2018 International Computers, Signals and Systems Conference (ICOMSSC)

A Method of Faint Trajectory Feature Detection

Based on Sliding Window

Wang Baojun, Liu Heng, Xu Fang Wang Baojun, Wang Lidong

Huayin Ordnance Test Center Army Engineering University Shijiazhuang Campus
Shan Xi, China Hebei, China
1986lh1986@163.com

Abstract—Aiming at the difficulty of small changes in speed

after flameout and the inability of traditional methods to II. CHANGES IN THE FAINT CHARACTERISTICS OF THE
identify faint feature changes of extinguishing points, a STATIONARY PROCESS
detection method of sequence feature changes based on sliding The so-called "faint characteristic change of the
windows is proposed, a change detection model is established,
stationary process" refers to a characteristic change that
and detection parameters and criteria are defined. The
verification of simulation data shows that this method can
occurs abruptly within a short time interval, either
effectively detect the changes of the sequence features and instantaneously or relative to the sampling interval. For a
identify the changes of the faint features in the stationary stationary process, data mining technology based on
process. This method is feasible in the detection of flameout sequence features has been widely used in the field of
features. similarity search of data features, sequence pattern mining,
and cycle analysis. The purpose of most of the sequence
Keywords—sliding window, feature detection feature data mining is to find frequently occurring patterns
and expect to find some rules. The feature change data is
I. INTRODUCTION usually ignored as noise. The changes in the test data
sequence are data objects that are significantly different or
The ballistic measurement radar is very sensitive to inconsistent with the motion or characteristic rules before
radial velocity changes. The velocity of the projectile and after the feature of the tested target feature, and may
changes at the 1 m/s level, and the radar can sense, measure have more valuable and meaningful action feature events.
and identify feature points such as open positions and Therefore, it is a very meaningful job to discover the
flameouts through data processing. The ending process of outliers in the test data sequence and analyze the impact of
the bottom flameout device is a “gradation process”. When the characteristic events that are implicated and represented
the bottom row is working, the air resistance of the behind it. In this paper, feature points detected by sliding
projectile is small. At the end of the bottom row device, the window based sequence feature changes are used. The
air resistance is increased, but the end of the bottom algorithm predicts future data values based on the sliding
discharge column burns is a slow process, which is different window method and determines the features by setting the
from the “deep change process” in which the projectile is difference between predicted values and actual
opened and the engine is extinguished, and it does not show measurements. This method can effectively detect the
any significant change in the projectile movement speed. feature changes existing in the data sequence, so as to
That is, the projectile is a smooth movement and is a achieve the goal of improving the event recognition.
relatively stable process. Solving the flameout time point
can be attributed to the faint change feature detection Feature change detection can be described formally as
problem. The extinguishing point is detected by analyzing follows: Given a set of n data points or objects and expected
and processing the change of acceleration, but due to the feature points (at k), it is found to be significantly abnormal,
noise signal and the error of the test equipment, the isolated, or inconsistent with the remaining data. Feature
extinguishing point cannot be detected by the traditional detection requires solving two main problems: defining
method. The difficulties faced by faint feature change what kind of data is a feature point in a given data set;
detection mainly include: small amount of speed change finding an effective method to detect such feature data.
after flameout, traditional techniques such as recognizing According to the different representations of feature data in
opening positions and engine shutdown points through the sequence, the sequence can be divided into three types:
speed change methods, which cannot be applied to sequence features, point features, and pattern features. At
extinguishing point tests, and random errors of acceleration present, the sequence feature detection methods mainly
data obtained by differentially speeds are large, so new include the three methods.
detection techniques need to be studied. (1) Window-based method. The sequence is divided into
In this paper, the difficult of the detection of faint a number of fixed-size subsequences (window), and the
trajectory feature points in the "stationary process" of the feature points are located in each subsequence. The basis of
end-fired flameout point is studied. The detection method of this method is that feature points in the sequence may be
the sequence feature change based on sliding window is feature points in one or more sub-sequences.
proposed. The detection model is established and the sliding (2) Distance-based method. This method uses quadratic
window is designed. Experimental verification shows that regression model to divide the inequalities of the sequence.
the method is effective.

978-1-5386-6751-4/18/$31.00 ©2018 IEEE 750 Dalian, China

September 28-30, 2018
Authorized licensed use limited to: Ming Chi University of Technology. Downloaded on May 21,2021 at 13:58:28 UTC from IEEE Xplore. Restrictions apply.
It calculates the feature score of the subsequence based on the confidence interval vi '±τ corresponding to the value of
the Dynamic Time Warping (DTW) distance, and selects the
largest feature score of k values to determine the feature point d i .and the threshold τ can be calculated from the
point. window width k and the confidence p of the prediction;
(3) Density-based method. This method does not simply when the test obtains the actual value vi , the relationship
use Yes or No to determine whether a point is an outlier, but
uses a weight to evaluate its degree of outliers. It is local, between it and the predicted value vi ' is compared. If vi is
meaning that the degree depends on the object's isolation
outside the interval of vi '±τ , then d i is determined as the
relative to its neighborhood. This method can detect both
global outliers and local outliers. characteristic change point, otherwise normal. The sliding
window is moved backwards, and the next point is judged
The window-based method depends on the window until all the test results are judged. The overall flow of the
width. If the window width is too large or too small, the algorithm is shown in Fig. 1.
accuracy of the detection result may be affected. The
test sequence to be
distance-based method has high time complexity and the
tested
efficiency cannot be guaranteed. The density-based method
is reasonable, but it depends on the rationality of the weight define the nearest neighbor
setting through prior knowledge. window prediction test value to
be scheduled and the

the measured value

of the scheduled
determination threshold
III. SEQUENCE FEATURE CHANGE DETECTION TECHNOLOGY compare measured
BASED ON SLIDING WINDOW

point
value and predicted
value
A. Model Establishment
The test data is time-varying data. The data element may Is the difference yes
be time series of the target's speed, acceleration or gray greater than
the threshold?
value of the video image, imaging size, and the like. A data
sequence is an ordered collection of elements consisting of
test values and time, and is written as Equation (1). mark as an ordinary point

D=
n d=
1 ( v1 , t1 ) , d=2 ( v2 , t2 ) , ⋅⋅⋅, d=n ( vn , tn ) (1) replace the nearest neighbor window of
the predetermined point value, with the
measured value to participate in the next
test
In which d i = (vi , ti ) represents the corresponding test Fig. 1. Flow of Feature Detection Algorithm Based on Sliding Window
value vi at the time ti , and the recording time is strictly
increased C. Sliding window design
(k )
η i ( k ) = {d i − 2 k , d i − 2 k +1 ,  , d i −1 } represents the k Define the sliding neighbor node window η i for the

nearest-neighbor window of the point d i ,and the set of test point d i to be detected in the feature sequence D . In order n

to reduce the computational complexity of the algorithm,

values is {vi −2 k , vi −2 k +1 ,, vi −1 }. If the difference between
only the k nearest neighbor node of the point d i to be
the actual test value of d i and the predicted value of the k
detected is used as the input parameter of the prediction
nearest neighbor window model exceeds a certain threshold, model for calculation. Therefore, it is especially important
the point is determined as the feature change point, (k )
to define a suitable nearest neighbor window η i .
according to the model idea.
The nearest neighbor window k and the threshold The neighbor node window can be divided into two
types: unilateral and bilateral. The bilateral neighbor node
τ become the basis for determining whether d i is a feature window is suitable for the case where both the pre- and
change point. These parameters can be adjusted according to subsequent window data of the node to be detected are
different test results and action characteristics to achieve the known; the single-side neighbor node window only contains
optimal detection result. the node's predecessor data. Based on the predictive feature
detection algorithm, only the left neighbor window of the
B. Algorithm Description node to be predicted (abnormal data and outlier data is
The core idea of the sequence feature detection method excluded) is used as the input parameter of the algorithm;
based on the sliding window is: according to different other algorithms can select the bilateral data detection. The
application requirements, define the nearest neighbor one-sided neighbor window is given by Equation (2).
(k )
window η i of the point d i , establish the prediction model
and use the set of η i
(k )
test values as input parameters to η i ( k ) = {d i −2 k , d i −2 k +1 ,, d i −1 } (2)

predict the observation value vi ' of the point d i ; calculate Where 2k is the length of the sliding window of the

751

Authorized licensed use limited to: Ming Chi University of Technology. Downloaded on May 21,2021 at 13:58:28 UTC from IEEE Xplore. Restrictions apply.
node d i , from {i − 2k } to {i − 1}. the confidence threshold τ and the confidence interval
PCI can be calculated as Equation (5).
D. Prediction Model Establishment
The core of the feature change detection algorithm is to = t a / 2, 2 k −1 × s 1 + 1 /(2k ) , PCI = v τ +1 ± τ (5)
establish a predictive model and use the sliding window

t
η i (k ) as an input parameter to predict the test values of
Where, v t +1 is the predicted value of the prediction
subsequent nodes. The predictive model predicts
observations at a future observation point in the time series model calculated according to the sliding adjacent window
through a set of specific input values—the sliding of the point to be tested; t a / 2, 2 k −1 is the percentile with the
(k )
window η i . The predictive model can reduce the number degree of freedom 2k − 1 ; s is the standard deviation of the
that cannot be predicted due to insufficient data. model residual; k is the size of the sliding window. If the
actual measured value falls within the confidence interval of
Using the sliding window η i = {d1 , d 2 ,  , di} as the
(k )
the prediction value, the undetermined point is marked as a
input parameter of the feature sequence prediction model to normal point; otherwise, it is marked as a characteristic
change point and the point is further analyzed and processed.
predict the observed value d i +1 , the budget method can be the threshold PCI for determining whether a node is a
expressed as Equation (3). feature change point can adjust the width of the window
effectively, so as to avoid false detection probability events
(k ) caused by threshold selection.
d i +1 = M (η i ) (3)
F. Preferences
(k )
Among them M (η i ) is the prediction model. The In order to detect the feature change point in the test
premise of this algorithm is that the test value at the moment sequence, the feature point prediction method based on the
sliding neighbor window needs to calculate a reasonable
k + 1 can be described by k finite precursor measurement threshold τ of the test point. In order to improve the
value, that is, its implicit assumption feature sequence is a k efficiency of the detection method and improve the
Markov process. By comparing the detection efficiency of performance of the algorithm, a reasonable choice of the
the simple change model, the nearest-neighbor prediction
model, the single-layer linear network prediction model and two parameters k and p involved in the algorithm is the
the multi-layer perceptual prediction model on different data key to improving the anomaly detection algorithm.
sets, the single-layer linear network prediction model can Therefore, use the following method to select the
obtain the best detection efficiency. A single-layer linear appropriate parameter value.
network is used as the prediction model, and the
observations at the moment t are assumed to be linear a) The width of the window k . k determines the
combinations of their advancing neighboring windows, number of sliding abutment points participating in the
shown in Equation (4). prediction. The larger the value is, the more sliding
abutment points are involved in the calculation and the
computational complexity increases accordingly. To select
2k 2k the optimal sliding window width, the value k range is 3-15,
v = (∑ ( wt −i vt −i )) /(∑ wt −i ) (4) the increment is 1, that is k = {3，
4， 15} .
，
i =1 i =1
b) Confidence coefficient p . p defines the expected
Where wt − 2 k , wt − 2 k +1 , , wt −1 represents the frequency of the measured value within the confidence
neighboring window node weight vector, the closer the interval of the value. The greater the confidence coefficient,
distance between the nodes, the greater the weight. To the larger the confidence interval of the prediction value,
simplify the calculation, the weight vector and the range p of the value is 80% to 100% in increments
〈 wt − 2 k , wt − 2 k +1 , , wt −1 〉 is generally assigned to of 2%. which is p = {80%， 82%， ， 100%}.
〈1, 2, , 2k 〉 . The goal of feature detection is to detect as many feature
points as possible in the sequence and reduce the probability
E. Feature Point Determination of false detections. Therefore, the criterion of the parameter
The adjacent window of the point to be detected is used pair ( k , p ) is to maximize the ratio between the detection
as an input parameter, and the predicted value of the point to accuracy and the false detection rate.
be tested and the confidence interval based on the predicted
value are calculated according to the prediction model. The IV. SIMULATION VERIFICATION
prediction confidence interval gives a possible range of
values for the observation to be detected. The confidence A. Simulation Example 1
factor p = 100(1 − α ) indicates the expected frequency of
In order to verify the correctness of the above algorithm,
the actual measurement within the confidence interval of the missile simulation data is generated for detection, and
value. If it is assumed that the model residuals have a zero- simulation data is graphically displayed. The speed curve is
mean Gaussian distribution, the confidence level p % with shown in Fig. 2, where the vertical axis is the speed and the

752

Authorized licensed use limited to: Ming Chi University of Technology. Downloaded on May 21,2021 at 13:58:28 UTC from IEEE Xplore. Restrictions apply.
horizontal axis is the flight time. From the curve, the missile TABLE I. SIMULATION OF BALLISTIC DATA FOR A CERTAIN TYPE OF
BOTTOM BOMB
flight is stable. Use the model and method established in the
previous section to test and verify the characteristics of the Distance Azimuth Elevation
Radial velocity Acceleration
missile's flight-the engine off point. According to the missile value value
trajectory, when the engine stalls, the missile acceleration 16437.3 357.734 20.2282 449.13 -10.5
becomes zero and the speed reaches a maximum. The curve 16446.7 357.737 20.2208 448.92 -9.5
reflects this trend. At 172 frames corresponding to time
16455.5 357.74 20.2135 448.73 -9.5
6.76s, the missile speed reaches a maximum 242m/s.
16464.3 357.742 20.2061 448.54 -10
From the simulation results, the characteristic data points
16473.6 357.745 20.1987 448.34 -10
are easily deviated from their neighbor nodes, namely
"feature change points." Fig. 3 shows the test results. 16482.4 357.748 20.1912 448.14 -10
16491.8 357.751 20.1839 447.94 -10
Fig. 3 shows the simulation data, prediction values and
confidence interval in the given speed series data set under 16500.6 357.753 20.1764 447.74 -11
the condition of the sliding window width k = 6, the 16509.4 357.756 20.169 447.52 -10.5
confidence p = 95%. From Fig. 3, it can be seen that most of 16518.2 357.759 20.1615 447.31 -11.5
the ballistic data are very close to the predicted value. The
16527.5 357.762 20.1541 447.08 -11.5
actual value of the feature point exceeds the confidence
interval, which is not the same as the predicted value, is 16535.7 357.764 20.1465 446.85 -11.5
detected. The detected engine flameout time is 6.75s, with 16544.5 357.767 20.1391 446.62 -12
the corresponding speed 241.6m/s, which 6.76s and 242m/s
600
in theory. The small errors of the data indicate that this
method can detect the mutation feature points. 500

400
B. Simulation Example 2
300
Compared with the detection of the characteristics of the
misfire point of a missile engine, the characteristics of the 200
flameout point of the bottom bomb are much fainter. This
100
simulates a certain type of bottom trajectory data. Table 1 is
a partial data list. The extinguishing point time of the 0
simulation data is approximately 27.7s.
111
133
155
177
199
221
243
265
287
309
331
353
375
397
419
441
463
485
507
529
1
23
45
67
89

In order to observe the trend more clearly, the data is Fig. 4. Speed simulation curve of a certain bottom bombing zone
plotted in the form of a curve, as shown in Fig. 4 and 5 . It is
very difficult to observe and detect the extinguishing point
only from the speed of the projectile. The speed curve is
very smooth and is a typical “smooth process”. The marked
part in Fig. 4 is the extinguishing process. This can also be
seen from data list 1.
300

250

200

150

acceleration 3-point 5-point 10-point

100

50 Fig. 5. Acceleration simulation data curve of a certain bottom bomb

0
0 100 200 300 400 500 600 700 800

Fig. 2. Simulation of a missile's speed and time

Feature point confidence interval

prediction value
measured data

feature point

confidence
acceleration prediction value
interval

Fig. 6. The acceleration detection result of a certain type of bottom

displacement under (k, p) = (6, 95%)
Fig. 3. speed characteristic change detection result under (k, p) = (6, 95%)

753

Authorized licensed use limited to: Ming Chi University of Technology. Downloaded on May 21,2021 at 13:58:28 UTC from IEEE Xplore. Restrictions apply.
 In order to show the characteristic change of V. CONCLUSION
extinguishing point, the speed data is differentiated to get Aiming at the difficulty of the small change of speed
the acceleration of the projectile, as shown in Fig. 5. The after flameout, the traditional method of identifying the
acceleration of the extinguishing process is mainly analyzed. extinguishing point, the faint change of the video image, this
Acceleration data obtained by directly differentiating the paper studies the sequence feature change detection
speed contains many high-frequency random error technology based on the sliding window and establishes the
components, and the maximum error value reaches 6m/s. sequence feature change detection model, clear detection
This data sequence is not conducive to the detection using parameters and criteria. Combined with simulation test
the “faint feature detection” method established in this paper. data for verification calculations, it shows that this method
so the data sequence is preprocessed. Using multi-point can indeed effectively detect changes in the data sequence
smoothing to eliminate the influence of high-frequency and identify changes in the faint characteristics of the
random errors brought about by the differential, 3, 5, and 10 process. This method is feasible in the detection of flameout
points are used to smooth the acceleration data sequence. By features.
comparison, the 10-point smoothing has the most obvious
suppression effect on random errors and can reach 0.3m/s.
The 5-point smoothing can reach 0.5m/s, and the 3-point REFERENCES
smoothing is about 0.8m/s; however, the 10-point [1] Li Zhengdong, He Wuliang. An improved data fusion algorithm.
smoothing result has brought error to the time information. Optoelectronic Engineering. 2001.
That is to say, the time corresponding to the “feature point” [2] Li Yang, Zhang Jing. Track Association Algorithm Based on
will be earlier than the actual time, increasing the time error. Adaptive Sliding Window OSAR. Chinese Journal of Electronics.
2016(2).
Consider comprehensively choosing the 5-point method to
[3] Luo Qian. Clutter suppression and faint moving target detection
smooth the data. algorithm for early warning radar. Journal of Beijing University of
Fig. 6 shows the measured data, prediction values and Technology. 2012. 3.
confidence interval in the given speed series data set under [4] Sun Lihong, Wang Jun. Improved TBD algorithm for radar faint
target detection[J]. Radar Science and Technology. 2007.5(4): 292-
the condition of the sliding window width k =5, the 296.
confidence p = 95%. From Fig. 6, it can be seen that most of [5] Li Tao, Zeng Haibin, Tian Zhongcheng et al. A new algorithm for
the test values are very close to their respective predicted pre-detection tracking of high-velocity faint target radar[J]. Journal of
values. The actual value of the feature points exceeds the Data Acquisition and Processing. 2008.23(6): 641-645.
confidence interval and is detected. The detected flameout [6] Zhao Mengmeng, Zhang Qun et al. A sliding window trajectory
time point is 27.75s, with the corresponding acceleration - tracking algorithm for space group target resolution. Journal of
9.8m/s2, and the theoretical ballistic data is 27.7s. Aerospace. 2015(10).

754

Authorized licensed use limited to: Ming Chi University of Technology. Downloaded on May 21,2021 at 13:58:28 UTC from IEEE Xplore. Restrictions apply.