Performance Analysis and Comparison for High Maneuver Target Track Based on Different Jerk Models

The Jerk model is widely used for the track of the maneuvering targets. Different Jerk model has its own state expression and is suitable to different track situation. In this paper, four Jerk models commonly used in the maneuvering target track are advanced. The performances of different Jerk models for target track with the state variables and the characters are compared. The corresponding limit conditions in the practical applications are also analyzed. Besides, the filter track is designed with UKF algorithm based on the four different models for the high-maneuvering target. The simplified dynamic model is used to gain the standard trajectory with Runge-Kutta numerical integration method. The mathematical simulations show that Jerk model with self-adaptive noise variance has the best robustness while other models may diverge when the initial error is much larger. If the process noise level is much lower, the track accuracy for four Jerk models is similar and stationary in the steady track situation, but it will be descended greatly in the much highly maneuvering situation.


Introduction
Target tracking refers to estimating the motion parameters, such as the position and the velocity for a target through the noise-containing measurement data acquired by the measurement device in real time.Because of the uncertainty of the maneuvering target motion, the measurement process uncertainty, and the difficulty of estimating the nonlinear system, the maneuvering target tracking has always been a research focus for more than half a century and there are a lot of proposed algorithms.To sum up, the track algorithms study has been focused on two parts, i.e., maneuver target modeling and nonlinear filter design [1,2].
For the nonlinear filter design, there are three main methods, including Extended Kalman filter (EKF), Unscented Kalman filter (UKF), and Particle filter (PF) [3,4].Many articles have sufficiently studied the three algorithms in the nonlinear fitness, filter accuracy, filter stability, computational complexity, and other aspects [5,6].The conclusion is that UKF has the best performance in much practical applications [7].
For the maneuver target modeling, there are lots of motion model describing the maneuver process of the target.The accuracy of the target modeling directly affects the tracking performance of maneuvering target, detection of the target [8,9], and the fault diagnosis of the target [10,11].According to their modeling state dimensions, they can be divided into second-order model, three-order model and four-order model.Two-order models contain CV (Constant Velocity) model and CT (Constant Turn) model; three-order models contain CA (Constant Acceleration) model, Singer model, CS (Current Statistic) model, semi-Markov model, and so on [12,13].Four-order models contain Jerk (Jerk denotes the rate of the acceleration change) model and its corresponding improved version.The higher the order of maneuvering model is, the higher the order of the target is described.Jerk model extends the target maneuvering form via estimating the acceleration changing rate in real time.
Theoretically, it can be applied on the highly maneuvering target tracking much better [14,15].Therefore, in this paper, we mainly focused on the maneuvering target track with Jerk model and UKF filter algorithm.
The references related to Jerk model has shown the good simulation results.However, due to the simple and the special simulation background, the trajectory is different from the true target.Therefore, the results cannot be convincing.This paper summarizes the various Jerk models in the references.Firstly, the advantages and disadvantages are analyzed in theory.Then every model is applied on the near-space highspeed maneuvering target, and the simulation results are compared.The track trajectory is generated through the integral of the simplified dynamic equation, which is close to the true target.Therefore, it is convictive to some extent.

Description for Different Jerk Model
Jerk denotes the acceleration changing rate.Jerk model is a model which describes the target Jerk mathematically.The models in the references mainly contain the following different kinds.

SJ Model.
The earliest Jerk model is proposed by Mehrota, etc. [16].They used Singer model for reference and modeled the Jerk model of the target as a zero-mean and first-order time related process.To distinguish other Jerk models, it is marked as SJ (Singer Jerk) model.Taking a one-dimension linear motion as an example, SJ model is expressed as where () = ...

CSJ Model.
Qiao [15] used the analyzing method which is also applied on Singer model tracking accuracy in [17] to analyze the SJ model.
where  CSJ () denotes the nonzero mean of Jerk. CSJ () denotes the covariance of the zero-mean colored Jerk noise, and () denotes zero-mean Gaussian white noise.Other parameters are defined as SJ model.
The model should set  and  2  previously before the practical application.Therefore, Pan proposed a novel CSJ algorithm to describe the probability density of Jerk according to truncated normal distribution and construct the connection between  2  and current Jerk estimation.It is In this way, the probable extreme Jerk can be predefined as  max .That is the covariance of the state noise can be selfadapted according to the Jerk estimation during the filter process to adapt to different maneuvering situations.The model is marked as MCSJ (Modified CSJ) model.

2.
3. J Model.The Jerk models above cannot avoid the problem that the Jerk frequency  should be predefined.However,  cannot be directly measured and it is constantly changing in the target practical motion process.For this reason, Luo [7] considered  in SJ model as an estimated parameter and took it as the extension variable.Therefore, it can be estimated in real time during the filter process. is modeled as nonzero-mean Gaussian white noise and the derivative of it is zero mean Gaussian white noise which can be considered as where () denotes the zero-mean input noise, whose variance is  2  .This is called  J (Alpha Jerk) model. J model can estimate  in real time, but  2  and  2  need to be predefined carefully.If the values of them are not proper, the estimation accuracy of  will be decreased severely and even cause the divergence of the filter.
Take the one-dimension motion as an example.(Threedimension situation has the same principle as the onedimension motion.)The state variable, state equation, and the characteristic of the four models are shown in Table 1.The specific form of the matrices in Table 1 can be found in the related references [14,16].

Trajectory State Equation.
Tracking target is the nearspace and high-speed target whose true trajectory comes from the integral of the dynamic equation.To simplify the dynamic equation of the target, the earth is assumed as an irrational ball and the influence on the center of mass motion of the target of the controlling force is neglected.The target has nonliteral movement and the simple dynamic equation in the launching frame is expressed as where  denotes the atmosphere density which is used as American 1976 standard atmosphere model. denotes the resultant velocity;  denotes the aerodynamic reference area    need to be predefined. needs to be estimated in real time.The target parameter  and  and the aerodynamic parameters   and   use the public values in American high-performance general air vehicle (CAV-H).The longitude, the latitude, the height, and the launching angle are 0 ∘ , 0 ∘ , 0m, and 90 ∘ .The initial longitude, the latitude, and the height of the target are 0 ∘ , 0 ∘ , and 30 km, respectively.The initial Mach number is 10.The local velocity angle is -1 ∘ , and the velocity drift angle is 0 ∘ .The controlled quantity is the attacking angle  and is constant with 15 ∘ .The integral method of the trajectory is fourth-order Runge-Kutta numerical integration.The height and the voyage are shown in Figure 1.

Measurement Setting Condition.
The measurement device is a pulse radar and the measurement elements are the distance R, the azimuth angle A, the pitch angle E, and the radial velocity V.The measure error standard deviations of the distance, the angle, and the velocity are 10m, 0.5mrad, and 1m/s, respectively.The longitude, the latitude, and the height of the radar station are 14 ∘ E, 1 ∘ N, and 0m, respectively.The sampling period is 0.05s and the tracking period is 70∼790s.To compare the tracking performance of the models (including the tracking accuracy, the filter robustness, and the computation complexity), the simulation results cannot be compared as the general references setting a parameter.Mehrotra proposed that "Jerk" frequency  has a good robustness in [7] via the simulation.That is, the filter performance can maintain stable when  changes in a large range.Therefore, the paper focuses on the standard deviation of Jerk and the robustness of the filter initial value error.The initial position and velocity in filter are equal to the true value adding the random error.The standard deviations of  3, when the filter initial value is large, only the self-adaption state noise of MCSJ model can maintain the convergence and the initial value robustness is much better.But the other three models will present filter divergence performance during the 100 Monte Carlo steps and the convergence is influenced by the initial value which is not proper for the practical tracking application.On the other hand, because the only difference between P2 and P3 is   .The parameter robustness of the four models can be compared.When comparing Figures 3 and 4, when   is small, the tracking accuracy of the four models in the steady period of the target motion can be improved.However, the filter error has a large jump when the target has a severe maneuvering motion.The computation complexity ratio of SJ, CSJ, MCSJ, and J is 8:8:9:12.

Conclusion
The paper summarizes the proposed Jerk models which are applied on the high-maneuvering target tracking.The tracking accuracy, the filter robustness, and the computation complexity of the models are compared in the theoretical analysis and the tracking simulation.The tracking simulation trajectory is created by the integral of the dynamic equation which is similar to the true process and is pretty persuasive.The simulation results demonstrate that the self-adaptive noise variance method has the best tracking performance.However, in the highly maneuvering period, all models have the error jumping problem.Controlling the filter jumping error is the key to improve the near-space and high-speed target tracking performance in the future.
)Tracking step Jerk signal shows steady-state deterministic error. 2  and  2
(a) RMSE of position (b) RMSE of velocity

Figure 4 :
Figure 4: Filtering RMSE for four models in P3 parameter.

Table 1 :
Description for four different Jerk models.

Table 2 :
Filter parameters.position and the velocity are  0 and  V0 , respectively.The acceleration and the Jerk are 0.  J model needs to set the initial value of  and the standard deviation   of the input noise ().MCSJ model needs to set the extreme Jerk value  max .The parameter sets of the three filters are expressed as P1, P2, and P3, respectively.The specific parameter values are shown in Table2.The Monte Carlo simulation number is 100.3.3.Simulation Results.The filter root mean square error (RMSE) of the position and the velocity under the parameter sets is shown in Figures2, 3, and 4, respectively.Because of the different values of P1 and P2, the initial value robustness of the four models can be compared.Comparing Figures2 and the