JP5501555B2 - Multi-target tracking device - Google Patents

Description

  The present invention relates to a multi-target tracking device that observes a target using a sensor and tracks the target using the obtained observation information (observed value).

  Many papers, patents, etc. have already mentioned the technology for tracking the target using the observation values obtained from the sensor and estimating the motion specifications. Proposals have been made. As a problem in tracking a plurality of close targets, it is easy to cause erroneous tracking by erroneously correlating the target track and the observed value. In a multi-target tracking device (for example, see Patent Document 1) for dealing with such a case, a plurality of hypotheses are generated for the correlation between the target track and the observed value, and the final correlation is determined while calculating the reliability. .

  Here, the conventional multi-target tracking device described above will be described with reference to FIGS. FIG. 6 is a block diagram showing a configuration of a conventional multi-target tracking device.

  In FIG. 6, the conventional multi-target tracking device 10A reads the observation value at the latest observation time from the sensor 20, calculates the gate of the existing track, and further checks whether the input observation value is within the gate, A position information correlation determining unit 2 that determines which wakes the incoming observation value can correlate with, a wake generation unit 3A that generates a wake based on the determination result inside and outside the gate of the observation value, and the generated wake Correlation hypothesis generation unit 4A that generates hypotheses in combination, hypothesis selection unit 5 that reduces hypotheses by deleting hypotheses with low reliability, and integration of similar hypotheses, wake file 8 that stores wakes, and hypotheses A hypothesis file 9 to be stored is provided.

  Next, the operation of the conventional multi-target tracking device will be described with reference to the drawings. FIG. 7 is a flowchart showing the operation of the conventional multi-target tracking device.

  First, in step 901 (observation value input), the position information correlation determination unit 2 reads the observation value at the latest observation time from the sensor 20.

  Next, in step 902 (gate inside / outside determination), the position information correlation determination unit 2 calculates the gate of the existing wake, further checks whether the input observation value is within the gate, and enters the observation value that has entered. To which wakes can be correlated. Here, the wake is assumed to be a vector having six elements of the following expression (a-1), which is composed of a target position and velocity in a three-dimensional space.

  FIG. 8 is a diagram illustrating an example of existing tracks and observed values on the xy plane. In this example, there are T1 and T2 as existing tracks, and three observation values O1, O2, and O3 are obtained. A gate, which is a spatial region where observation values may be obtained, is calculated from the predicted position calculated using the motion parameters of the existing wake and the error covariance of the estimated motion parameters. Is done. In this example, it is assumed that the observation values O1 and O2 enter the gates of the existing tracks T1 and T2. In this case, the observed value O1 can be correlated with the existing tracks T1 and T2, and the observed value O2 can be correlated with the existing tracks T1 and T2.

The details of the method for judging inside / outside of a set of wakes and observations are described below. The predicted value (^) X _k (−) is calculated by the following equation (a-2). In the specification, (^) X indicates that there is ^ on X.

Here, Φ _k−1 is a transition matrix, and (^) X _k−1 (+) is a smooth value before one observation time.

The prediction error covariance P _k (−) is calculated by the following equation (a-3).

Here, P _k−1 (+) is a smoothing error covariance matrix before one observation time, Q _k−1 is a driving noise covariance matrix, and Γ ₁ (k−1) is a driving noise conversion matrix.

The residual covariance matrix S _k is calculated by the following equation (a-4).

Here, H _k is an observation matrix, R _k is an observation error covariance matrix, and Γ ₂ (k) is a conversion matrix of observation noise.

Using the above residual covariance matrix, the gate outside judgment of observations Z _k is determined by the success or failure conditions of the following formulas (a-5).

Here, Z _k (−) is a predicted observation value, and can be calculated by the following equation (a-6).

  Moreover, d of Formula (a-5) is a parameter of a boundary value determined by the significance level in the χ square test. If this equation (a-5) holds (the observed value is within the range of the error covariance matrix), “the observed value and the wake are correlated”; otherwise, the observed value is outside the range of the error covariance matrix. ), It is determined that “the observed value and the wake have no correlation”.

  Next, in step 903 (wake generation), the wake generation unit 3A generates a wake based on the determination result of the observation value inside / outside the gate. There are three types of tracks to be created: “update track”, “new track”, and “memory track track”. The smoothness specifications at the observation time of these tracks are calculated, and the likelihood of the track corresponding to the correlation result is calculated for the updated track.

  The “update wake” is a wake generated by adding any observation value in the gate to the existing wake. In the example of FIG. 8, the wake that updated the existing wake T1 using the observed value O1, the wake that updated the existing wake T1 using the observed value O2, the wake that updated the existing wake T2 using the observed value O1, and the observed value A total of four wakes are generated by updating the existing wake T2 using O2.

  The smooth vector is calculated by the following equation (a-7).

Here, K _k is a filter gain and can be calculated by the following equation (a-8).

  The smooth error covariance matrix is calculated by the following equation (a-9).

  The likelihood of a wake is calculated on the assumption that the probability distribution of observed values is a Gaussian distribution centered on the predicted position (three-dimensional position).

  The “new wake” is a wake starting from the observation value obtained at the latest observation time. In the example of FIG. 8, a total of three tracks are generated: a new track starting from the observation value O1, a new track starting from the observation value O2, and a new track starting from the observation value O3. The smooth vector is initially set so that the position component is equal to the observed value and the velocity component is zero. In addition, the smooth error covariance matrix cannot be calculated at this time, and is calculated using the observation error covariance matrix when the second observation value is determined.

  The “memory track wake” is a wake that “there was no correlated observation value at the corresponding time” with respect to the existing wake. In the example of FIG. 8, a total of two tracks are generated: a memory track track of the existing track T1 and a memory track track of the existing track T2.

  The smooth vector is set to a value equivalent to the prediction vector, and the smooth error covariance matrix is set to a value equivalent to the prediction error covariance matrix.

  Next, in step 904 (correlation hypothesis generation), first, the correlation hypothesis generation unit 4A generates a hypothesis by combining the wakes generated in step 903. This hypothesis must satisfy the following conditions:

  The hypothesis must be an extension of one of the existing hypotheses generated one observation time ago. FIG. 9 is a diagram illustrating two examples of correlation hypotheses. Correlation hypothesis 1 shown in FIG. 6A adopts two wakes, a wake obtained by updating the existing wake T1 with the observation value O2 and a wake obtained by updating the existing wake T2 with the observation value O1, and the observation value O3 is an unnecessary signal. I reckon. Since this correlation hypothesis is based on the premise that T1 and T2 exist as existing tracks, the correlation hypothesis is generated by the development of the existing hypothesis 1 shown in FIG. FIG. 10 is a diagram showing three examples of existing hypotheses. Therefore, if the existing hypothesis 1 is not included in the existing hypothesis group, the correlation hypothesis 1 cannot be generated. Correlation hypothesis 2 adopts two tracks, a track obtained by updating the existing track T1 with the observation value O1 and a track T3 starting from the observation value O3, and regards O2 as an unnecessary signal. Since this correlation hypothesis is based on the premise that only T1 exists as an existing track, it is generated by the development of the existing hypothesis 2 shown in FIG.

  For the selected observation and wake combination, the observation must be in the wake gate. For example, a hypothesis for adopting a track obtained by updating the existing track T1 with the observation value O3 is not generated.

  An observation is associated with any one wake or is considered an unnecessary signal. For example, in one hypothesis, the observed value O1 is not associated with both the existing tracks T1 and T2, and the observed value O2 is not considered to be part of the track or an unnecessary signal.

  At most one observation value is used for updating one existing track. For example, both O1 and O2 are not associated with the existing track T1 in one hypothesis.

  For each hypothesis, the reliability is calculated as the evaluation value. The reliability is calculated by calculating the product of the selected observation value and the likelihood of the wake, and further normalizing so that the sum of the reliability of all hypotheses is 1.0. The equation for calculating the reliability before normalization is the following equation (a-11).

The meaning of each stage on the right side of the formula (a-11) is as follows.
First stage: A term to be accumulated when the existing track and the position information of the observation value are correlated.
Second stage: This is a term to be accumulated when the existing wake becomes missing.
Third stage: A term that is accumulated when the observed value is set as a new target.
Fourth stage: A term that is integrated when an observed value is an unnecessary signal.
5th stage: The reliability of the parent hypothesis.

The meanings of symbols in formula (a-11) are as follows.
P ^k _i : The i-th hypothesis at the observation time k.
P _D : A detection probability of the sensor, which is a parameter set before the apparatus is activated.
β _FT : Unnecessary signal generation density, which is a parameter set before starting the apparatus.
β _NT : A new target generation density, which is a parameter set before the apparatus is started.
N _TGT : Number of wakes (referred to as targets) referenced in the parent hypothesis.
N _DT : Number of existing wakes assumed to be correlated with any observation value at the latest time in the hypothesis.
N _NT : The number of observations that were considered as new targets at the latest time in the hypothesis.
N _FT is the number of observed values that are considered as unnecessary signals at the latest time in the hypothesis.

Next, in step 905 (hypothesis selection integration), the hypothesis selection integration unit 5 reduces the hypothesis group by deleting hypotheses with low reliability and integrating similar hypotheses. The following methods are well known for this reduction.
・ Set a threshold for reliability and delete all hypotheses with reliability less than that.
Set an upper limit for the number of hypotheses, leave only the set number of hypotheses in descending order of reliability, and delete others.
-Integrate hypotheses with the same correlation between observation values for the past several observation times.

  In step 906 (wake display), the wake display unit 30 selects a hypothesis corresponding to the reliability, and displays the wake included in the hypothesis.

  The above is the flow of data processing for one observation time. When this one cycle ends, the next observation value is read and the processing of the next cycle is executed.

  When attribute information other than information related to the position of the target, for example, a target shape by the image sensor, is obtained from the sensor 20, it can be said that there is a high possibility that a correlation result regarding observation values having similar attributes as the same target is correct. Another conventional multi-target tracking device (for example, refer to Patent Document 2) that takes correlation by including this attribute in a tracking state variable has been proposed. This other conventional multi-target tracking apparatus also shows a correlation hypothesis reliability calculation method, and can be easily applied to the conventional multi-target tracking apparatus described above.

Japanese Patent No. 3145893 Japanese Patent No. 3457174

  When tracking with the observation value from the sensor obtained by applying another conventional multi-target tracking device as described above to the conventional multi-target tracking device described above, the attribute is abrupt, For example, if a situation occurs in which the target turns sharply and its shape changes completely, the track that has been tracked up to that point cannot be correlated with the observed value, and is recognized as a separate target.

  For example, consider a case where two targets are observed and the transition of the position and attribute is as shown in FIG. Here, it is assumed that the attribute information is obtained as two-dimensional numerical information of attribute A and attribute B. It is assumed that a trajectory in which two targets approach each other on the position space is drawn halfway, and that a significant value change occurs in the attribute A of the target (target 1) on the attribute space. In the extension of the prior art, the attribute information of the observed value of the target 1 cannot be correlated before and after the change of the attribute A, and an erroneous hypothesis indicating the existence of three targets as shown in FIG. That is, there is a problem that the tracking performance may be deteriorated by using the attribute as compared with the case of tracking only by the position information.

  The present invention has been made to solve the above-described problems, and its purpose is to perform correct correlation determination even when a sudden change in the attribute occurs while using attribute information obtained from the sensor. A multi-target tracking device that can be obtained is obtained.

The multi-target tracking device according to the present invention includes a sensor that generates target position information and attribute information related to the target other than the position information as observation values, an observation value at the latest observation time that is input, and each in the wake file Attribute correlation determination unit that determines the possibility of correlation of wakes with respect to attributes, and calculates a gate related to position information of each wake in the wake file, and whether the position information of the input observation value is in the gate And a position information correlation determination unit that determines which wake the input observation value can correlate with, a correlation result regarding the observation value and the attribute of the wake, and a position information of the observation value and the wake Based on the correlation results, the combined wake generation unit that generates the wake that estimates the target motion specifications at the latest observation time is combined with the wake generated by the integrated wake generation unit to generate a hypothesis. In addition, an integrated correlation hypothesis generation unit that calculates the reliability of each hypothesis using an attribute variation probability parameter that is a probability that a sudden change of the attribute occurs, and a hypothesis of low reliability calculated by the integrated correlation hypothesis generation unit And a hypothesis selection integration unit that reduces hypothesis groups by integrating similar hypotheses, and the integrated correlation hypothesis generation unit uses the attribute information change probability parameter as a sudden change probability P _{T of the} attribute information and the attribute using a body volume V _R of the range information can take, from a sudden change probability P _T of the attribute information, and calculates the likelihood of the case where the correlation is taken for the attribute, before Kitai product V _R, the correlation for the attribute Likelihood is calculated when it is not possible, and hypothesis reliability is calculated using the product of the respective likelihoods in the two cases.

  The multi-target tracking device according to the present invention has an effect that correct correlation determination can be performed even when an abrupt change of the attribute occurs while using attribute information obtained from the sensor.

Embodiment 1 FIG.
A multi-target tracking device according to Embodiment 1 of the present invention will be described with reference to FIGS. FIG. 1 is a block diagram showing the configuration of the multitarget tracking apparatus according to Embodiment 1 of the present invention. In the following, in each figure, the same reference numerals indicate the same or corresponding parts.

In FIG. 1, the multi-target tracking device 10 according to the first embodiment of the present invention includes an attribute correlation determination unit 1 that performs attribute determination on the inside / outside determination of each track in the track file 8 and the observation value at the latest observation time, Position information that calculates the gate related to the position information of the wake, checks whether the position information of the input observation value is in the gate, and determines which wake the incoming observation value can correlate with A correlation determination unit 2, an integrated wake generation unit 3 that generates a wake that estimates the motion specifications of the target at the latest observation time, based on the correlation result regarding the position information of the wake and the observation value;
An integrated correlation hypothesis generation unit 4 that generates hypotheses by combining the generated wakes, a hypothesis selection integration unit 5 that reduces hypotheses by deleting hypotheses with low reliability, and integrating similar hypotheses, and a wake that stores wakes A file 8 and a hypothesis file 9 for storing hypotheses are provided.

  Next, the operation of the multi-target tracking device according to the first embodiment will be described with reference to the drawings. FIG. 2 is a flowchart showing the operation of the multitarget tracking apparatus according to Embodiment 1 of the present invention.

  In the prior art, the target position and speed are estimated parameters, but in the first embodiment, an attribute consisting of attribute A and attribute B is further added thereto. The state variable can be described as the following equation (b-1).

X _{trk, k} corresponds to a state variable in the prior art, and X _{att, k} represents a state variable related to attribute information.

  A processing procedure when an observation value is input from the sensor 20 will be described below.

  First, in step 101 (observation value input), the multi-target tracking device 10 reads an observation value at the latest observation time of the sensor 20. Information on the target position obtained from the observed values is taken into the position information correlation determining unit 2 and attributes are taken into the attribute correlation determining unit 1.

  Next, in step 102 (attribute comparison determination), the attribute correlation determination unit 1 performs gate inside / outside determination of each track in the track file 8 and the observation value at the latest observation time with respect to the attribute.

The details of the method for judging inside / outside of a set of wakes and observations are described below. The predicted value (^) X _{att, k} (−) regarding the attribute is calculated by the following equation (b-4).

Here, Φ _{att, k−1} is a transition matrix related to the attribute, and (^) X _{att, k−1} (+) is a smooth value one observation time before the attribute.

The prediction error covariance P _{att, k} (+) regarding the attribute is calculated by the following equation (b-5).

Here, P _{att, k-1} (+) is the smoothing error covariance matrix related to the attribute before one observation time, Q _{att, k-1} is the driving noise covariance matrix related to the attribute, and Γ _{att, 1} (k-1). Is a transformation matrix of driving noise related to attributes.

The residual covariance matrix S _{att, k for the} attribute is calculated by the following equation (b-6).

Here, H _{att, k} is an observation matrix relating to attributes, R _{att, k} is an observation error covariance matrix relating to attributes, and Γ _{att, 2} (k) is a transformation matrix of observation noise relating to attributes.

Using the above residual covariance _, the inside and outside of the gate of the observation value Z _{att, k} related to the attribute is determined by the success or failure of the condition of the following equation (b-7).

Here, Z _{att, k} (−) is a predicted observation value, and is calculated by the following equation (b-8).

Further, d _att is a boundary value parameter determined by the significance level in the χ square test.

  If this expression (b-7) is satisfied (the observed value is within the range of the error covariance matrix), “there is a correlation with the attribute”, and if not (the observed value is outside the range of the error covariance matrix) In this case, it is determined that “there is no correlation with respect to attributes”.

  Next, in step 103 (gate inside / outside determination), the position information correlation determination unit 2 calculates a gate related to the position information of each wake in the wake file 8. This gate is an area on the space where the spread is designated by the residual covariance, which is the sum of the wake prediction error covariance and the observation error covariance. Further, it is checked whether or not the position information of the input observation value is in the gate, and it is determined which wake the incoming observation value can be correlated with.

The details of the method for judging inside / outside of a set of wakes and observations are described below. The predicted value (^) X _{trk, k} (−) regarding the target motion is calculated by the following equation (b-9).

Here, Φ _{trk, k−1} is a transition matrix related to the target motion, and (^) X _{trk, k−1} (+) is a smooth value one observation time before the target motion.

The prediction error covariance P _{trk, k} (−) regarding the target motion is calculated by the following equation (b-10).

Here, P _{trk, k-1} (+) is a smooth error covariance matrix related to the target motion before one observation time, Q _{trk, k-1} is a drive noise covariance matrix related to the target motion, and Γ _{trk, 1} (k− 1) is a drive noise conversion matrix for the target motion.

The residual covariance matrix S _{trk, k} regarding the target motion is calculated by the following equation (b-11).

Here, H _{trk, k} is an observation matrix related to the position of the target motion, R _{trk, k} is an observation error covariance matrix related to the position of the target motion, and Γ _{trk, 2} (k) is a conversion matrix of observation noise related to the position of the target motion. It is.

Based on the determination using the residual covariance matrix, whether the observed value Z _{trk, k} relating to the position of the target motion is inside or outside the gate is determined by the success or failure of the condition of the following equation (b-12).

Here, Z _{trk, k} (−) is a predicted observation value and is calculated by the following equation (b-13).

Further, d _trk is a boundary value parameter determined by the significance level in the χ square test.

  When this formula (b-12) is satisfied (the observed value is within the range of the error covariance matrix), “there is a correlation with respect to the position”, and when it is not satisfied (the observed value is outside the range of the error covariance matrix) In this case, it is determined that “there is no correlation with respect to the position”.

  Next, in step 104 (integrated wake generation), the integrated wake generation unit 3 determines the target of the latest observation time based on the correlation result regarding the observation value and the attribute of the wake and the correlation result regarding the observation value and the position information of the wake. A wake that estimates the motion specifications is generated. There are three types of tracks that are created: “update track”, “new track”, and “memory track track”. For these tracks, smooth specifications (smooth vector, smooth error covariance matrix) that are estimated values of motion and attributes at the latest observation time are calculated.

  (1) The updated wake is a wake that is generated by adding the observation value that has entered the gate related to the position information of the existing wake in step 103 (gate inside / outside determination). The smooth vector is calculated by the following equation (b-14).

Here, K _{trk, k} and K _{att, k} are filter gains relating to position and attributes, and can be calculated by the following equations (b-16) and (b-17).

  The smooth error covariance matrix is calculated by the following equations (b-18) and (b-19).

  The likelihood of a wake is calculated as follows assuming that the probability distribution of observed values is a Gaussian distribution centered on the true value. There are two kinds of likelihood of the wake, that is, the likelihood related to position information and the likelihood related to attributes described below, and both are calculated.

  The likelihood functioning on the position information is calculated by the following equation (b-20).

  The likelihood related to the attribute is calculated by the following formula (b-21) assuming a Gaussian distribution, for example, by applying the attribute distribution function.

  When it is determined in step 102 (attribute comparison determination) that the observation value for the attribute is outside the wake gate, the smooth value of the attribute is the latest observation value, and the error covariance is the observation error. Set to be equivalent to covariance (only attribute is calculated with the estimated value equivalent to the following new track).

  (2) The new wake is a wake starting from the observation value obtained at the latest observation time. The smooth vector is initialized to the position component equivalent to the observed value, the velocity component is set to 0, and the attribute-related component is initialized to the observed value. Further, the smooth error covariance matrix cannot be calculated at this time, and is calculated using the observation error covariance matrix when the second observation value is determined.

  (3) The memory track track is a track indicating that “there was no correlated observation value at the corresponding time” with respect to the existing track. The smooth vector is set to a value equivalent to the prediction vector, and the smooth error covariance matrix is set to a value equivalent to the prediction error covariance matrix.

Next, in step 105 (integrated correlation hypothesis generation), first, the integrated correlation hypothesis generation unit 4 generates a hypothesis by combining the wakes generated in step 104 (integrated wake generation). This hypothesis must satisfy the following conditions:
・ The hypothesis must be an extension of one of the existing hypotheses generated one observation time ago.
• For the combination of selected observations and existing wakes, the position information of the observations must be in the position gate of the existing wakes.
All observations are associated with any one wake or are considered unnecessary signals.
• There is at most one observation that can be correlated to one existing track.

  For each hypothesis, the reliability is calculated as the evaluation value. The reliability is obtained by calculating the product of the selected observation value and the likelihood of the wake, and further normalizing so that the sum of the reliability of all hypotheses is 1.0. The calculation formula of the reliability before normalization is the following formula (b-22).

The meaning of each stage on the right side of the above formula (b-22) is as follows.
First stage: A term to be accumulated when the existing track and the position information of the observation value are correlated. Further, the attribute is limited to the case where correlation is obtained.
Second stage: A term that is accumulated when the existing track and the position information of the observation value are correlated. However, the attribute is limited to a case where correlation cannot be obtained.
Third stage: This term is accumulated when the existing wake becomes a detection failure.
Fourth stage: A term to be accumulated when the observed value is set as a new target.
Fifth stage: A term that is integrated when an observed value is an unnecessary signal.
6th stage: The reliability of the parent hypothesis.

The following two points are used as parameters used in the second-stage calculation on the right side.
P _T : A probability that an abrupt change of an attribute occurs, and is a parameter (attribute variation probability parameter) set before the apparatus is activated. In addition, when the attribute change is similar to a sudden change in the past, a larger value (weighting) than other changes may be set.
V _R is a volume in a range that can be taken by the target attribute to be tracked, and is a parameter set before starting the apparatus.

Next, in step 106 (hypothesis integration selection), the hypothesis selection integration unit 5 reduces the hypothesis group by deleting hypotheses with low reliability and integrating similar hypotheses. For this reduction, the following well-known method is used.
・ Set a threshold for reliability and delete all hypotheses with reliability less than that.
Set an upper limit for the number of hypotheses, leave only the set number of hypotheses in descending order of reliability, and delete others.
-Integrate hypotheses with the same correlation between observation values for the past several observation times.

  In step 107 (wake display), the wake display unit 30 selects a hypothesis corresponding to the reliability, and displays the wake included in the hypothesis.

  The above is the flow of data processing for one observation time of the sensor 20, and when this one cycle ends, the observation value at the next observation time is read and the processing of the next cycle is executed.

  As described above, according to the multi-target tracking device according to the first embodiment, even when the normal correlation determination method cannot obtain the correlation between the track and the observation value due to the sudden change of the attribute obtained from the same target, the same target is obtained. It is possible to maintain the hypothesis to recognize with high reliability.

Embodiment 2. FIG.
A multi-target tracking device according to Embodiment 2 of the present invention will be described with reference to FIGS. 3, 4, and 5. FIG. FIG. 3 is a block diagram showing the configuration of the multitarget tracking apparatus according to Embodiment 2 of the present invention.

In FIG. 3, the multi-target tracking device 10 according to the second embodiment of the present invention includes an attribute correlation determination unit 1 that performs attribute determination on the inside / outside determination of each track in the track file 8 and the observation value at the latest observation time, Position information that calculates the gate related to the position information of the wake, checks whether the position information of the input observation value is in the gate, and determines which wake the incoming observation value can correlate with A correlation determination unit 2, an integrated wake generation unit 3 that generates a wake that estimates the motion specifications of the target at the latest observation time, based on the correlation result regarding the position information of the wake and the observation value;
An integrated correlation hypothesis generation unit 4 that generates hypotheses by combining the generated wakes, a hypothesis selection integration unit 5 that reduces hypotheses by deleting hypotheses with low reliability, and integrating similar hypotheses, A target type determination unit 6 that determines the target type from the attribute value of the observed value of the sensor 20 used for the estimation process, a wake file 8 that stores a wake, and a hypothesis file 9 that stores a hypothesis are provided. Yes.

  Next, the operation of the multi-target tracking device according to the second embodiment will be described with reference to the drawings. FIG. 4 is a flowchart showing the operation of the multitarget tracking apparatus according to Embodiment 2 of the present invention.

  The information obtained from the sensor 20 and the specifications estimated by the tracking device are the same as those in the first embodiment.

  First, in step 201 (observation value input), the multi-target tracking device 10 reads an observation value at the latest observation time of the sensor 20. The target position information of the observed value is taken into the position information correlation determining unit 2 and the attribute is taken into the attribute correlation determining unit 1.

  Next, in step 202 (attribute comparison determination), the attribute correlation determination unit 1 performs gate inside / outside determination of each track in the track file 8 and the observation value at the latest observation time with respect to the attribute.

The details of the method for judging inside / outside of a set of wakes and observations are described below. The predicted value (^) X _{att, k} (−) regarding the attribute is calculated by the following equation (c-1).

Here, Φ _{att, k−1} is a transition matrix related to the attribute, and (^) X _{att, k−1} (+) is a smooth value one observation time before the attribute.

The prediction error covariance P _{att, k} (+) regarding the attribute is calculated by the following equation (c-2).

Here, P _{att, k-1} (+) is the smoothing error covariance matrix related to the attribute before one observation time, Q _{att, k-1} is the driving noise covariance matrix related to the attribute, and Γ _{att, 1} (k-1). Is a transformation matrix of driving noise related to attributes.

The residual covariance matrix S _{att, k for the} attribute is calculated by the following equation (c-3).

Here, H _{att, k} is an observation matrix relating to attributes, R _{att, k} is an observation error covariance matrix relating to attributes, and Γ _{att, 2} (k) is a transformation matrix of observation noise relating to attributes.

Based on the determination using the residual covariance matrix, whether the observed value Z _{att, k} regarding the attribute is inside or outside the gate is determined by the success or failure of the condition of the following equation (c-4).

Here, Z _{att, k} (−) is a predicted observation value, and is calculated by the following equation (c-5).

D _att is a parameter of a boundary value determined by the significance level in the χ square test.

  If this expression (c-4) is satisfied (the observed value is within the range of the error covariance matrix), “there is a correlation with respect to the attribute”, and if not (the observed value is outside the range of the error covariance matrix) In this case, it is determined that “there is no correlation with respect to attributes”.

  Next, in step 203 (gate inside / outside determination), the position information correlation determination unit 2 calculates a gate related to position information of each wake in the wake file 8. This gate is an area whose spread is designated by the residual covariance, which is the sum of the wake prediction error covariance and the observation error covariance. Further, it is checked whether or not the position information of the input observation value is in the gate, and it is determined which wake the incoming observation value can be correlated with.

The details of the method for judging inside / outside of a set of wakes and observations are described below. The predicted value (^) X _{trk, k} (−) regarding the target motion is calculated by the following equation (c-6).

Here, Φ _{trk, k−1} is a transition matrix related to the target motion, and (^) X _{trk, k−1} (+) is a smooth value one observation time before the target motion.

The prediction error covariance P _{trk, k} (−) regarding the target motion is calculated by the following equation (c-7).

Here, P _{trk, k-1} (+) is a smooth error covariance matrix related to the target motion before one observation time, Q _{trk, k-1} is a drive noise covariance matrix related to the target motion, and Γ _{trk, 1} (k− 1) is a drive noise conversion matrix for the target motion.

The residual covariance matrix S _{trk, k} regarding the position of the target motion is calculated by the following equation (c-8).

Here, H _{trk, k} is an observation matrix related to the position of the target motion, R _{trk, k} is an observation error covariance matrix related to the position of the target motion, and Γ _{trk, 2} (k) is a conversion matrix of observation noise related to the position of the target motion. It is.

Using the above residual covariance _, the gate inside / outside determination of the observed value Z _{trk, k} regarding the position of the target motion is determined by the success or failure of the condition of the following equation (c-9).

Here, Z _{trk, k} (−) is a predicted observation value, and is calculated by the following equation (c-10).

Further, d _trk is a boundary value parameter determined by the significance level in the χ square test.

  If this expression (c-9) is satisfied (the observed value is within the range of the error covariance matrix), “there is a correlation with respect to the position”, and if not (the observed value is outside the range of the error covariance matrix) In this case, it is determined that “there is no correlation for the position”.

  Next, in step 204 (integrated wake generation), the integrated wake generation unit 3 sets the target of the latest observation time based on the correlation result regarding the observation value and the attribute of the wake and the correlation result regarding the observation value and the position information of the wake. A wake that estimates the motion specifications is generated. There are three types of tracks to be created: “update track”, “new track”, and “memory track track”. The motion parameters and attribute estimates at the observation time of these tracks are calculated.

  (1) The updated wake is a wake that is generated by adding the observation value that has entered the gate related to the position information of the existing wake in step 203 (gate inside / outside determination). The smooth vector is calculated by the following equations (c-11) and (c-12).

Here, K _{trk, k} and K _{att, k} are filter gains relating to position and attributes, and can be calculated by the following equations (c-13) and (c-14).

  The smooth error covariance matrix is calculated by the following equations (c-15) and (c-16).

  The likelihood of a wake is calculated as follows assuming that the probability distribution of observed values is a Gaussian distribution centered on the true value. There are two kinds of likelihood of the wake, that is, the likelihood related to position information and the likelihood related to attribute information described below, and both are calculated.

  The likelihood functioning on the position information is calculated by the following equation (c-17).

  The likelihood regarding the attribute is calculated by the following equation (c-18).

  When it is determined in step 202 (attribute comparison determination) that the observation value for the attribute is outside the wake gate, the smooth value of the attribute is the latest observation value, and the error covariance is the observation error. Set to be equivalent to covariance (only attribute is calculated with the estimated value equivalent to the following new track).

  (2) The new wake is a wake starting from the observation value obtained at the latest observation time. The smooth vector is initialized to the position component equivalent to the observed value, the velocity component is set to 0, and the attribute-related component is initialized to the observed value. In addition, the smooth error covariance matrix cannot be calculated at this time, and is calculated using the observation error covariance matrix when the second observation value is determined.

  (3) The memory track track is a track indicating that “there was no correlated observation value at the corresponding time” with respect to the existing track. The smooth vector is set to a value equivalent to the prediction vector, and the smooth error covariance matrix is set to a value equivalent to the prediction error covariance matrix.

  Next, in step 205 (target type determination), the target type determination unit 6 determines the target type for each track from the attribute value of the observation value of the sensor 20 used for the past estimation process.

  FIG. 5 is a diagram showing an example of attribute ranges that can be taken for each target type of the multi-target tracking device according to Embodiment 2 of the present invention. From FIG. 5, since all the attributes of the observation values (triangle marks in FIG. 5) used by the wake 1 belong to the target type (1), it is determined that the target type of the wake 1 is (1). In addition, since the attribute of the observation value (square mark in FIG. 5) used by Wake 2 is 1: 2 and there are more observation values belonging to the target type (2), the target type of Wake 2 is (2). judge. Furthermore, since the attributes of the observation values (circles in FIG. 5) used by the wake 3 belong to both the target types (1) and (2), the wake has a high importance for each preset target type. Select the direction.

Next, in step 206 (integrated correlation hypothesis generation), first, the integrated correlation hypothesis generation unit 4 generates a hypothesis by combining the wakes generated in step 204 (integrated wake generation). This hypothesis must satisfy the following conditions:
・ The hypothesis must be an extension of one of the existing hypotheses generated one observation time ago.
• For the combination of selected observations and existing wakes, the position information of the observations must be in the position gate of the existing wakes.
All observations are associated with any one wake or are considered unnecessary signals.
• There is at most one observation that can be correlated to one existing track.

  For each hypothesis, the reliability is calculated as the evaluation value. The reliability is calculated by calculating the product of the selected observed value and the likelihood of the wake, and further normalized so that the sum of the reliability of all hypotheses is 1.0. The calculation formula of the reliability before normalization is the following formula (c-19).

The meaning of each stage on the right side of the above formula (c-19) is as follows.
First stage: A term to be accumulated when the existing track and the position information of the observation value are correlated. However, the attribute is limited to when the correlation is obtained.
Second stage: This term is accumulated when the existing track and the observed value are correlated. However, the attributes are limited to cases where correlation cannot be obtained.
Third stage: This term is accumulated when the existing wake becomes a detection failure.
Fourth stage: A term to be accumulated when the observed value is set as a new target.
Fifth stage: A term that is integrated when an observed value is an unnecessary signal.
6th stage: The reliability of the parent hypothesis.

The following two points are used as parameters used in the second-stage calculation on the right side.
P _T : The probability that an abrupt change of the attribute occurs, and is a parameter (attribute change probability parameter) set before the apparatus is started. However, for each track, if the observed value indicates an attribute that cannot be taken by the target type determined in step 205 (target type determination), this value is set to zero. For example, if the type is determined to be type (1) in step 205 and the latest observed value is outside the range of the solid rectangle, this parameter is set to 0.
V _R is a volume in the attribute space that can be observed from the target to be tracked, and a value corresponding to the target type determination result in step 205 (target type determination) is set for each track. For example, in the example of FIG. 5, when the type of the target track is determined to be (1), V _R becomes the area of the solid line of the rectangle, it is determined the type of target to be (2) If it is in, V _R is the area of the dashed rectangle.

Next, in step 207 (hypothesis integration selection), the hypothesis selection integration unit 5 reduces the hypothesis group by deleting hypotheses with low reliability and integrating similar hypotheses. For this reduction, the following well-known method is used.
・ Set a threshold for reliability and delete all hypotheses with reliability less than that.
Set an upper limit for the number of hypotheses, leave only the set number of hypotheses in descending order of reliability, and delete others.
-Integrate hypotheses with the same correlation between observation values for the past several observation times.

  In step 208 (wake display), the wake display unit 30 selects a hypothesis corresponding to the reliability and displays the wake included in the hypothesis.

  The above is the flow of data processing for one observation time of the sensor 20, and when this one cycle ends, the observation value at the next observation time is read and the processing of the next cycle is executed.

  As described above, according to the multi-target tracking device according to the second embodiment, even if a normal correlation determination method cannot obtain a correlation between a wake and an observation value due to a large variation in attributes observed from the same target, It becomes possible to maintain the hypothesis recognized as the target with high reliability.

  In addition, since the target type is determined for each track and the reliability of the hypothesis is calculated by changing the probability of attribute variation for each model, a more accurate correlation result can be obtained.

  Furthermore, since the target type is determined for each track and the hypothesis reliability calculation is performed by limiting the volume of the possible range in the attribute space, a more accurate correlation result can be obtained.

It is a block diagram which shows the structure of the multi-target tracking apparatus which concerns on Embodiment 1 of this invention. It is a flowchart which shows operation | movement of the multi target tracking apparatus which concerns on Embodiment 1 of this invention. It is a block diagram which shows the structure of the multi-target tracking apparatus which concerns on Embodiment 2 of this invention. It is a flowchart which shows operation | movement of the multi-target tracking apparatus which concerns on Embodiment 2 of this invention. It is a figure which shows the example of the range of the attribute which can be taken for every target classification of the multi-target tracking apparatus which concerns on Embodiment 2 of this invention. It is a block diagram which shows the structure of the conventional multi-target tracking apparatus. It is a flowchart which shows operation | movement of the conventional multi-target tracking apparatus. It is a figure which shows the example of an existing track and an observed value on xy plane. It is a figure which shows two examples of a correlation hypothesis. It is a figure which shows three examples of the existing hypothesis. It is a figure which is observing 2 targets and shows transition of the position and an attribute. It is a figure for demonstrating that the false hypothesis which shows existence of a target will be made into a dominant hypothesis.

Explanation of symbols

  1 attribute correlation determination unit, 2 position information correlation determination unit, 3 integrated track generation unit, 4 integrated correlation hypothesis generation unit, 5 hypothesis selection integration unit, 6 target type determination unit, 8 track file, 9 hypothesis file, 10 multi-target tracking apparatus.

Claims (5)

A sensor that generates, as observation values, target position information and attribute information related to the target other than the position information;
An attribute correlation determination unit that performs the gate internal / external determination of each track in the track file and the observation value at the latest observation time input,
Calculate a gate related to the position information of each track in the track file, check whether the position information of the input observation value is in the gate, and determine which track the input observation value can correlate with. A position information correlation determination unit to determine;
An integrated wake generation unit that generates a wake that estimates the motion parameters of the target at the latest observation time based on the correlation result regarding the observation value and the attribute of the wake, and the correlation result regarding the position information of the observation value and the wake. When,
An integrated correlation hypothesis generator that generates a hypothesis by combining the wakes generated by the integrated wake generator, and calculates the reliability of each hypothesis using an attribute variation probability parameter that is a probability of abrupt attribute variation; ,
A hypothesis selection integration unit that deletes hypotheses with low reliability calculated by the integrated correlation hypothesis generation unit, and reduces hypothesis groups by integration of similar hypotheses;
The integrated correlation hypothesis generation unit
As the attribute change probability parameters, using the sudden change probability P _T of the attribute information, and a body volume V _R of a range in which the attribute information is possible,
From the sudden change probability _{PT of the} attribute information, the likelihood when the attribute is correlated is calculated,
Before Kitai product V _R, and it calculates the likelihood of the case it is not possible to correlate the attribute,
A multi-target tracking device that calculates the reliability of a hypothesis using a product of likelihoods in the two cases. The multi-target according to claim 1, wherein the integrated correlation hypothesis generation unit calculates the reliability of the hypothesis by weighting the attribute variation probability parameter for the sudden change of the attribute detected in the past tracking result. Tracking device. 2. A target type determination unit that determines a target type for each wake generated by the integrated correlation hypothesis generation unit from a value of an attribute of an observation value used for past estimation processing. Multi-target tracking device. The multi-target tracking according to claim 3, wherein the integrated correlation hypothesis generation unit calculates the reliability of the hypothesis by changing the attribute variation probability parameter according to the target type determined by the target type determination unit. apparatus. The multi-target tracking according to claim 3, wherein the integrated correlation hypothesis generation unit calculates a reliability of the hypothesis by changing a possible range of the attribute according to the target type determined by the target type determination unit. apparatus.

Images