JP2001099925A - Parallel multi-target tracking device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a parallel multi-target tracking device capable of conducting optimum load dispersion in parallel multi-target tracking processing even in a condition where another processing is executed on a processor for executing the processing or where another communication is carried out on a network. SOLUTION: As to a cluster generated from observation point information and predicted areas by a cluster generating part 3, when the nodes of a search tree are generated by a node dividing/integrating part 4 to be distributed plural task processing parts 5a-5d to be processed, the distribution of the processing to each of the respective task processing parts 5a-5d is assigned dynamically in response to a load condition in each of the task processing pats 51-5d by a load dispersing part 9.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for tracking a plurality of targets in a radar, which is a JPDA (Joint).
Probabilistic Data Associate
The present invention relates to a parallel multi-target tracking apparatus that employs a combination of a plurality of target objects such as a plurality of target objects and dynamically allocates processing to a plurality of task processing units to distribute a load.

[0002]

2. Description of the Related Art As a conventional tracking method for performing multi-target tracking in parallel, for example, "On Mapping a Tracking" in which processing is load-balanced by functional division.
Algorithm Onto Parallel P
processors ”(IEEE TRANSACTI
ON ON AEROSPACE AND ELECT
RONIC SYSTEMS VOL. 26, NO.
5, 1990). 5 and 6 are explanatory diagrams showing a tracking method by the conventional parallel multi-target tracking device, and show a tracking process and a processor as a graph. In this tracking method, the processing is represented as a task graph as shown in FIG. 5, and the load of each task is divided into processors based on the information flow.

FIG. 5 is a task graph showing the relationship between the processes. The task graph is a graph showing the correlation between the processes, and the processes (1) to (8) indicated by circles;
Arrow 110 indicating the connection of each of the processes (1) to (8)
To 119. Processing (1)
To (8) indicate processing itself performed by a computer or the like. In this case, tasks (1) 101 to task (8) 10
It is set to 8. Arrows 110 to 119 indicate the destinations where the calculation results in the tasks (1) 101 to (8) 108 are used.

That is, for example, if there are two processes (A) and (I), C = A + B (A) E = C + D (I)
It is represented by the arrow (a) → (i). Here,
“(A)” and “(i)” indicate a state in which “a” and “i” are circled.

[0005] All of the processing represented by the task graph can be divided and processed by a plurality of processors. In the previous example, the calculation of the process (A) and the process (I)
Can be performed by different processors. In this case, the processor that calculates the processing (i)
The calculation result of the process (a) is received from the processor that performs the calculation of the process (a). As described above, in the task graph shown in FIG. 5, tasks (1) 101 to task (8) 108 connected by arrows 110 to 119 need to exchange calculation results.

Of course, arrows 110 to 11 appear in the task graph.
There are some connected by 9 and others not connected, but those not connected by arrows 110 to 119 need not exchange data with each other. In the example shown in FIG. 5, task (1) 101 and task (2) 102 to
Task (5) 105 has arrows 110, 113, 115, 1
17, the task (6) 106 is connected to the task (2) 102 via an arrow 111, and the task (7) 10 is connected to the task (5) 105 via an arrow 118.
7 are connected. Task (6) 10
6, task (3) 103, task (4) 104, task (7) 107, and task (8)
They are directly connected via 12, 114, 116 and 119. Accordingly, pairs not connected by arrows are a pair of task (2) 102 and task (5) 105 and a pair of task (6) 106 and task (7) 107.

On the other hand, task (1) 101 and task (2)
102 pairs, task (2) 102 and task (6) 1
06, and task (5) 105 and task (7)
The pairs 107 are connected by arrows 110, 111 or 118. For example, if the task (1) 101 and the task (2) 102, the calculation of the task (2) 102 is also performed unless the calculation of the task (1) 101 is completed. The relationship is not over.

At this time, if there are two types of divisions shown below, the latter division is selected. This is because in the former method, it is necessary to pass the calculation result between the processors, but in the latter method, the necessity does not occur.

Task (2) 102 and Task (5) 1
05 to one processor, task (6) 10
6 and task (7) 107 are assigned to another processor. Assign task (2) 102 and task (6) 106 to one processor, and assign task (5) 105 and task (7) 107 to another processor.

Further, the required amount of calculation is usually shown for each of the tasks 101 to 108 (omitted in FIG. 5). Therefore, when dividing the processing given to the task graph into a plurality of processors, the processing is performed based on the following two policies.

The tasks are allocated to the respective processors so that the total required calculation amount of the tasks is the same as much as possible.

The tasks 101 to 108 connected by arrows 110 to 119 are assigned to the same processor.

FIG. 6 is a processor graph showing an example in which four processors (P1) 121 to (P2) 124 are connected in a ring. That is, the processor (P
1) The processor (P2) 122 is connected to 121, the processor (P3) 123 is connected to the processor (P2) 122, and the processor (P4) is connected to the processor (P3) 123.
The processor (P1) 121 is connected to the processor (P4) 124.

As described above, tasks 101 to 108 connected by arrows 110 to 119 are connected to the same processor 121.
Assigning to .about.124 is a tracking method used in a conventional parallel multi-target tracking apparatus. In the task graph shown in FIG. 5, the tasks 101 to 108 represented by circles are processes that are functionally grouped. JP
When the tracking process is performed in consideration of the combination of the target and the observation point, such as DA, if the target problem becomes large, the task of obtaining the combination of the target and the observation point is more difficult than the other tasks. , The processing load becomes very large.

[0015]

Since the conventional parallel multi-target tracking apparatus is configured as described above, in a method of performing tracking processing in consideration of a combination of a plurality of targets and observation points like JPDA. In some cases, the processing load of one specific task 101 to 108 for obtaining a combination of a target and an observation point becomes extremely large. In this case, as described above, the plurality of processors 121 ~ 124
Tasks 101 to 10 whose load becomes extraordinarily large even when parallel processing of tasks 101 to 108 is performed using
Only the processors 121 to 124 that have performed the processing of No. 8 do not finish the processing, and as a result, it takes a long time until the entire processing is completely completed, and the load distribution effect using the plurality of processors 121 to 124 is obtained. However, there is a problem that the meaning of performing the parallel processing is lost.

Tasks 101 to 108 for obtaining a combination of a target and an observation point are divided so that the processing load is as equal as possible, and the tasks are separated into separate processors 121 to 124.
It is possible to process with. However, when each of the processors 121 to 124 operates on a general-purpose workstation, a task 10 for determining a combination of a target and an observation point is required.
Processes other than 1 to 108 and processes of other users may also be executed on the same processor 121 to 124, and these other processes are not necessarily constant in each processor 121 to 124. There has been a problem that the execution times of the tasks 101 to 108 processed by the respective processors 121 to 124 are not uniform, and the load cannot be optimally distributed.

Even if the processors (workstations) 121 to 124 do not perform any processing other than the tasks 101 to 108 for obtaining the combination of the target and the observation point, When connected via a network, tasks 101 to
It is conceivable that processors other than the processors 121 to 124 that perform the processing of 108 share the same network, and in such a case,
Since the communication time between the processors 121 to 124 does not become constant, the processing of the tasks 101 to 108 in each processor 121 to 124 does not become constant, and there is a problem that the load is not optimally distributed. Was.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and performs load distribution on a task that has a particularly high load, such as a task for obtaining a combination of a target and an observation point, and performs task distribution. Even if other processing is running on the processor that executes the processing or other communication is being performed on the network, the parallel multi-target tracking can perform optimal load balancing in the parallel multi-target tracking processing. The aim is to obtain a target tracking device.

[0019]

A parallel multi-target tracking apparatus according to the present invention generates a node of a search tree in a node dividing / integrating section for a cluster generated from observation point information and a prediction area by a cluster generating section. When distributing it to a plurality of task processing units as a task and processing it, the load distribution unit distributes the processing of the node to each task processing unit according to the load situation of each task processing unit. It was done.

The parallel multi-target tracking apparatus according to the present invention periodically monitors the load status of each task processing unit, and moves the load from the heavily loaded task processing unit when the load on the task processing unit decreases. Thus, the load on each task processing unit is equalized.

The parallel multi-target tracking apparatus according to the present invention distributes node processing to each task processing unit one unit at a time, and the task processing unit that has completed processing of the distributed node transmits the next node distribution. Is received and the processing of a new unit of processing is distributed to distribute the load.

In the parallel multi-target tracking apparatus according to the present invention, the processing of a plurality of units of nodes is first distributed to each task processing unit from the load distribution unit, and each task processing unit distributes the received node among the received nodes. When the processing of one section is started and the rest is held, and when the processing of the section ends, the distribution of the next section is requested to the load distribution unit, and the distribution is held before the response to the request is returned. The processing of the next unit is started, and the load distribution unit that receives the distribution request distributes the next node to the task processing unit that has requested the distribution of the node. is there.

In the parallel multi-target tracking apparatus according to the present invention, the unit of load distribution is reduced with time.

The parallel multi-target tracking apparatus according to the present invention
When the processing time of one load balancing unit exceeds a certain time,
The processing of a part of the group of nodes under the processing node is returned to the load distribution unit.

In the parallel multi-target tracking apparatus according to the present invention, the upper limit of the processing time when it is determined that the processing of some of the nodes is returned to the load distribution unit is reduced with time. is there.

In the parallel multi-target tracking apparatus according to the present invention, one of the task processing units is operated on the same processor as the load distribution unit.

[0027]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of a parallel multi-target tracking apparatus according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes an input line for inputting observation point information from the outside. Reference numeral 2 denotes a prediction area generation unit that predicts movement of a target and generates a prediction area according to the prediction result. Three
Is a cluster generation unit that generates a cluster from the observation point information input from the input line 1 and the prediction region generated by the prediction region generation unit 2. Reference numeral 4 denotes a node division / integration unit that generates a node of the search tree for the cluster generated by the cluster generation unit 3 and performs division / integration using the node as a task. 5a ~
Reference numeral 5d denotes a task processing unit that performs processing for each task (section) divided by the node division / integration unit 4. Reference numeral 6 denotes task processing units 5a to 5d integrated by the node division / integration unit 4.
Is a target track update unit that updates information (track) of the target based on the processing result of (1). Reference numeral 7 denotes an output line for outputting the information updated by the target object track updating unit 6 to the outside as a tracking result. Reference numerals 8a to 8d denote transfer lines used for transferring the processing results of these units.

Reference numeral 9 denotes a load distribution unit which distributes the processing for the nodes divided by the node division / integration unit 4 to the task processing units 5a to 5d based on the respective load conditions. 8f is a transfer line connecting the load distribution unit 9 and the node division / integration unit 4, and 8g is a transfer line connecting the load distribution unit 9 and each of the task processing units 5a to 5d.

Next, the operation will be described. The cluster generation unit 3 receives observation point information from the outside via the input line 1,
The predicted area created by the predicted area generation unit 2 is the transfer path 8a.
Respectively. When receiving the observation point information and the prediction region, the cluster generation unit 3 combines the prediction regions when the prediction regions of the plurality of targets overlap and the observation point exists in the overlapping region, and forms one cluster. Generate. The information of one cluster (observation point and predicted area of the target) generated by the cluster generation unit 3 is sent to the node division / integration unit 4 via the transfer path 8b, and the node division / integration unit 4 receives the information. A node of the search tree is generated from the information of the cluster.

The load distribution unit 9 fetches the node generated by the node division / integration unit 4 via the transfer path 8f and, depending on the load status of each of the task processing units 5a to 5d, transfers the processing of the node to the transfer path 8g. Are distributed to the respective task processing units 5a to 5d via. Each of the task processing units 5a to 5d obtains a leaf of the search tree (corresponding to one of combinations of tracking) from the received nodes, and sends the processing result to the node division / integration unit 4 via the transfer paths 8g and 8f. The node division / integration unit 4 integrates the processing results sent from the task processing units 5a to 5d and sends them to the target track update unit 6 via the transfer path 8c. The target track update unit 6 has information on the target and updates the information on the target based on the possibility of each sent combination.
The information on the target updated by the target track update unit 6 is output to the outside via the output line 7 and is also sent to the prediction area generating unit 2 via the transfer line 8d. The prediction area generation unit 2 performs movement prediction for each target based on the information on the target updated by the target track update unit 6,
The prediction area is generated and sent to the cluster generation unit 3 via the transfer line 8a. This device operates as a parallel multi-target tracking device using dynamic load balancing by repeating the above operation.

The operation of each unit will be described below, taking the case of using JPDA as an example. Here, FIG. 2 is an explanatory diagram showing a relationship between a target and an observation point in JPDA, FIG. 3 is an explanatory diagram showing a matrix Ω representing a relationship between the target and an observation point, and FIG. It is explanatory drawing which shows the search tree showing the combination. In FIG. 2, t (1) and t (2) are targets, and an ellipse indicated by a black circle and centered on predicted coordinates of the targets t (1) and t (2) is a prediction range. Also, y
(1) to y (4) are observation points.

Note that the matrix Ω shown in FIG. 3 has targets t (0) to t (2) on the vertical and observation points y (1) to y on the horizontal.
The target object t (0) indicates that there is no corresponding target. In the matrix Ω, “1” indicates a possibility and “0” indicates no possibility, and the matrix Ω indicates the following contents. Observation point y (1) has no target t (2) or corresponding target. Observation point y (2) has target t (1) or target t (2) or no corresponding target. -Observation point y (3) has no target t (1) or no corresponding target.-Observation point y (4) has no target t (1) or no corresponding target.

There are the following cases as a combination of the target and the observation point satisfying the above. The target t (1) is at the observation point y (2) and the target t
(2) is the observation point y (1) • The target t (1) is the observation point y (2) and the target t
(2) No response ・ Target t (1) is observation point y (3) and target t
(2) is the observation point y (1). • The target t (1) is the observation point y (3) and the target t
(2) is the observation point y (2). • The target t (1) is the observation point y (3) and the target t
(2) No response ・ Target t (1) is observation point y (4) and target t
(2) is the observation point y (1).-The target t (1) is the observation point y (4) and the target t
(2) is the observation point y (2) • The target t (1) is the observation point y (4) and the target t
(2) does not correspond ・ Target t (1) has no correspondence, target t (2) has observation point y (1) ・ Target t (1) has no correspondence and target t (2) has no correspondence Observation point y (2)-Target t (1) does not correspond, target t (2) does not correspond

FIG. 4 shows this combination as a search tree. Each name in this search tree corresponds to the following part in FIG. 4, and the combination corresponds to the state of each node from the root to the leaf.・ Root → Root ・ Knot → Middle of tree structure branch, Fig. 4
Ellipse in the figure ・ Leaf → Located at the tip of the tree structure branch, Fig. 4
・ Branches → Lines connecting roots and nodes, nodes and nodes, and nodes and leaves in Fig. 4 ・ Nodes → concept including both nodes and leaves ・ Brothers → Number of branches passing from root to node Equal nodes

Based on the above, the operation of each unit will be specifically described. First, the prediction area generation unit 2 calculates the targets t (1) and t (1) corresponding to the time at which the observation point was observed, based on information (coordinates, speed, and the like) of the target held therein.
The predicted coordinates of (2) are calculated. Similarly, a prediction area (ellipse in FIG. 2) that is equal to or more than a certain existence probability and is centered on the prediction coordinates of the target objects t (1) and t (2) is calculated.
Next, the cluster generation unit 3 calculates the prediction area and the observation point y.
The inclusion relationship with (1) to y (4) is examined. When there are observation points included in a plurality of prediction regions, clusters are formed by combining the prediction regions including the observation points. In the example shown in FIG. 2, since the observation point y (2) is included in the two prediction regions, the two prediction regions are combined as one cluster.

The generated cluster is sent to the node division / integration unit 4, which generates the nodes of the search tree shown in FIG. The load distribution unit 9 is a node division / integration unit 4
Captures the clauses generated by the task processing sections 5a to 5d, and processes the clauses according to the load conditions of the task processing sections 5a to 5d.
Distribute to 5d. Here, the level of the node distributed to the task processing units 5a to 5d may be the observation point y (1), y (2), or y (3) in FIG. . It should be noted that the processing amount for each node usually decreases as the level becomes lower. Each task processing unit 5a to 5d
Calculates the leaves from the nodes distributed by the load distribution unit 9, calculates the reliability for each combination indicated by each leaf, and weights the correspondence between the target and each observation point.

For example, in FIG. 4, the task processing unit that receives the clause 10 calculates the following reliability E (1) to E (3) as the reliability for each combination. E (1) → Observation point y (1) does not correspond, observation point y
(2) The reliability of the case where the target t (2) and the observation points y (3) and y (4) do not correspond E (2) → the observation point y (1) does not correspond and the observation point y
(2) the target t (2), the observation point y (3) does not correspond,
The reliability of the case where the observation point y (4) is the target t (1) E (3) → the observation point y (1) does not correspond, the observation point y
(2) is the target t (2), and the observation point y (3) is the target t.
(1), the reliability of the case where the observation point y (4) does not correspond

Next, the task processing units 5a to 5d perform the following weighting on the correspondence between the target and each observation point in the node 10 shown in FIG. 4 based on the calculated reliability values. . The target object t (1) has a weight corresponding to the observation point y (3), E (3). The target object t (1) has a weight corresponding to the observation point y (4), E (2). The weight at which the object t (1) does not correspond is E (1). The weight at which the target t (2) corresponds to the observation point y (2) is E (1) + E (2) + E (3).

Next, the node division / integration section 4 compiles the "weight of the correspondence between the target and each observation point" calculated by each of the task processing sections 5a to 5d. For example, in the case of the target t (2), the weight corresponding to the observation point y (1) and the observation point y
The weight corresponding to (2) and the weight not corresponding are obtained.

The target track update unit 6 receives the information obtained by the node division / integration unit 4 after integration, calculates the position of the target, and updates the internal information such as the coordinates and speed of the target.
For example, in the case of the target t (2), the coordinates of the target t (2) are obtained based on the above three weightings. In addition, the calculation result based on the weight is regarded as the coordinates of the target. Thus, the operation of each unit is completed.

Next, in the task processing units 5a to 5d,
The distribution of nodes according to each load situation will be described. First, the load distribution unit 9 equally distributes the processing of all the nodes of a certain level of the search tree generated by the node division / integration unit 4 to the task processing units 5a to 5d. As a method of distributing the processing of the nodes evenly, it is conceivable to equalize the number of nodes or equalize the number of lower leaves of the nodes.

Next, the load distribution unit 9 issues a load report request to each of the task processing units 5a to 5d via the transfer line 8g, and each of the task processing units 5a to 5d distributes the load status in response to the request. Report to Part 9. Here, the load status includes the length per unit time of the job queue managed by the operating system, that is, the number of jobs in the job queue, or each of the task processing units 5a to 5a.
The number of nodes and leaves not yet processed by 5d may be considered.

Next, the load distribution unit 9 is connected to each task processing unit 5a.
-5d is requested, based on the information, to move the nodes to the task processing units 5a-5d so that the loads of the task processing units 5a-5d are equalized. Here, for example, the load status is stored in each of the task processing units 5a to 5
d, the number of unprocessed nodes is equal to
The load status reports from 5d are 3, 5, 4, and 4, respectively.
It is assumed that From this, all task processing units 5
The number of unprocessed nodes in a to 5d is 3 + 5 + 4 + 4 = 1
Since the load state of the four task processing units 5a to 5d is equalized, the number of nodes to be assigned to each of the task processing units 5a to 5d is four. Therefore, the load distribution unit 9 instructs the task processing unit 5b to move the processing of one node to the task processing unit 5a via the transfer line 8g. As a result, the number of unprocessed nodes in each of the task processing units 5a to 5d becomes four.
One by one. The collection of the load status and the movement of the nodes between the task processing units 5a to 5d are periodically performed.

As described above, according to the first embodiment, the load distribution unit 9 distributes node processing to the task processing units 5a to 5d according to the load status of each of the task processing units 5a to 5d. Therefore, even when each of the task processing units 5a to 5d operates on a machine on which other processing is performed, such as a general-purpose workstation, the processing time of each of the task processing units 5a to 5d becomes equal, An effect that optimal load distribution can be performed is obtained.

Since the number of unprocessed nodes in each of the task processing units 5a to 5d is regularly equalized, processing other than the processing of the nodes is performed on the same processor as the task processing units 5a to 5d. However, the processing of all the nodes of each of the task processing units 5a to 5d is completed in substantially the same time, and an effect that optimum load distribution can be obtained.

Embodiment 2 In the first embodiment,
The load distribution unit 9 takes the initiative in the timing of load distribution and periodically equalizes the loads of the task processing units 5a to 5d, so that the processing of the nodes in each of the task processing units 5a to 5d is performed for substantially the same time. However, the task processing units 5a to 5d may take the initiative in the timing of load distribution, and make a request to transfer the work to the load distribution unit 9. The second embodiment describes such a parallel multi-target tracking apparatus, and its configuration is the same as that of the first embodiment shown in FIG.
Here, the description is omitted.

Next, the operation will be described. Here, a description will be given of portions different from the first embodiment.
First, the load distribution unit 9 distributes the processing of the nodes of a certain level in the search tree generated by the node division / integration unit 4 to the task processing units 5a to 5d one by one. Each task processing unit 5a to 5d
Then, when the node of the search tree is received, the processing is started.
When the processing of the received node is completed and the load is reduced, each of the task processing units 5a to 5d requests the load distribution unit 9 for the next job, that is, transfer of the node to be processed. Upon receiving the transfer request, the load distribution unit 9 gives the processing (load) of one unit of the node to the task processing units 5a (to 5d) that issued the request. The above processing is repeated until all the nodes are gone.

As described above, according to the second embodiment, each of the task processing units 5a to 5d issues a transfer request for the next work to the load distribution unit 9 when there is no more work. Is performed on the same machine, the processing of all the nodes of each of the task processing units 5a to 5d is completed in substantially the same time, and an effect that optimal load distribution can be performed is obtained. Even when the processing amount corresponding to the processing is small, there is no need to periodically collect the load status, so that the effect of reducing the overhead of load distribution can be obtained.

Embodiment 3 In each of the above embodiments, the case has been described where the task processing units 5a to 5d issue a transfer request for the processing of the next node to the load distribution unit 9 after the processing of one node is completed. It is also possible to secure a certain amount of work in 5d and sequentially process it. The third embodiment describes such a parallel multi-target tracking apparatus. Since the configuration is the same as that of the first embodiment shown in FIG. 1, the description is omitted here.

Next, the operation will be described. Note that, in this case as well, portions different from those in Embodiment 1 will be described. First, the load distribution unit 9 first distributes the processing of the node at a certain level of the search tree generated by the node division / integration unit 4 to each of the task processing units 5a to 5d by a plurality of units. When each of the task processing units 5a to 5d receives the plurality of nodes, the task processing units 5a to 5d hold the remaining nodes except for one unit, and execute the processing of the remaining one unit node that is not held. Each task processing unit 5a-
When the processing of the node is completed, 5d requests the load distribution unit 9 for the next job, that is, transfer of the node to be processed next. Before a response to the transfer request is returned from the load distribution unit 9, one unit of the stored nodes is extracted and the processing is started. When the load distribution unit 9 receives a transfer request for a node to be processed next from the task processing unit 5a (負荷 5d), the task processing unit 5a (〜
For 5d), the processing of the next section is distributed. Further, upon receiving the distribution of the node to be processed next from the load distribution unit 9, the task processing unit 5a (to 5d) which issued the request holds the distribution. The above processing is repeated until all the nodes are gone.

As described above, according to the third embodiment, when the task processing units 5a to 5d run out of work, they issue a transfer request for the next work to the load distribution unit 9 and hold the job. Since the processing of the set section is started immediately, the processing of the other section is executed even after issuing the transfer request for the next job to the load distribution unit 9 and receiving the response, and Since there is almost no time when the processing units 5a to 5d are not performing the processing of the node, an effect that optimal load distribution can be performed can be obtained.

Embodiment 4 FIG. In each of the above embodiments, the case where the unit of load distribution is fixed has been described. However, the unit of load distribution may be reduced with time. The fourth embodiment describes such a parallel multi-target tracking apparatus. Since the configuration is the same as that of the first embodiment shown in FIG. 1, the description is omitted here.

Here, the unit of load distribution in the parallel multi-target tracking apparatus is a node, and the size corresponding to the processing amount can be measured, for example, by the number of leaves below the node. FIG.
As can be seen from the search tree shown in (1), the number of leaves increases as the level of the node increases, and decreases as the level decreases.
Therefore, when distributing the load, it is easy to generate a work having a smaller size by decomposing the nodes and using the lower-level nodes as units of the load distribution.

Here, the smaller the size of the unit of the load distribution, the greater the number of times the load is distributed, so that the overhead for the load distribution increases, but the end of the processing is uniformly performed by the task processing units 5a to 5d. The smaller the better, the better. Therefore, at first, the high-level clause is used as the unit of load distribution, the clause is decomposed as the processing progresses, and the lower-level node is used as the unit of load distribution, so that the load distribution overhead is suppressed and each task Processing units 5a to 5
It is possible to equalize the end of the processing of the node d and perform an optimal load distribution.

When the unit of load distribution is reduced with time, each task processing unit 5a is used as in the first embodiment.
5d hold work (that is, processing of a clause), the task processing units 5a to 5d decompose the clause. Further, when the load distribution unit 9 holds the processing of a node as in the second embodiment, the load distribution unit 9 decomposes the node, and as in the third embodiment, the load distribution unit 9 and each task processing unit 5a. ~
When the processing of the node is held in 5d, the load distribution unit 9
And the task processing units 5a to 5d decompose the clause.

As described above, according to the fourth embodiment, since the unit of load distribution is reduced with time, the overhead of load distribution is suppressed, and the processing of the nodes of the respective task processing units 5a to 5d is performed. An effect is obtained in which the end of the processing can be made uniform and the optimum load distribution can be performed.

Embodiment 5 In the fourth embodiment, the case where the unit of load distribution is reduced with time has been described. However, regardless of time, when the processing time of one load distribution unit becomes longer than a certain value, the load distribution unit is reduced. You may do so. The fifth embodiment describes such a parallel multi-target tracking apparatus. Since the configuration is the same as that of the first embodiment shown in FIG. 1, the description is omitted here.

When the processing of a node at the same level in the search tree is used as a unit of load distribution, the number of lower leaves in some nodes is not uniform. Therefore, if the number of leaves under a certain node is very large, the processing of that node takes a very long time. In the worst case, the processing of all other nodes has been completed, but the processing of that node has been completed. It is possible that only processing remains.

Therefore, each of the task processing units 5a to 5d monitors the processing time of a node, and while the processing of a certain node is being performed, the execution time from the start of the processing of the node is fixed. If the time has elapsed, a part of the unprocessed node group under the node is returned to the load distribution unit 9. Then, as in the case of the first embodiment, when each of the task processing units 5a to 5d holds work (processing of a node), the node group returned until the next load distribution timing is held. At the time of the next load distribution, the processing of the node group is distributed to the task processing units 5a to 5d having a low load. Embodiment 2
When the load distribution unit 9 holds the processing of the clause as in the case of (3), the returned clause group is held as it is and treated in the same way as the clause already held, and Similarly, in the case where the load distribution unit 9 and each of the task processing units 5a to 5d hold the processing of a clause, the returned clause group is retained as it is and handled in the same manner as the clause already retained.

As described above, according to the fifth embodiment, a part of the processing of a node which requires a long processing time is returned to the load distribution unit 9, so that another task processing unit having a sufficient processing capacity can be provided. Since the processing of the node groups is distributed to 5a to 5d, the load is distributed, the processing is completed as a whole quickly, and the effect that the optimal load distribution can be performed is obtained.

Embodiment 6 FIG. In the fifth embodiment, a description will be given of a case where the upper limit value of the processing time of one load distribution unit for determining that the task processing units 5a to 5d return the processing of the unprocessed node group to the load distribution unit 9 is fixed. However, the upper limit may be reduced with time. The sixth embodiment describes such a parallel multi-target tracking apparatus. Since the configuration is the same as that of the first embodiment shown in FIG. 1, the description is omitted here.

When the task processing units 5a to 5d return the processing of the unprocessed node group to the load distribution unit 9, the upper limit of the processing time of one load distribution unit determined to be returned is large.
When it is determined to return to the load distribution unit 9, there is a risk that the processing of the other task processing units 5a to 5d has already been completed, and if the processing is small, the processing of the other task processing units 5a to 5d may not be completed for the time being. Regardless, the processing of the unprocessed node group is performed by the load distribution unit 9.
There is a danger that the work will be wasted.

Therefore, initially, the upper limit value for determining that the processing of the unprocessed node group is returned to the load distribution unit 9 is increased, and the upper limit value is reduced over time. As a result, useless return of the node group to the load distribution unit 9 can be suppressed by a large upper limit value at first. Further, at the end of the entire process, a part of the work of the task processing units 5a to 5d which take a long time to process before the other task processing units 5a to 5d complete the process due to the small upper limit value. The load can be distributed to the other task processing units 5a to 5d, and the optimum load distribution can be achieved.

As described above, according to the sixth embodiment, since the upper limit value of the load amount when returning a node group to the load distribution unit 9 is reduced with time, it is useless at first. The return of the node group can be suppressed, and in the last part of the processing, the task processing units 5a to 5
A part of the task d can be surely distributed to the other task processing units 5a to 5d, so that an effect that optimal load distribution can be performed can be obtained.

Embodiment 7 FIG. In the above embodiments, the case where each of the task processing units 5a to 5d and the load distribution unit 9 are operated on different processors has been described. You may make it operate on the same processor as the part 9. The load at the time of performing the processing of the node is the task processing units 5a to 5d.
And the load distribution unit 9 has almost no load. Therefore, by operating the load distribution unit 9 on the same processor as one of the task processing units 5a to 5d, the load of all the processors can be increased.

As described above, according to the seventh embodiment, the load distribution unit 9 is operated on the same processor as one of the task processing units 5a to 5d. And the effect of enabling parallel processing without waste is obtained.

[0067]

As described above, according to the present invention, when a load distribution unit is provided and nodes of a search tree generated for a cluster are distributed to a plurality of task processing units and processed, each task processing unit Is configured to distribute the processing of the nodes according to the load status of each task processing unit, so that even if each task processing unit operates on a processor that is executing other processing, the processing time of each task processing unit can be reduced. There is an effect that a parallel multi-target tracking device that can equalize and perform optimal load distribution can be obtained.

According to the present invention, the load status of each task processing unit is periodically monitored, and when the load is reduced, the processing of the node is moved from the task processing unit with the heavy load. The number of unprocessed clauses in the processing unit is regularly equalized, and even if processing other than the processing of clauses is performed on the same processor as the task processing unit, processing of all clauses in each task processing unit is performed. The processing is completed in substantially the same time, and there is an effect that optimum load distribution can be performed.

According to the present invention, the processing of a node is distributed to each task processing unit by one unit, and when the processing of the distributed node is completed, the task processing unit requests the distribution of the next node and newly allocates the next node. Since the processing of a single node is distributed, even if processing other than the processing of the node is performed on the same processor, the processing of all the nodes of each task processing unit takes about the same time. Finished, allowing for optimal load balancing,
Even when the processing amount corresponding to the processing of one node is small, there is no need to periodically collect the load status, so that the load distribution overhead is reduced.

According to the present invention, the processing of a plurality of nodes is first distributed to each task processing unit from the load distribution unit, and each task processing unit starts processing of one unit of the nodes. Then, the rest is held, and when the processing of the section is completed, the distribution of the next section is requested to the load distribution unit,
Since the processing of the next one unit held before the response to the request is returned is started, when each task processing unit runs out of work, each task processing unit transmits a transfer request of the next work to the load distribution unit. And the processing of the held clause is started immediately, so it is retained from the time the request for transfer of the clause to the load balancer is issued until the transfer of the corresponding clause is received. There is an effect that the time when the task processing unit is not performing the processing of the clause is almost eliminated when the processing of the other clause is executed, and the optimal load distribution can be performed.

According to the present invention, since the unit of the load distribution is made smaller with time, the overhead of the load distribution is suppressed and the end of the processing of the node in each task processing unit is made uniform. This has the effect of enabling optimal load distribution.

According to the present invention, when the processing time of one load distribution unit becomes equal to or more than a certain value, the processing of a part of the subordinate clause groups of the processing node is returned to the load distribution unit. By load distribution, it is possible to distribute the processing of these nodes to other task processing units that have extra processing capacity, and the processing is completed early as a whole, so that optimal load distribution is possible. effective.

According to the present invention, the upper limit value of the processing time when it is determined that the processing of some of the nodes is to be returned to the load distribution unit is configured to decrease with time. With a large upper limit, unnecessary return of nodes to the load balancer can be suppressed.At the end of the overall processing, a small upper limit allows processing to be performed before the work of other task processing units is completed. Since part of the time-consuming work of the task processing unit can be distributed to other task processing units, there is an effect that optimum load distribution can be achieved.

According to the present invention, since one of the task processing units and the load distribution unit are operated on the same processor, the load can be made high in all the processors, and the wasteful operation can be achieved. There is an effect that no parallel processing is possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a parallel multi-target tracking apparatus according to the present invention.

FIG. 2 shows JP in the parallel multi-target tracking apparatus of the present invention.
It is explanatory drawing which shows the relationship of the target and observation point by DA.

FIG. 3 is an explanatory diagram showing a matrix Ω representing a relationship between a target and an observation point in the parallel multi-target tracking apparatus of the present invention.

FIG. 4 is an explanatory diagram showing a search tree representing a combination of a target and an observation point in the parallel multi-target tracking apparatus of the present invention.

FIG. 5 is an explanatory diagram showing a task graph of a tracking process in a conventional parallel multi-target tracking device.

FIG. 6 is an explanatory diagram showing a processor graph of tracking processing in a conventional parallel multi-target tracking device.

[Explanation of symbols]

1 input line, 2 prediction region generation unit, 3 cluster generation unit, 4 node division / integration unit, 5a to 5d task processing unit,
6 Track update section, 7 output lines, 8a-8d, 8
f, 8g transfer path, 9 load balancer, 10 clauses, t
(1), t (2) target, y (1) to y (4) observation points.

Claims (8)

[Claims] 1. A parallel multi-target tracking apparatus that tracks a plurality of targets by distributing loads on a plurality of task processing units, predicts the movement of the targets, and sets a prediction area according to the prediction result. A prediction region generation unit to generate; a cluster generation unit to generate a cluster from the input observation point information; and a prediction region generated by the prediction region generation unit; and a search for a cluster generated by the cluster generation unit. A node division / integration unit that generates a node of a tree and divides / integrates the node as a task; a plurality of task processing units that perform processing on each of the nodes divided by the node division / integration unit; A load distribution unit that dynamically assigns processing to the nodes divided by the division / integration unit based on the load status of each of the task processing units; and a processing result of the node division / integration unit. A parallel update comprising updating the information of the target, outputting the result as a tracking result, and inputting the target track update unit to the prediction area generating unit for predicting the movement of the target. Multi-target tracking device. 2. A load distribution unit first distributes node processing to each task processing unit so that the loads of a plurality of task processing units are equalized, and thereafter periodically loads the task processing units. By monitoring the situation and moving the processing of the node from the task processing unit with a large load to the task processing unit with a small load, the load distribution is performed while equalizing the load of each task processing unit. The parallel multi-target tracking apparatus according to claim 1. 3. The load distribution unit distributes the processing of a node to each task processing unit that processes a node, one unit at a time. When the processing of the distributed node ends, the task processing units Requesting the distribution of nodes to the load distribution unit; upon receiving the distribution request of the next node, the load distribution unit distributes one unit of node to the task processing unit that requested the distribution of the node. 2. The parallel multi-target tracking apparatus according to claim 1, wherein 4. A load distribution unit first distributes processing of a plurality of nodes to each task processing unit that processes a node, and each of the task processing units is distributed by the load distribution unit. After processing the remaining one-unit clauses while retaining the multiple-unit clauses except for one unit, when the processing of the corresponding clause is completed, a request is made to the load distribution unit to distribute the next clause, and When the load distribution unit receives the distribution request of the next node, it starts processing of the next unit of the node held before the response to the request is returned. 2. The parallel multi-target tracking apparatus according to claim 1, wherein the following sections are distributed to the sections. 5. The parallel multi-target tracking apparatus according to claim 1, wherein the unit of load distribution is reduced with time. 6. When the processing time of one load distribution unit in the task processing unit becomes longer than a predetermined time, processing of a part of the lower-order clause group of the clause being processed by the task processing unit is performed. 2. The parallel multi-target tracking apparatus according to claim 1, wherein the information is returned to a dispersing unit. 7. An apparatus according to claim 6, wherein the upper limit value of the processing time of one load distribution unit for determining that the task processing unit returns the processing of the node group to the load distribution unit is reduced with time. Parallel multi-target tracking device. 8. The parallel multi-target tracking apparatus according to claim 1, wherein one of the plurality of task processing units is operated on the same processor as the load distribution unit.