > 5 ms, of crash pulses. These two sets are given as follows: 1) crash
all the folds show great consistency for a fully connected model. pulses captured through actual accelerometer measurements in
Fig. 6(II) shows the LL(N, F ) and LL(L, F ) of crash a crash or 2) the set of pulses generated by the simulation of
pulses for the left–right HMMs. From this, we conclude that CAE models. These two sets are independently treated. When
an increase beyond N ≥ 7 and L > 5 ms (an increase in the the crash pulse model is induced using the set of pulses from
number of states to more than seven or the buffer size with actual crash data, model validation is performed using the set of
storage larger than 5 ms) does not result in further performance CAE pulses. Similarly, when the crash pulse model is induced
improvements. Furthermore, Fig. 6 demonstrates that with a using the CAE pulses, model validation is performed using
sufficient number of states and sequence length, all training actual crash pulses.
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 SINGH AND SONG: USING HMMs IN VEHICULAR CRASH DETECTION 1127

TABLE I TABLE II
CLASSIFICATION OF “REAL” CRASH PULSES USING AN HMM INDUCED CLASSIFICATION OF CAE-GENERATED SIMULATED CRASH PULSES USING
FROM CAE-GENERATED SIMULATED CRASH PULSE DATA AN HMM I NDUCED F ROM “R EAL ” C RASH P ULSE D ATA

V. C ONCLUSION
Crash data were modeled as a 3-D time series composed of a
time stamp, an acceleration reading from the tunnel (noncrush
zone), and an acceleration reading from the crush zone. This
paper has presented a methodology for building a crash detec-
tion system using continuous-mode HMMs and has established
the proposed strategy to be a robust methodology for early
In addition to the crash pulses used for training that re-
detection of automotive crashes. A data-validating strategy
quire airbag deployment, crash pulses that do not need airbag
for establishing that the crash pulse train is a sample from a
deployment are included in each set. Six actual crash pulses
multivariate normal distribution was presented. The process of
not requiring bag deployment, and three CAE nondeployment
optimizing a real-valued HMM parameter was also presented;
pulses are added to each set, respectively. Detection time is
a left–right HMM with seven states and a sample buffer with
defined as the duration of time that elapses between the point of
6-ms crash pulse storage was shown to be a simple, yet robust,
impact and the time that the system can classify the acceleration
architecture for automotive crash detection. In our experiment,
pulse as a being a crash. Again, the collision time is set
no incidence of false-positive or false-negative detection was
to t = 0.
observed, and thus, 100% accuracy was reported for all the
These results are shown in Table I. All CAE-generated crash
pulses. It was also demonstrated that CAE pulses from FEA
and noncrash pulses are correctly classified by the HMMs
can provide valuable information in crash evaluation. Although
within 6 ms.
the number of cases examined was limited and not adequate for
Similarly, when the crash pulse model is induced using
generalization, CAE pulses may be more suitable for training
the set of pulses from CAE simulations of crashes, model
crash pulse models with the actual crash pulses used to evaluate
validation is performed with the actual crash pulses. These
the model’s performance.
results are shown in Table II. All crash and noncrash pulses
are correctly classified by the HMMs within 6 ms. It might be
noted that in some cases, classification is correctly achieved in ACKNOWLEDGMENT
1 ms. This corresponds to making a decision based on about
12–13 samples corresponding to the sample interval of 0.08 ms. The authors would like to thank Dr. C. C. Chou from Ford
However, the decision is based on the entire 6-ms buffer of Motor Company for the many valuable discussions.
which is filled with near-zero acceleration values corresponding
to the constant speed of the vehicle prior to the crash. The HMM R EFERENCES
learns to recognize this transition and correctly classifies the [1] L. Evans, “Airbag benefits, airbag costs,” presented at the Soc.
pulse correspondingly. Automotive Eng. World Congr. Exhib., Air Bags Belt Restraints, Detroit,
MI, 2004, SAE Paper 2004-01-0840.
Results shown in Tables I and II demonstrate that [2] C.-Y. Chan, “On the detection of vehicular crashes—System cha-
HMM-based models are quite effective in early and accurate racteristics and architecture,” IEEE Trans. Veh. Technol., vol. 51, no. 1,
classification of crash pulses. Furthermore, the HMM trained by pp. 180–193, Jan. 2002.
[3] R. Barnard and M. Riesner, “Vehicular air-bag control based on energy,
CAE pulses outperforms the HMM trained by true pulses when momentum, and semimetric spaces,” IEEE Trans. Veh. Technol., vol. 49,
the average detection time is used as a measure of performance. no. 5, pp. 1641–1649, Sep. 2000.
Finally, for a given pulse, regardless of whether it is a true pulse [4] G. Singh and H. Song, “Intelligent algorithms for early detection of
automotive crashes,” presented at the Soc. Automotive Eng. World
or a pulse generated from CAE models, there were no false Congr. Exhib., Air Bags (Part A and B), Detroit, MI, 2002, SAE Paper
negatives. 2002-01-0190.
orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap
 1128 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58, NO. 3, MARCH 2009

[5] W. Yin, J. Zhang, and S. Huang, “Algorithm for airbag actuation [22] M. Huang, Vehicle Crash Mechanics, 1st ed. Boca Raton, FL: CRC,
based on artificial neural network,” J. Tsinghua Univ., vol. 39, no. 11, Jun. 19, 2002, ch. 1.
pp. 73–75, 1999. [23] T. Poggio, V. Torre, and C. Koch, “Computational vision and regulariza-
[6] S. Fine, Y. Singer, and N. Tishby, “The hierarchical hidden tion theory,” Nature, vol. 317, no. 26, pp. 314–319, Sep. 1985.
Markov model: Analysis and applications,” Mach. Learn., vol. 32, no. 1, [24] F.-J. Stützler, C. C. Chou, J. Le, and P. Chen, “Development of CAE-based
pp. 41–62, Jul. 1998. crash sensing algorithm and system calibration,” presented at the Soc.
[7] T. Ivanova, V. Mottle, and I. Muchnik, “Estimating the parameters of Automotive Eng. World Congr. Exhib., Airbags Safety Test Methodology,
hidden Markov models of noise like signals with abruptly changing prob- Detroit, MI, 2003, SAE Paper 2004-01-0509.
abilistic properties,” Autom. Remote Control, vol. 55, no. 9, p. 1299, 1994. [25] J. Stevens, Applied Multivariate Statistics for the Social Sciences, 2nd ed.
[8] J. Yang, Y. Xu, and C. Chen, “Human action learning via hidden Markov Mahwah, NJ: Lawrence Erlbaum, 1992, pp. 245–251.
model,” IEEE Trans. Syst., Man, Cybern. A, Syst., Humans, vol. 27, no. 1, [26] M. S. Srivastava and E. M. Carter, An Introduction to Applied Multivariate
pp. 34–44, Jan. 1997. Statistics. New York: Elsevier, 1983, pp. 25–35.
[9] J. Yang, Y. Xu, and C. Chen, “Hidden Markov model approach to skill [27] J. Ziv and N. Merhav, “Estimating the number of states of a finite-state
learning and its application to telerobotics,” IEEE Trans. Robot. Autom., source,” IEEE Trans. Inf. Theory, vol. 38, no. 1, pp. 61–65, Jan. 1992.
vol. 10, no. 5, pp. 621–631, Oct. 1994. [28] C.-C. Liu and P. Narayan, “Order estimation and sequential universal data
[10] L. R. Rabiner, “A tutorial on hidden Markov models and selected appli- compression of a hidden Markov source by the method of mixtures,” IEEE
cations in speech recognition,” Proc. IEEE, vol. 77, no. 2, pp. 257–286, Trans. Inf. Theory, vol. 40, no. 4, pp. 1167–1180, Jul. 1994.
Feb. 1989.
[11] M. Brown, J. Foote, G. Jone, K. Spärck Jones, and S. Young, “Open-
vocabulary speech indexing for voice and video mail retrieval,” in Proc.
ACM-Multimedia Conf., 1996, pp. 307–316. Gautam B. Singh (SM’07) received the B.Tech.
[12] E. Mittendorf and P. Schauble, “Document and passage retrieval based degree in electrical engineering from the Indian In-
on hidden Markov models,” in Proc. 17th ACM-SIGIR Int. Conf. Inf. stitute of Technology, Kanpur, India, in 1983 and the
Retrieval, 1994, pp. 318–327. Ph.D. degree in computer engineering from Wayne
[13] S. Oh, J. Ha, and J. Kim, “Context dependent search in intercon- State University, Detroit, MI, in 1993.
nected hidden Markov model for unconstrained handwriting recognition,” He is currently an Associate Professor of computer
Pattern Recognit., vol. 28, no. 11, pp. 1693–1704, Nov. 1995. science and engineering with Oakland University,
[14] R. Khattree and D. N. Naik, Applied Multivariate Statistics With SAS Rochester, MI. Prior to joining Oakland University,
Software. Cary, NC: SAS Inst. Inc., 1995, pp. 9–12. he was a Senior Scientist with the National Center for
[15] J. Chen and A. Kundu, “Rotation and gray scale transform invari- Genome Resource, Santa Fe, NM, and held a faulty
ant texture identification using wavelet decomposition and hidden position with Wayne State University. He conducts
Markov model,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 16, no. 2, research in machine learning and data mining and consults with Chrysler
pp. 208–214, Feb. 1994. Corporation.
[16] V. Krishnamurthy, “Hidden Markov model signal processing in presence Dr. Singh is a member of Tau Beta Pi and the Association for Computing
of unknown deterministic interferences,” IEEE Trans. Autom. Control, Machinery.
vol. 38, no. 1, pp. 146–152, Jan. 1993.
[17] A. Krogh, M. Brown, S. Mian, K. Sjolander, and D. Haussler, “Hidden
Markov models in computational biology—Applications to protein mod-
eling,” J. Mol. Biol., vol. 235, no. 5, pp. 1501–1531, Feb. 1994. Haiping Song received the B.S., M.S., and Ph.D.
[18] A. Krogh, M. Brown, S. Mian, and D. Haussler, “A hidden Markov degrees in electrical engineering from East China
model that finds genes in E. coli DNA,” Nucleic Acids Res., vol. 22, Normal University, Shanghai, China, in 1992, 1995,
pp. 4768–4778, 1994. and 1998, respectively, and the M.S. degree in com-
[19] P. Baldi, Y. Chauvin, T. Hunkapiller, and M. McClure, “Hidden Markov puter science from Oakland University, Rochester,
models of biological primary sequence information,” Proc. Nat. Acad. Sci. MI, in 2002. She is currently working toward the
USA, vol. 91, no. 3, pp. 1059–1063, Feb. 1994. Ph.D. degree with the Department of Computer Sci-
[20] J. Bilmes, “What HMMs can do,” Univ. Washington, Seattle, WA, Tech. ence and Engineering, Oakland University.
Rep. UWEETR-2002-0003, Jan. 2003. Her research interests include using machine
[21] U. Glavitsch and P. Schauble, “A system for retrieving speech learning and data-mining techniques and applica-
documents,” in Proc. ACM-SIGIR 15th. Int. Conf. Inf. Retrieval, 1992, tions of intelligent systems technology to vehicle
pp. 168–176. safety and in airbag system design.

orized licensed use limited to: AMRITA VISHWA VIDYAPEETHAM AMRITA SCHOOL OF ENGINEERING. Downloaded on October 09,2024 at 09:32:50 UTC from IEEE Xplore. Restrictions ap