KR102353619B1 - Apparatus and method for radar target tracking - Google Patents

Abstract

The present invention relates to target tracking of a radar, and more particularly, to setting coefficients of a Kalman filter used for target tracking and setting a search radius. In the case of the existing target tracking, the present invention automatically calculates and operates the optimal value according to the operation and sea conditions without user intervention, compared to the user manually setting the coefficients from time to time according to the operation and sea conditions during radar operation. Adaptive target tracking. That is, it has the characteristic of adaptively adjusting the step size and the search radius of the Kalman filter according to the operation of the target, input noise, etc. generated during operation.

Description

Apparatus and method for radar target tracking {APPARATUS AND METHOD FOR RADAR TARGET TRACKING}

The present invention relates to radar target tracking (Target Tracking, for short), and more particularly, to setting coefficients of a Kalman filter used for target tracking and setting a search radius.

for k = 1, 2, 3, …
tracking error:

filtering:

prediction:

end
여기서, k는 프레임 인덱스, T는 프레임 시간 주기, z_k는 BLOB 분석 결과 측정된 타겟의 위치 벡터, p_k｜k(p_k+1｜k)는 k-시점에서 필터링(예측)된 타겟의 위치 벡터, v_k｜k(v_k+1｜k)는 k-시점에서 필터링(예측)된 타겟의 속도 벡터, α(β)는 위치(속도) 스텝 사이즈를 나타낸다. 여기서 벡터란 말은 상기 위치, 속도 등이 2차원상의 x, y 좌표 성분을 갖는 벡터라는 뜻이다.
상기 수학식들의 물리적 의미를 살펴보면 다음과 같다. 상기 알파-베타 필터는 기본적으로 타겟이 등속운동을 하고 있다고 가정하고 있다. (수학식 1)에서는 타겟의 위치, 속도를 적당한 값으로 초기화시킨다. 초기에는 아직 예측된 값이 없으므로 예를 들어 P₀은 측정된 타겟 위치 z_k로, V₀은 0으로 둘 수 있다. (수학식 2)에서는 측정된 타겟 위치 벡터 z_k와 이전 프레임에서 예측된 타겟의 위치 벡터 p_k｜k-1간의 차이, 즉 위치 추정 오차(또는 타겟 트랙킹 위치 오차)를 계산한다. (수학식 3)에서는 위치 추정 오차에 위치 스텝 사이즈를 곱한 만큼을 이전에 예측된 위치값에 더하여 필터링된 위치값을 계산한다. 비슷한 방법으로 (수학식 4)에서는 위치 추정 오차에 속도 스텝 사이즈를 곱한 만큼을 이전에 예측된 속도값에 더하여 필터링된 속도값을 계산한다. (수학식 5) 및 (수학식 6)에서는 가정된 타겟의 운동 모델(등속도운동)에 따라 위치와 속도의 차기 예측값을 계산한다. 이 과정을 매 k마다 반복함으로써 원래 타겟 트랙킹의 목적인 타겟의 위치와 속도를 추정할 수 있다.
상기 수학식에서 α, β의 역할에 대해 살펴본다. (수학식 3) 및 (수학식 4)의 필터링 과정에서 α, β가 크면 위치, 속도의 필터링 값이 크게 변화하고(즉 위치 추정 오차에 민감하게 반응), α, β가 작으면 위치, 속도의 필터링 값이 작게 변화함(즉 둔감하게 반응)을 알 수 있다. 타겟이 등속운동을 하고 있다고 가정하면, α, β를 아주 작게 잡음으로써 측정된 위치 벡터에 포함되어 있는 노이즈를 평균화하여 노이즈를 크게 줄일 수 있다. 반대로 타겟이 급선회를 한다고 가정하면, α, β를 크게 잡음으로써 타겟 위치 및 속도 추정값이 빨리 변하게 함으로써 급선회하는 타겟을 빨리 따라잡을 수 있게 할 수 있다. 전자를 정확도 향상이라고 하고, 후자를 속응성 향상이라고 부른다. 따라서 α, β를 어떻게 설정하느냐에 따라 위치 및 속도의 정확도와 속응성을 조정할 수 있다. 그런데 상기 설명으로부터 상기 정확도와 속응성은 서로 trade off 관계에 있음을 알 수 있다. 즉 한쪽이 좋으면 다른 쪽이 나빠지고, 다른 쪽이 좋아지면 이쪽이 나빠진다. 요약하자면 α, β가 크면 위치 및 속도 추정의 속응성은 향상되지만 정확도는 떨어지고, 반대로 α, β가 작으면 위치 및 속도의 속응성은 느려지지만 정확도는 향상된다. 따라서 상기 정확도와 속응성간의 상반된 추세를 고려하여 어떻게 α, β를 결정하느냐가 타겟 트랙킹의 주요 관건이다.
기존 방식에서는 α, β를(또는 이와 관련된 파라미터를) 사용자가 직접 세팅하도록 되어 있다. 따라서 속응성을 높이기 위해 α, β를 크게 잡았다면 측정된 타겟 위치가 노이지하다면 추정된 위치 및 속도값도 노이지해질 것이다. 반대로 정확도를 높이기 위해 α, β를 작게 잡았다면 자선이 등속도운동(순항)을 하다가 급선회할 경우 그 변화를 빨리 따라가지 못함으로써 속응성이 떨어질 것이다. 따라서 사용자는 그때그때의 운행 상황 및 해면 상황을 주시하면서 수동으로 α, β를(또는 이와 관련된 파라미터를) 조정해야 하는 불편함을 감수해야 한다. 뿐만아니라 사용자가 수동으로 상기 파라미터들을 결정하면 최적점을 찾기 어려울 수도 있다. 예를 들어, 자선과 타겟간의 거리가 멀어질수록 거리에 비례하여 접선방향의 노이즈가 커지는데 이런 노이즈는 사용자에게 미처 인지되지 않아 사용자가 적절히 파라미터 세팅을 못할 수 있다.
어쨌든, 일단 매 프레임마다 입력되는 측정된 타겟 위치 신호열(즉 시계열 신호)로부터 타겟의 위치 및 속도를 추정했다고 했을 때 다음으로 고려해야 할 사항이 서치 반경(Search Radius)를 결정하는 것이다. 다음 프레임에 수신되는 프레임 영상에 수많은 BLOB들이 포함되어 있을 텐데 이들 중 어느 것이 타겟인지 판단해야 하는데 한 프레임 내의 모든 BLOB에 대해 이들을 체크하는 것은 무리이다. 따라서 상기 알파-베타 필터를 통해 추정된 위치 예측치를 중심으로 일정한 서치 반경내에서 다음 프레임의 타겟을 서치하는 것이 합리적이다. 기존의 경우 이 서치 반경 역시 사용자가 수동으로 설정하도록 되어 있다. 서치 반경을 작게 잡으면 그 속에 들어있는 BLOB수가 적으므로 타겟을 찾는 시간을 줄일 수 있는 장점이 있으나 타겟을 놓칠 확률이 커진다. 반대로 서치 반경을 크게 잡으면 타겟을 놓칠 확률은 작아지지만 서치 반경에 들어있는 BLOB 수가 늘어나서 타겟을 찾는 시간이 증가하는 문제가 있다. 따라서 사용자가 그때그때 상황을 주시하면서 직접 서치 반경을 세팅해야 하는 불편함을 감수해야 한다.
어쨌든, 상기 기술한 타겟의 위치 및 속도 추정, 서치 반경 결정등을 포함하는 대표적인 기존 기술에 의한 타겟 트랙킹 과정을 (그림1)에 흐름도로 나타내었다. 이중 서치 반경 내에 서치해 보았더니 원하는 타겟이 발견되지 않은 경우에 대한 조치가 필요해서 그 과정을 나타내었다. 즉, 타겟의 크기, 해상 상황등의 이유로 타겟 BLOB가 일시적으로 사라지는 현상이 발생할 수 있는데 이 경우 타겟 트랙킹은 보통 그때의 측정 타겟 위치 신호가 없다고 판단하고 (수학식 3) 및 (수학식 4)에 의한 필터링 과정을 생략해버린다. 대신 다음 예측값을 (수학식 5) 및 (수학식 6)에 의해서만 계산한다. 몇 번 프레임을 반복해도 못찾으면 해당 타겟을 잃어버렸다고 판단하고 그 타겟에 대한 추적을 종료한다. 이를 포함하여 타겟의 설정부터 타겟 추적, 타겟 종료에 이르는 타겟 생존 기간 동안의 과정에 대한 제어는 흐름도 내의 상태 천이(State Transition) 과정을 통해 수행된다. 이외 자선 운동 보상(Compensate Own Ship Motion) 등이 있으나 이는 본 발명의 핵심적 내용과는 직접 관련이 없으므로 설명을 생략한다.Radar, especially magnetron radar, which is currently widely used, basically fires high-power magnetron pulses from a 360-degree rotating transmitting antenna and hits a nearby object (fixed object such as land or island, or a moving object such as a ship or buoy) and receives the antenna A device that displays the fixed or moving object on a screen from an echo signal captured by (usually the same transmitting and receiving antenna). From the point of view of signal processing (especially digital signal processing (DSP)), the operation can be roughly divided into two stages.
The first is a step of displaying a fixed or moving object on the screen (Image Display), and the second is an object desired or specified by the user, in particular, among the fixed or moving objects (this is called a target or target). ) is the stage of continuous tracking (Target Tracking, TT for short). There are various purposes of target tracking, but the representative purpose is to prevent collision with other ships.
In the video display stage, the main goal is to suppress noise mixed with echo signals such as additive noise and sea waves, and to enhance signal components such as fixed or moving objects to increase the possible SNR (Signal-to-Noise Ratio) to express the image clearly. , the target tracking step is a step of extracting information required by the user, such as the location and speed of the target from the image. In other words, it can be seen as a stage in which a certain level of intelligence is given to the radar. This target tracking is also called ARPA (Automatic Radar Plotting Aid).
I would like to briefly explain the process of target tracking. First, an image is received from the image display step, and a binary image called BLOB (Binary Large Object Block) is created through a CFAR (Constant False Alarm Rate) discriminator. BLOB is updated once every time the antenna rotates (usually several seconds), and an image corresponding to one rotation time is called a one-frame image. In general, the number of BLOBs in one frame image may vary greatly depending on sea conditions, and there may be 0 to hundreds or more. Given a single frame image, first, the BLOB analyzer calculates and outputs the features of the BLOB, such as the location (or center of gravity) and size of each BLOB in the frame. When the frame index is k (that is, the current frame is the kth frame), the BLOBs corresponding to fixed objects such as islands and moving objects such as ships will move every k, and the user-specified BLOB (ie, target BLOB) will also move. The primary goal of target tracking is to receive the output of the BLOB analyzer and extract the position and velocity of the target while tracking the target every k. Furthermore, the secondary goal is to use this information to calculate information for collision avoidance such as TCPA (Time to Closest Point of Approach). There can be 0, 1 or multiple BLOBs, and each BLOB (ie target) specified by the user will be tracked separately. For convenience, it is assumed that there is only one target and will be described.
A Kalman filter is usually used for target tracking (of course, other methods or similar methods are also possible). The Kalman filter was originally derived from the point of view of statistical signals and is known as an optimal filter under the environment where the system is linear and the noise is Gaussian. However, in order to design this optimal filter, there are many difficulties in practical application, such as needing to know in advance the statistical constant values of the environment (for example, process noise covariance matrix, measurement noise covariance matrix, etc.) A relatively simple alpha-beta filter (α-β filter) is widely used. Therefore, in the present invention, an alpha-beta filter will be described as an example of the conventional method. The alpha-beta filter can be expressed mathematically as follows.
initialization:

for k = 1, 2, 3, …
tracking error:

filtering:

prediction:

end
Here, k is the frame index, T is the frame time period, z _k is the position vector of the target measured as a result of BLOB analysis, and p _k|k (p _k+1|k ) is the filtered (predicted) target at the k-point. The position vector, v _k|k (v _k+1|k ) is the velocity vector of the filtered (predicted) target at the k-time point, and α(β) denotes the position (velocity) step size. Here, the term "vector" means that the position, velocity, and the like are vectors having two-dimensional x and y coordinate components.
The physical meanings of the above equations are as follows. The alpha-beta filter basically assumes that the target is moving at a constant velocity. In (Equation 1), the position and velocity of the target are initialized to appropriate values. Initially, since there is no predicted value yet, for example, P ₀ can be set to the measured target position z _k and V ₀ can be set to 0. In (Equation 2), a difference between the measured target position vector z _k and the target position vector p _k|k-1 predicted in the previous frame, that is, a position estimation error (or target tracking position error) is calculated. In (Equation 3), a filtered position value is calculated by adding an amount obtained by multiplying the position estimation error by the position step size to the previously predicted position value. In a similar way (Equation 4), the filtered speed value is calculated by adding the value obtained by multiplying the position estimation error by the speed step size to the previously predicted speed value. In (Equation 5) and (Equation 6), next predicted values of position and velocity are calculated according to the assumed motion model (uniform velocity motion) of the target. By repeating this process every k, the position and velocity of the target, which is the original target tracking purpose, can be estimated.
Let's look at the roles of α and β in the above equation. In the filtering process of (Equation 3) and (Equation 4), when α and β are large, the filtering values of position and velocity change significantly (that is, sensitive to position estimation error), and when α and β are small, position and velocity It can be seen that the filtering value of is changed small (that is, it responds insensitively). Assuming that the target is moving at a constant velocity, it is possible to significantly reduce noise by averaging noise included in the measured position vector by making α and β very small. Conversely, if it is assumed that the target is making a sharp turn, it is possible to quickly catch up with the sharp turning target by making the target position and velocity estimates change quickly by setting α and β large. The former is called accuracy improvement, and the latter is called rapid response improvement. Therefore, depending on how α and β are set, the accuracy and quick response of position and velocity can be adjusted. However, it can be seen from the above description that the accuracy and responsiveness are in a trade-off relationship with each other. That is, when one is good, the other is bad, and when the other is good, this is bad. In summary, when α and β are large, the responsiveness of the position and velocity estimation is improved but the accuracy is low. Therefore, how to determine α and β in consideration of the opposite trend between the accuracy and quick response is the main key to target tracking.
In the existing method, the user directly sets α and β (or related parameters). Therefore, if α and β are set large to increase the quick response, if the measured target position is noisy, the estimated position and velocity values will also become noisy. Conversely, if α and β are set small to increase the accuracy, if the own ship makes a sharp turn during constant velocity motion (cruising), it will not be able to follow the change quickly, so the quick response will be poor. Therefore, the user must bear the inconvenience of manually adjusting α and β (or related parameters) while monitoring the driving situation and sea level situation at that time. In addition, if the user manually determines the parameters, it may be difficult to find an optimal point. For example, as the distance between the own ship and the target increases, the noise in the tangential direction increases in proportion to the distance. This noise is not recognized by the user and thus the user may not be able to properly set the parameters.
Anyway, once the position and velocity of the target are estimated from the measured target position signal sequence (ie, time series signal) input for every frame, the next thing to consider is to determine the search radius. A number of BLOBs must be included in the frame image received in the next frame, and it is unreasonable to check all BLOBs in one frame to determine which one is the target. Therefore, it is reasonable to search the target of the next frame within a constant search radius based on the position estimate estimated through the alpha-beta filter. In the existing case, this search radius is also set manually by the user. If the search radius is small, the number of BLOBs contained therein is small, so it has the advantage of reducing the time to find the target, but the probability of missing the target increases. Conversely, if the search radius is set large, the probability of missing the target decreases, but there is a problem in that the number of BLOBs in the search radius increases and the time to find the target increases. Therefore, the user must bear the inconvenience of having to set the search radius directly while keeping an eye on the situation.
In any case, the target tracking process by a representative existing technique including the above-described target position and velocity estimation, and search radius determination is shown in a flowchart (Fig. 1). When I searched within the double search radius, I needed to take action when the desired target was not found, so the process is shown. That is, a phenomenon may occur that the target BLOB temporarily disappears due to the size of the target, sea conditions, etc. In this case, the target tracking usually determines that there is no measurement target position signal at that time, and is The filtering process by Instead, the next predicted value is calculated only by (Equation 5) and (Equation 6). If the target is not found even after repeating several frames, it is determined that the target is lost and the tracking of the target is terminated. Including this, control of the process during the target survival period from target setting to target tracking and target termination is performed through a state transition process in the flowchart. In addition, there is Compensate Own Ship Motion, etc., but the description is omitted because it is not directly related to the core content of the present invention.

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

Kalata, P.R., “The Tracking Index: A Generalized Parameter for α-β and α-β-γ Target Trackers,” IEEE Trans. Aero. Elect. Syst., Vol. AES-20, March 1984, pp. 174-182.

The present invention has been devised to solve the problems that the existing method of target tracking has in the past. In order to overcome the inconvenience and inaccuracy of adjusting the tracking operation, the present invention provides an apparatus and method for adaptively finding and determining the optimal value of the above parameters according to the operation and sea conditions according to the operation and sea conditions under a certain standard.

Apparatus and method for radar target tracking according to the present invention estimate the position and velocity of the target from the received echo signal of the radar or the measured target position signal from the BLOB analyzer described in the text, and based on this, various maritime information In the derived target tracking system, a technical solution means to automatically or adaptively determine the main coefficients of the filter for estimating the position and speed of the target without user intervention.

delete

Compared to the existing target tracking method in which the user sets the main parameters of the target tracking, that is, the step size and the search radius of the tracking filter, according to the driving situation or the sea situation, the target tracking of the present invention tracks the target without user intervention. Algorithm itself adaptively determines and operates the parameters optimally. As a result, it is possible not only to reduce the operational inconvenience that the user has to experience through the existing method, but also to find the optimal parameter of the target tracking algorithm itself, thereby bringing more reliable target tracking performance. It is expected that this will not only secure the safety of vessel operation, but also contribute to autonomous driving of vessels in the future.

1 is a flowchart of overall target tracking according to the prior art;
2 is a target tracking trajectory when the target moves at a constant velocity;
3 is a target tracking trajectory when the target turns rapidly;
4 shows a search area and a search radius;
5 is an overall target tracking flow chart according to the present invention;
Figure 6 is the position of the measured target and the target tracking position by the target tracking of the present invention;
7 is a comparison of the performance of the target tracking of the conventional method and the target tracking of the present invention

여기서 m은 타겟 트랙킹 오차(정확히는 타겟 트랙킹 오차 벡터의 크기)의 평균값이고 σ은 타겟 트랙킹 오차의 표준 편차이다. κ는 상기 유동성을 산출함에 있어 m과 σ 간의 가중치를 나타내는 상수로서 하나의 디자인 파라미터이다.
일반적으로 일정한 운동 모델과 측정 모델을 가진 타겟을 추적하는데 칼만 필터가 많이 쓰인다. 특히 운동 모델이 등속도 모델인 경우 칼만 필터의 일종인 알파-베타 필터가 많이 쓰인다. 알파-베타 필터는 필터 계수가 α와 β로 결정되는 2차 필터이다. 이 필터 계수값을 결정하는데 있어 (수학식 7)과 같이 유동성이 정의된 경우, 속응성과 정확도를 절충하여 다음과 같이 결정하는 방법이 (참고문헌1)에서 제시되었다. 본 발명에서 이를 Kalata 방법이라고 부른다.

Kalata 방법이 속응성과 정확도를 최적화하는 방법이라고 하나, 실제 운용을 해보면 이에 따른 필터의 과도응답 특성이 다소 진동적 특성을 보임을 알 수 있다.(즉 underdamped response). 이는 정상 상태에 이르기까지의 정착시간이 길어짐을 의미하고 이를 보완하기 위해 critical damping 방법을 구상해본다. critical damping 방법을 구상하는 이유는 필터 특성이 critical damping 일 때 정상 상태에 이르는 정착시간이 가장 짧기 때문이다.(즉 속응성의 강화). 알파-베타 필터에서 critical damping이 되려면 (α+β)² = 4β 또는 β = 2 - α - 2SQRT(1 - α) 식을 만족해야 한다. 여기서 critical damping이 되려면 이 조건만 만족하면 되나 Kalata 방법과 유사한 추적 노이즈 특성(즉 대역폭)을 갖기 위해 상기 두 계수중 α는 Kalata 방법의 (수학식 8)로 정하기로 한다. 따라서 critical damping 방법의 경우 α와 β를 다음과 같이 정한다.

본 발명에서는 상기 Kalata 방법과 critical damping 방법을 다음과 같이 선형적으로 혼용(mixing)하기로 한다.

여기서 ε은 Kalata 방법과 critical damping 방법의 가중치를 나타내는 0과 1 사이의 상수로서 디자인 파라미터이다. 예를 들어 ε이 1이면 Kalata 방법이 되고, ε이 0이면 critical damping 방법이고 그 사이값이면 두 방법의 절충 방법이 된다. 본 발명은 ε을 적절히 줌으로써(예를 들어 0.5) 최적의 타겟 트랙킹 성능을 얻도록 한다.
문제는 (수학식 7)의 유동성을 정의할 때 필요한 m과 σ를 구하는 것이다. 타겟은 임의의 운동을 할 수 있으므로 m을 미리 알기 힘들다. 그리고 타겟이 임의의 거리에 있을 수 있고 해상 상황도 불규칙적이므로 σ도 미리 알기 힘들다. 따라서 m과 σ를 on-line(실제 상황에서)으로 적응적으로 구해야 한다. 앞서 기술했듯이 m은 타겟 트랙킹 위치 추적 오차의 평균이므로 e_k를 on-line으로 LPF 함으로써 아래와 같이 추정할 수 있다. 추정치는 매 프레임마다 변할 수 있으므로 인덱스 k를 붙인다.

여기서 η는 1차 LPF의 대역폭을 결정하는 상수로서 0과 1사이의 값을 가진다. η가 0에 가까우면 e_k에 민감하므로 대역폭이 넓은 필터가 되고, η가 1에 가까우면 e_k에 둔감하므로 대역폭이 좁은 필터가 된다. 상기 식은 e_k를 간단한 1차 LPF한 것으로 필요하면 고차 또는 다른 종류의 LPF하는 것도 가능하다.
다음으로 표준편차 σ를 구해야하는데 우선 e_k의 평균값을 제거해야 한다. 이를 위해 e_k를 HPF(High Pass Filtering)해야 한다. 본 실시예에서는 Butterworth 3차 HPF를 통과시켰는데 다른 방법도 가능하다. 그 HPF 출력을 e_k ^(hpf)라고 했을 때 우선 분산값 σ²을 e_k ^(hpf)의 제곱의 평균 또는 LPF로써 구하고 그 출력의 제곱근을 취하는 식으로 다음과 같이 σ를 구할 수 있다.

(수학식 13) 내지 (수학식 15)에 의해 m_k, σ_k를 구하면 k 시점에서의 유동성을 다음과 같이 구한다.

(수학식 16)을 (수학식 8) 내지 (수학식 12)에 Λ 대신 대입하면 k 시점에서의 알파-베타 필터의 스텝 사이즈 α_k, β_k를 구할 수 있고 다시 이를 (수학식 1) 내지 (수학식 6)에 α, β 대신 대입하면, 본 발명에 의한 알파-베타 필터의 알고리즘이 완성된다. 그 결과 매 프레임마다 유동성이 운전 상황과 해상 상황에 따라 적절히 반영되고 이에 따라 최적 스텝 사이즈가 결정되므로 발명된 알파-베터 필터에 의한 타겟 트랙킹은 사용자의 개입없이 어떠한 운전 상황과 해상 상황에서도 자동적으로 적응적으로 최적적으로 동작할 수 있다.
한편, 일단 상기 방법에 의해 칼만 필터링을 했다면, 다음 프레임에 대한 예측된 타겟 위치 벡터 p_k+1｜k가 계산되었을 것이다. 이것이 다음 프레임의 서치 센터가 될 것인데, 이와 아울러 서치 반경이 필요하다. 전술하였듯이 기존에는 서치 반경도 사용자가 수동으로 직접 정해야 했는데, 상기 방법에 의해 현재 운전 상황과 해상 상황에 대한 통계량을 안다면 이것 역시 적응적으로 구할 수 있다.
서치 반경은 다음 프레임에 수신될 타겟의 측정 위치 벡터의 산포도에 비례하여 정하는 것이 합리적이라고 생각되므로 다음과 같이 서치 반경을 상기 산포도를 나타내는 σ_k에 비례하도록 정한다.

여기서 μ는 서치 반경과 표준 편차간의 스케일링 상수인데, 통계학적으로 노이즈가 가우시안이라고 가정하면 표준편차의 3배 내의 반경내에 들 확률이 0.99 이므로 μ를 3 ~ 5 정도로 잡는다.
한가지 주의할 점이 있는데, 설명의 편의상 (수학식 14) 및 (수학식 15)에서 가산 노이즈 및 해면파로 이루어진 측정 노이즈의 방향성을 무시했는데, 실제로는 자선과 타겟 간의 기하학적 관계에 따른 방향성을 갖게 된다.(그림4)에 나타났듯이, 자선을 중심으로 타겟을 지나는 원을 생각해본다. 기본적으로 측정오차는 거리 방향의 측정 불확실성과 방위각 방향의 측정 불확실성에 기인하며 이들은 별개의 것이고 각자 일정한 크기의 표준 편차를 갖고 있다고 볼 수 있다. 따라서, 원에 직각 방향(즉 거리 방향)의 노이즈는 원의 반경에 상관없이 일정함에 비해, 원에 접선 방향(즉 방위각 방향)의 노이즈는 원의 반경에 비례하여 커지게 된다. 따라서 서치 반경을 결정할 때, 접선 방향 서치 반경 r_i,k와 수직 방향 서치 반경 r_n,k로 나누어 생각하는 것이 합리적이다. 그 결과 서치 영역은 서치 센터를 중심으로 타원 또는 직사각형 또는 부채꼴이 된다. r_i,k, r_n,k를 위한 구체적인 수학식 전개는 본 발명을 설명하는데 큰 관련이 없으므로 생략한다.
만약 믿을만한 측정 노이즈 표준 편차 σ_k에 대한 선행 정보가 있다면 상기와 같이 굳이 on-line으로 추정하는 대신 이값을 바로 알고리즘에 적용할 수 있다. 그러면 σ_k가 정확한 만큼 유동성도 보다 정확할 것이므로 이방법이 더 유리할 수 있다. 그렇지 않다면 당연히 전술한 것처럼 σ_k를 on-line으로 직접 추정하는 것이 낫다.
상기 과정을 종합하면(수학식 1) 내지 (수학식 6)의 기존의 타겟 트랙킹 용 알파-베타 필터를 수정해 다음과 같은 본 발명에 의한 타겟 트랙킹 용 알파-베타 필터를 얻을 수 있다. 이 수정된 알파-베타 필터가 본 발명에 의한 타겟 트랙킹 알고리즘의 근간이 된다.
initialization:

m₀, σ₀, ε, η, κ, μ
for k = 1, 2, 3, …
tracking error:

filtering:

prediction:

statistics:

mobility:

step size:

search radius:

end
상기 설명된 본 발명의 방법을 적용한 실시예를 보임으로써 본 발명의 실용가능성 및 효과를 보다 구체적으로 드러내고자 한다. 실시예는 (그림 5)에 나타난 본 발명에 의한 전체 타겟 트랙킹 흐름도를 통해 설명하고자 한다. 이 흐름도에 의한 구현은 원칙적으로 Hardware나 Software 둘 다 가능하나 최근의 processor(또는 CPU)의 발전에 비추어 보면 Software에 의한 구현이 보다 효과적일 것으로 보여서 본 발명에서는 Software적인 실시예를 보기기로 한다. 즉 흐름도에 나타난 과정은 모두 디지털 영역이고 코딩에 의해서 구현되며 해당 타겟 트랙킹 program 내용은 compile 후 특정 processor 에 다운로드되어 전원을 켜면 흐름도에 나타난 순서대로 동작함을 의미한다.
본 발명에 의한 타겟 트랙킹 program은 전원을 켜면 우선 초기화를 한다. 여기에는 k(=0), p_1｜0, v_1｜0, m₀, σ₀, ε, η, κ, μ 등이 포함된다. 여기서 특히 p_1｜0는 초기 서치 센터에 해당한다. 매 프레임(k)마다 BLOB 분석기로부터 BLOB 데이터가 입력되면 동시에 사용자가 지정한 타겟 데이터가 입력된다. 그러나 기존 타겟 트랙킹 흐름도(그림1)과 비교하면 알 수 있듯이 사용자에 의한 스텝 사이즈, 서치 반경 등은 입력되지 않는다. 대신 이들은 on-line으로 이후 자동적으로 적응적으로 계산될 것이다. 여기서 타겟 데이터는 보통 사용자가 표시기 화면에 커서를 갖다 대고 원하는 BLOB 상에 클릭함으로써 입력된다. BLOB 데이터와는 달리 타겟 데이타는 매 프레임마다 입력되는 것이 아니고 간헐적으로 필요시 사용자가 원할 때 입력된다. 찍다보면 BLOB 상에 찍히지 않을 수도 있으나 큰 상관은 없다. 꼭 그럴 필요는 없지만 편이상 상기 입력들은 정해진 주기 T마다 한 프레임 전제 데이터가 동시에 입력된다고 가정한다.
이전 프레임(또는 초기화)에서 설정된 서치 센터를 중심으로 역시 이전 프레임(또는 초기화)에서 설정된 서치 반경을 잡아 해당 서치 영역내에 존재하는 BLOB들 중 지정된 타겟과 가장 가까운 BLOB를 찾는다. 다른 기준으로 찾을 수도 있으나 편이상 가장 가까운 BLOB를 찾기로 한다. 지정된 타겟에 가장 가까운 BLOB가 발견되면 이 BLOB의 무게중심을 지정된 타겟의 현재 프레임에서 측정된 타겟의 위치 벡터라고 간주하고 BLOB가 발견되었다는 flag를 올린다.
BLOB가 발견되면(수학식 19) 내지 (수학식 30)에 이르는 알파-베타 필터링, 통계량 계산, 유동성 계산, 스텝 사이즈, 서치 반경 결정 등을 수행한다. 여기서 계산된 스텝 사이즈 및 서치 반경 등은 다음 프레임에서 알파-베타 필터링할 때 사용될 것이다.
만약 BLOB가 발견되지 않았다면 측정치 없는 알파-베타 필터링을 수행한다. 이 경우(수학식 20) 및 (수학식 21)의 filtering에 의한 위치, 속도의 update는 이루어지지않고 이전값을 그대로 유지하고(즉, p_k｜k = p_k｜k-1, v_k｜k = v_{k｜k-1) 및 (}수학식 22) 및 (수학식 23)에 의한 prediction만 이루어진다.(수학식 24) 내지 (수학식 30)에서의 통계량, 유동성, 서치반경 등의 update는 이루어지지 않고 이전값을 그대로 유지한다.
이후 자선 움직임 보상, 상태 천이 등이 있으나 이는 본 발명의 내용과 직접 관련있는 사항이 아니므로 설명을 생략한다. 마지막으로 상기 과정을 통해 이번 프레임에서 추출된 타겟의 위치, 속도 등의 정보를 출력한다. 이는 TCPA등 2차 정보 추출에 쓰일 수 있다. 그리고 나서 k를 1만큼 증가시켜 다음 프레임 동작을 기다린다.
기존의 타겟 트랙킹 흐름도(그림1)과 본 발명에 의한 타겟 트랙킹 흐름도(그림5)를 비교해보면 가장 두드러진 차이점을 발견할 수 있는데 굵은선 박스로 표시된 통계량 계산, 유동성 계산, 스텝 사이즈 및 서치 반경 결정 등이다. 본 발명에 따르면 이들을 통해 타겟 트랙킹 알고리즘 자체적으로 적응적으로 최적의 스텝 사이즈, 서치 반경등을 결정해가면서 자동적으로 동작하므로 기존의 타겟 트랙킹 흐름도에서 필요했던 사용자에 의한 스텝 사이즈 및 서치 반경 입력이 필요없게 된다. 그럼으로써 사용자의 불편함을 해소할 수 있고, 보다 정확하고 빠른 성능으로 타겟 트랙킹을 수행해 나갈 수 있다는 것이 본 발명에 의한 타겟 트랙킹의 큰 장점이라고 말할 수 있다.
본 발명에 의한 타겟 트랙킹의 특징을 한층 잘 드러내기 위해 컴퓨터 시뮬레이션을 통해 동일 환경 조건하에서 기존 방식의 타겟 트랙킹과 본 발명의 타겟 트랙킹의 성능을 비교해본다. 실험 조건은 다음과 같다.
타겟은 자선과 동일한 위치인 원점에서 출발하여 지그재그식으로 등속도운동과 급회전을 반복하면서 자선으로부터 멀어지는 진로를 가지고 운항한다. 즉, 등속도 운동(14 meter/sec)을 일정 기간하다가 일정한 거리(7 km) 간격마다 90도씩 급선회하는 운동을 한다. 측정 노이즈는 거리 방향으로 10 meter 만큼, 방위각 방향으로 1 도만큼의 표준 편차를 가진 정규분포를 가지면서 발생한다. 따라서 자선으로부터 멀어질수록 접선 방향의 측정 노이즈가 증가함을 알 수 있다.
(그림6)에 상기의 조건으로부터 발생한 측정된 타겟의 위치(o)와 본 발명의 타겟 트랙킹에 의한 타겟 추적 위치(+)가 나타나있다. 연속적인 부채꼴모양의 도형은 서치 영역을 나타낸다. 우선 측정 노이즈와 타겟의 급선회에도 불구하고 타겟 트랙킹이 원만하게 이루어짐을 알 수 있다. 이는 본 발명에서 제시된 유동성에 기초한 적응적 스텝 사이즈의 적용에 기인한 것이다. 그리고 상기 부채꼴 모양의 서치 영역을 살펴보면 거리 방향으로는 서치 반경이 거의 일정하나 방위각 방향으로는 자선에서 멀어질수록 커짐을 알 수 있다. 이것 역시 서치 반경이 추적 오차의 표준 편차에 기초하여 적응적으로 결정된 것으로 자선에서 멀어질수록 접선방향의 노이즈 성분이 증가하고 따라서 접선 방향의 추적 오차가 증가하기 때문이다. 이렇게 함으로써 타겟 서치시 최소한의 서치영역으로 최대한 타겟 발견 확률을 높이면서, 즉 최적적으로 타겟을 찾을 수 있다.
다음으로 (그림7)에 상기 조건하에 기존 방식의 타겟 트랙킹과 본 발명 방식의 타겟 트랙킹의 성능을 서로 비교하였다. 기존 방식은 일정한 유동성 조건하에서(수학식 8) 및 (수학식 9)에 의해 α, β 값을 미리 결정한 상황에서 운전한 방식임에 비해, 본 발명 방식은 운전중 유동성을 기반으로 적응적으로 α, β를 계산하고 적용하는 방식으로 운전한 방식이다.
(그림7-a)에 기존 방식에 의한 타겟 진로 추정(-)과 타겟의 실제 진로(--)가 나타나있다.(그림7-b)에 본 발명에 의한 타겟의 진로 추정(-) 실제 진로(--)가 나타나 있다. 두 경우 모두 타겟의 실제 진로를 어쨌든 추종하고 있다는 점에서는 동일하지만 그 성능을 보면 큰 차이를 보이고 있음을 알 수 있다. 즉 자선으로부터 거리가 가까운 영역(k < 500)에서는 즉 노이즈가 비교적 작은 영역에서는, 본 발명의 타겟 트랙킹의 노이즈 특성은 기존 방식의 그것과 거의 비슷한 특성을 보이는 반면, 속응성은 기존 방식보다 훨씬 빨리 실제 타겟을 추종하는 특성을 보임을 알 수 있다. 반대로 자선으로부터 거리가 먼 영역(k > 500)에서는 즉 노이즈가 비교적 큰 영역에서는, 본 발명 타겟 트랙킹의 속응성은 기존 방식의 그것과 거의 비슷한 특성을 보이는 반면, 노이즈 특성은 기존 방식보다 현저히 안정적인 특성을 보임을 알 수 있다. 다시 말해 이는 본 발명의 타겟 트랙킹이 유동성을 기초로 적응적으로 스텝 사이즈 및 서치 영역을 결정하기 때문이다.The method presented in the present invention to solve the problem of target tracking by the conventional technique described above will be described.
To this end, first, the mobility of the target is defined. In simple terms, the target fluidity refers to how much the target motion deviates from the constant velocity motion under a certain measurement noise environment, that is, how fluid it is. In other words, if the target is stationary or moving at a constant velocity, there is no fluidity. If the target is accelerated, such as linear acceleration, changing direction, or circular motion, fluidity is high.
When the fluidity is small (uniform motion), the target tracking can be optimally performed using a standard alpha-beta filter. However, when the fluidity is large (accelerated motion), it is difficult to apply a standard alpha-beta filter. Of course, in the case of linear constant acceleration motion, if the alpha-beta-damping filter is applied, target tracking can be optimally performed, but continuous constant acceleration motion is far from the actual target model. In the case of uniform circular motion, the extended Kalman filter model can be applied, but this is also far from the actual target model. The motion model of the target we are dealing with may be temporarily uniform acceleration motion or uniform circular motion, but this is a case that occurs intermittently at an unexpected time and most of them are in constant velocity motion (cruising). Target tracking is based on the basic idea of operating an alpha-beta filter and taking necessary actions during the intermittent acceleration motion. That is, what the present invention proposes is to provide a means for optimally tracking a target according to a specific defined criterion, regardless of when and where the target performs any movement.
If you think of a normal ship, most cruises (constant speed motion) but intermittently change course will make random acceleration motions. From the viewpoint of the above-described liquidity, most of them operate under conditions of low liquidity, and intermittently perform movement with large liquidity. In the case of target tracking through the radar, the position signal of the measurement target captured by the radar is not only the signal component due to the movement of the target itself, but also the noise component due to receiver noise and sea wave, etc. In this case, the target tracking algorithm generally operates in a direction to suppress the noise component and follow the signal component. Therefore, when the target moves at a constant velocity under a constant noise environment, as shown in (Fig. 2-a), when the target measurement position is input as shown in the dotted line, the target tracking trajectory is created as shown in the solid line. In this case, the error between the measurement target position signal and the target tracking trajectory (that is, the target position estimation error or the target tracking error) will look as shown in (Fig. 2-b), and in this case, the average of the target tracking error (i.e. the first moment) ) is close to zero, and the standard deviation (ie the square root of the second moment) will appear as a value equal to that of the additive noise.
On the other hand, if the target changes course during constant velocity motion, as shown in (Fig. 3-a), the measurement target position signal will appear as noise is added to the target's actual course (dotted line), and the target tracking signal tracking it will have a constant velocity. During exercise, as shown in (Fig. 2-a), it tracks well without a target tracking error, and then tracks the target with a large target tracking error for a certain period of time from the moment the course is changed. Of course, if you cruise again on this course, the target tracking error will eventually be close to zero. As a result, in the target tracking error as shown in (Fig. 3-b), in the constant velocity section (Fig. 2-b), the average is close to 0 and the standard deviation has the same value as that of the additive noise, but the moment the course is changed From there, the average value becomes as large as the target tracking error, and the standard deviation around it will have a value equal to that of the additive noise.
From the above description, in the present invention, the fluidity of the above-mentioned target is mathematically defined as follows.

Here, m is the average value of the target tracking error (more precisely, the magnitude of the target tracking error vector), and σ is the standard deviation of the target tracking error. κ is one design parameter as a constant indicating a weight between m and σ in calculating the fluidity.
In general, a Kalman filter is often used to track a target with a constant motion model and a measurement model. In particular, when the motion model is a constant velocity model, an alpha-beta filter, which is a kind of Kalman filter, is often used. The alpha-beta filter is a second-order filter whose filter coefficients are determined by α and β. In determining the filter coefficient value, when fluidity is defined as in (Equation 7), a method of determining the following by compromising quick response and accuracy is suggested in (Reference 1). In the present invention, this is called the Kalata method.

Although the Kalata method is a method for optimizing quick response and accuracy, in actual operation, it can be seen that the transient response characteristic of the filter is somewhat oscillatory (ie, underdamped response). This means that the settling time to reach a steady state is longer, and to compensate for this, consider a critical damping method. The reason for designing the critical damping method is that the settling time to the steady state is the shortest when the filter characteristic is critical damping (ie, reinforcement of quick response). For critical damping in the alpha-beta filter, (α+β) ² = 4β or β = 2 - α - 2SQRT(1 - α) must be satisfied. Here, for critical damping, only this condition needs to be satisfied, but in order to have a tracking noise characteristic (ie, bandwidth) similar to that of the Kalata method, α among the two coefficients is determined by (Equation 8) of the Kalata method. Therefore, in the case of the critical damping method, α and β are determined as follows.

In the present invention, the Kalata method and the critical damping method are linearly mixed as follows.

Here, ε is a design parameter as a constant between 0 and 1 representing the weight of the Kalata method and the critical damping method. For example, if ε is 1, it is the Kalata method, if ε is 0, it is a critical damping method, and if it is a value in between, it is a compromise method between the two methods. The present invention makes it possible to obtain optimal target tracking performance by appropriately giving ε (eg 0.5).
The problem is to find m and σ needed to define the fluidity of (Equation 7). Since the target can have arbitrary motion, it is difficult to know m in advance. And since the target can be at an arbitrary distance and the sea conditions are irregular, it is difficult to know σ in advance. Therefore, m and σ must be adaptively obtained on-line (in real situations). As described above, since m is the average of the target tracking position tracking error, it can be estimated as follows by LPFing _{e k on-line.} Since the estimate can change from frame to frame, index k is appended.

Here, η is a constant that determines the bandwidth of the primary LPF and has a value between 0 and 1. When η is close to 0, _{it is sensitive to e k} , so it becomes a wide-bandwidth filter, and when η is close to 1, it becomes a narrow-bandwidth filter because it is insensitive to _{e k .} The above equation _{is a simple first-order LPF of e k,} and if necessary, a higher-order or other type of LPF is also possible.
Next, the standard deviation σ must be calculated. First, the mean value of _{e k must be removed.} For this, e _k should be subjected to HPF (High Pass Filtering). In this example, the Butterworth 3rd order HPF was passed, but other methods are possible. Assuming that the HPF output is e _k ^(hpf) , first, the variance σ ² is _{obtained as the average or LPF of the squares of e k} ^(hpf) , and then the square root of the output is taken, σ can be obtained as follows.

_{When m k} , σ _k is obtained by (Equation 13) to (Equation 15), the fluidity at time k is obtained as follows.

By substituting (Equation 16) into (Equation 8) to (Equation 12) instead of Λ, the step sizes α _k , β _k of the alpha-beta filter at time k can be obtained, and again, from (Equation 1) to If α and β are substituted for (Equation 6), the alpha-beta filter algorithm according to the present invention is completed. As a result, the fluidity is properly reflected in each frame according to the driving and sea conditions, and the optimal step size is determined accordingly. Therefore, the target tracking by the invented alpha-better filter automatically adapts to any driving and sea conditions without user intervention. can operate optimally.
On the other hand, once Kalman filtering is performed by the above method, the predicted target position vector p _k+1|k for the next frame may have been calculated. This will be the search center for the next frame, but we also need a search radius. As described above, in the past, the user had to manually determine the search radius, but if the statistics on the current driving situation and the sea situation are known by the above method, this can also be adaptively obtained.
Since it is considered reasonable to determine the search radius in proportion to the scatter of the target's measurement position vector to be received in the next frame, the search radius is determined to be proportional to _{σ k indicating the scatter as follows.}

Here, μ is the scaling constant between the search radius and the standard deviation. Statistically, assuming that the noise is Gaussian, the probability of falling within a radius within three times the standard deviation is 0.99, so we take μ as 3 to 5.
There is one caveat, for convenience of explanation, the directionality of the measurement noise composed of the addition noise and the sea wave is ignored in (Equation 14) and (Equation 15). As shown in (Figure 4), consider a circle passing through the target centered on the ship. Basically, the measurement error is caused by the measurement uncertainty in the distance direction and the measurement uncertainty in the azimuth direction. Accordingly, noise in the direction perpendicular to the circle (ie, the distance direction) is constant regardless of the radius of the circle, whereas the noise in the direction tangential to the circle (ie, the azimuth direction) increases in proportion to the radius of the circle. Therefore, when determining the search radius, it is reasonable to divide it into a _{tangential search radius r i,k} and a vertical search radius r _{n,k .} As a result, the search area becomes an ellipse, a rectangle, or a sector around the search center. Specific expression development for r _i,k , r _n,k is omitted because it is not significantly related to the description of the present invention.
If there is prior information on reliable measurement noise standard deviation σ _k , this value can be directly applied to the algorithm instead of estimating on-line as described above. This method may then be more advantageous as the fluidity will be more accurate as _{σ k is correct.} Otherwise, of course, it is better to directly estimate _{σ k on-line as described above.}
Combining the above process, the alpha-beta filter for target tracking according to the present invention as follows can be obtained by modifying the existing alpha-beta filter for target tracking of (Equation 1) to (Equation 6). This modified alpha-beta filter is the basis of the target tracking algorithm according to the present invention.
initialization:

m ₀ , σ ₀ , ε, η, κ, μ
for k = 1, 2, 3, …
tracking error:

filtering:

prediction:

statistics:

mobility:

step size:

search radius:

end
It is intended to reveal the practicality and effect of the present invention in more detail by showing examples to which the method of the present invention described above is applied. The embodiment is to be described through the entire target tracking flow chart according to the present invention shown in (Fig. 5). Implementation by this flowchart is in principle possible in both hardware and software, but in view of the recent development of the processor (or CPU), the implementation by software seems to be more effective. That is, the process shown in the flowchart is all digital domain and implemented by coding, and the contents of the target tracking program are downloaded to a specific processor after compilation and are operated in the order shown in the flowchart when the power is turned on.
The target tracking program according to the present invention initializes first when the power is turned on. These include k(=0), p _1|0 , v _1|0 , m ₀ , σ ₀ , ε, η, κ, μ, and the like. In particular, p _{1 | 0} corresponds to the initial search center. When BLOB data is input from the BLOB analyzer every frame (k), the target data designated by the user is input at the same time. However, as can be seen from comparison with the existing target tracking flowchart (Figure 1), the user's step size and search radius are not input. Instead they will be automatically and adaptively calculated afterward on-line. Here, the target data is usually entered by the user placing the cursor on the indicator screen and clicking on the desired blob. Unlike BLOB data, the target data is not input every frame, but intermittently when the user wants it. If you take a picture, it may not appear on the BLOB, but it doesn't matter. Although it is not necessary, for the sake of convenience, it is assumed that one frame premise data is simultaneously input for each predetermined period T.
The search radius set in the previous frame (or initialization) is also taken centering on the search center set in the previous frame (or initialization) to find the BLOB closest to the designated target among the BLOBs existing in the corresponding search area. You can search by other criteria, but for convenience, we will find the nearest BLOB. When the BLOB closest to the designated target is found, the center of gravity of this BLOB is regarded as the target's position vector measured in the current frame of the designated target, and a flag indicating that the BLOB has been found is raised.
When a BLOB is found (Equation 19) to (Equation 30), alpha-beta filtering, statistical calculation, fluidity calculation, step size, search radius determination, and the like are performed. The calculated step size and search radius will be used for alpha-beta filtering in the next frame.
If no BLOB is found, alpha-beta filtering without measurement is performed. In this case (Equation 20) and (Equation 21), the position and velocity update by filtering is not performed and the previous values are maintained (ie, p _k|k = p _k|k-1 , v _{k| k} = v _{k|k-1) and (} Equation 22) and (Equation 23) only prediction is made. Updates of statistics, fluidity, search radius, etc. in (Equation 24) to (Equation 30) are It is not done and the previous value is kept as it is.
Thereafter, there are self-charity motion compensation and state transition, but these are not directly related to the contents of the present invention, and thus descriptions thereof will be omitted. Finally, information such as the position and velocity of the target extracted from this frame through the above process is output. This can be used to extract secondary information such as TCPA. Then, increment k by 1 and wait for the next frame operation.
Comparing the existing target tracking flow chart (Figure 1) and the target tracking flow chart according to the present invention (Figure 5), the most striking difference can be found. to be. According to the present invention, the target tracking algorithm automatically operates while adaptively determining the optimal step size, search radius, etc. through these, so there is no need to input the step size and search radius by the user, which was required in the existing target tracking flowchart. do. In this way, it can be said that the user's inconvenience can be solved, and that target tracking can be performed with more accurate and faster performance is a great advantage of target tracking according to the present invention.
In order to better reveal the characteristics of the target tracking according to the present invention, the performance of the conventional target tracking and the target tracking of the present invention is compared under the same environmental conditions through computer simulation. The experimental conditions are as follows.
The target starts from the origin, which is the same position as the own ship, and operates with a path away from the own ship while repeating constant velocity motion and sharp rotation in a zigzag manner. That is, a constant velocity motion (14 meter/sec) is performed for a certain period of time, and the exercise is performed by rapidly turning 90 degrees at intervals of a certain distance (7 km). Measurement noise is generated with a normal distribution with a standard deviation of 10 meters in the distance direction and 1 degree in the azimuth direction. Therefore, it can be seen that as the distance from the own ship increases, the measurement noise in the tangential direction increases.
In (Fig. 6), the measured target position (o) resulting from the above conditions and the target tracking position (+) by the target tracking of the present invention are shown. A continuous sector-shaped figure represents a search area. First, it can be seen that the target tracking is smoothly performed despite the measurement noise and the rapid turning of the target. This is due to the application of an adaptive step size based on the flexibility presented in the present invention. And, looking at the sector-shaped search area, it can be seen that the search radius is almost constant in the distance direction, but increases as the distance from the own ship increases in the azimuth direction. This is also because the search radius is adaptively determined based on the standard deviation of the tracking error, and the farther away from the ship, the more the noise component in the tangential direction increases, and thus the tracking error in the tangential direction increases. In this way, it is possible to find the target optimally while increasing the target discovery probability as much as possible with the minimum search area during target search.
Next, in (Fig. 7), the performance of the target tracking of the conventional method and the target tracking of the present invention is compared under the above conditions. In contrast to the existing method in which α and β values are previously determined by Equation 8 and Equation 9 under constant fluidity conditions, the present invention method adaptively α based on fluidity during operation. , β is calculated and applied.
(Figure 7-a) shows the target route estimation (-) and the target’s actual route (--) by the existing method. (Figure 7-b) shows the target’s route estimation (-) actual route according to the present invention (Figure 7-b) (--) appears. In both cases, they are the same in that they follow the target's actual course anyway, but if you look at their performance, you can see that there is a big difference. That is, in a region close to the ship (k < 500), i.e., in a region where noise is relatively small, the noise characteristic of the target tracking of the present invention is almost similar to that of the conventional method, whereas the rapid response is much faster than the conventional method. It can be seen that the characteristic of following the target is shown. Conversely, in a region far away from the ship (k > 500), that is, in a region where noise is relatively large, the rapid response of the target tracking of the present invention shows almost similar characteristics to that of the conventional method, whereas the noise characteristic is significantly more stable than that of the conventional method. it can be seen that In other words, this is because the target tracking of the present invention adaptively determines the step size and the search area based on flexibility.

Claims (22)

A target tracking device for estimating the position and velocity of a target from a received echo signal of a radar or an output of a BLOB analyzer, and deriving valid information based thereon,
Estimation of the position and velocity of the target is made by an alpha-beta filter among Kalman filters,
The alpha and beta coefficients of the alpha-beta filter are adaptively determined according to the fluidity constant (Λ) of [Equation 7] that is proportional to the ratio of the average value (m) and the standard deviation value (σ) of the target tracking error,
When the fluidity constant is relatively large, the alpha and beta values, which are the step sizes of the alpha-beta filter, are increased to increase the speed of target tracking, and when the fluidity constant is relatively small, target tracking is performed by decreasing the alpha and beta values. increase the accuracy of
The alpha and beta values {α ^(mixing) , β ^(mixing) } of the alpha-beta filter are derived from [Equation 8] and [Equation 9] by the Kalata method, the alpha and beta values {α ^(kalata) , β ^{(kalata) } and the alpha and beta values {α (critical)} , β ^(critical) } derived from [Equation 10] and [Equation 11] by the critical damping method to ε:1-ε(0 ≤ Radar target tracking device, characterized in that determined by [Equation 12], which is a value internally divided by ε ≤ 1) (however, κ is a proportionality constant).
[Equation 7]

[Equation 8]

[Equation 9]

[Equation 10]

[Equation 11]

[Equation 12]

delete delete The method according to claim 1,
The radar target tracking apparatus, characterized in that the average value and the standard deviation value of the target tracking error are determined from the estimation error of the alpha-beta filter.
5. The method according to claim 4,
_{The estimated average value (m k} ) of the target tracking error in the k-th frame of the radar reception echo signal or the BLOB analyzer output is derived by [Equation 13],
_{The estimated standard deviation value (σ k} ) of the target tracking error in the k-th frame of the radar received echo signal or the BLOB analyzer output is derived by [Equation 14] and [Equation 15],
In the k-th frame of the radar reception echo signal or the BLOB analyzer output, the fluidity constant (Λ _k ) is a radar target tracking device, characterized in that determined by [Equation 16] (however, 0 ≤ η ≤ 1, e _k is Estimation error of the target in the kth frame of the radar received echo signal or BLOB analyzer output, e _k ^(hpf) is the result of high pass filtering of e _k).
[Equation 13]

[Equation 14]

[Equation 15]

[Equation 16]

6. The method of claim 5,
When the standard deviation value of the target tracking error is known in advance as prior information, the known target tracking instead _{of the estimated standard deviation value (σ k ) of the target tracking error according to Equations 14 and 15 above} Radar target tracking apparatus, characterized in that by substituting the standard deviation value of the error into the denominator of [Equation 16] _{to determine the fluidity constant (Λ k).}
6. The method of claim 5,
_{Radar target tracking device, characterized in that the search radius (r k} ) of the target tracking in the k-th frame of the output of the radar received echo signal or the BLOB analyzer is determined by [Equation 17] (where μ is a proportional constant) .
[Equation 17]

delete A target tracking method for estimating the position and velocity of a target from a received echo signal of a radar or an output of a BLOB analyzer, and deriving valid information based thereon,
Estimation of the position and velocity of the target is made by an alpha-beta filter among Kalman filters,
The alpha and beta coefficients of the alpha-beta filter are adaptively determined according to the fluidity constant (Λ) of [Equation 7] that is proportional to the ratio of the average value (m) and the standard deviation value (σ) of the target tracking error,
When the fluidity constant is relatively large, the alpha and beta values, which are the step sizes of the alpha-beta filter, are increased to increase the speed of target tracking, and when the fluidity constant is relatively small, target tracking is performed by decreasing the alpha and beta values. increase the accuracy of
The alpha and beta values {α ^(mixing) , β ^(mixing) } of the alpha-beta filter are derived from [Equation 8] and [Equation 9] by the Kalata method, the alpha and beta values {α ^(kalata) , β ^{(kalata) } and the alpha and beta values {α (critical)} , β ^(critical) } derived from [Equation 10] and [Equation 11] by the critical damping method to ε:1-ε(0 ≤ A radar target tracking method, characterized in that it is determined by [Equation 12], which is a value internally divided by ε ≤ 1) (however, κ is a proportionality constant).
[Equation 7]

[Equation 8]

[Equation 9]

[Equation 10]

[Equation 11]

[Equation 12]

delete delete delete delete delete delete delete delete delete delete delete delete delete

Images