The parallel track initiation method of region rasterizing Multiple feature association based on Hough transformationTechnical field
The track initiation method based on Hough transformation that the present invention relates to a kind of.
Background technique
Track initiation, which refers to the process of, establishes track before not entering the track maintenance stage, be primarily asking for multiple target trackingTopic, directly affects the performance of multiple target tracking.Common track initiation method can be divided into sequential processes method and batch processingTwo kinds of method.Hough transformation is as a kind of typical batch processing method, because it is to the insensitive, to random noise of local defectRobustness and it is suitable for parallel processing, the in real time good characteristics such as application, it has also become the hot spot of current track initiation technique study.
The early 1980s, Smith and Winter for the first time introduce Hough transformation, as a kind of effective trackInitial mode.1994, Carlson, Evans and Wilson etc., from principle concept, accumulation method and starting effect isogonismDegree, application of the detailed analysis based on the initial mode of Hough transformation in single search radar.1998, Yankowich was moreHough transformation has been used in target multisensor track initiation blending algorithm.
Although many fruitful work have been made in forefathers, there is also all in a particular application for Hough transformationMore problems.Firstly, there is presently no a kind of choosing methods of generally acknowledged effective parameter space quantized interval；Secondly, as oneKind batch processing method, the naturally occurring calculation amount of Hough transformation, the huge defect of amount of storage are necessarily required to consider in practical applicationsReal time problems.
Summary of the invention
The purpose of the present invention is to provide a kind of, and the parallel track of region rasterizing Multiple feature association based on Hough transformation risesBeginning method.
The technical solution for realizing the aim of the invention is as follows: first according to the maximum speed v of expectation targetmaxAnd scanningThe minimum range span DivDis=10*v in cycle T s computation partition regionmax* Ts is set in advance if calculated DivDis is less thanThe minimum range span M set, then take DivDis=M.Search coverage in radar horizon R, which is divided into X span, isDivDis+δdOverlapped zonule, δdIt is the distance of overlapping region,For each cellDomain, according to observing environment reference cycle number Taging(reference cycle number refers to number effective period for participating in ballot accumulation) and peakValue extracts thresholding Thr, goes out Hough transformation minimum quantization interval delta θ and Δ ρ according to target error in measurement off-line calculation respectively.By thisX region carries out track initiation using the method for Hough transformation according to the parameter in respective region parallel.
For the track initiation method itself based on Hough transformation, parallel processing is also carried out.By parameter space cumulant matrixIt is divided into N number of submatrix, N is determined by computing resource.Each submatrix individually creates task, and multiple submatrix parallel processings are appointedThe synchronously control of business is realized by the way of semaphore+message queue.
Compared with prior art, the present invention its remarkable advantage are as follows:
Entire search coverage is divided into multiple zonules by the present invention, uses different reference cycle numbers to different regionsTagingWith peak extraction thresholding Thr, quickly starting can be taken into account and inhibit false.It is miscellaneous due in entire radar observation regionWave power is different, and the prior art all uses identical reference cycle number T in whole regionagingWith peak extraction thresholding Thr, ifUsing higher TagingWith Thr, if then can not achieve quick starting in clutter region, according to lower TagingWith Thr,It will lead to again in strong clutter region false too many.The partition method that the present invention uses uses high threshold in strong clutter regionTo inhibit false, quickly starting is realized using low threshold in weak clutter region.Secondly, in the prior art, in entire detecting areaIn domain, same group of quantized interval is taken to be handled.Since the error in measurement of radar target is observed in polar coordinate system,It is transformed into rectangular coordinate system, it is big the small remote error range of short range error range necessarily occurs.The present invention is smaller in short range useQuantized interval, remotely use larger quantization interval, be more in line with actual conditions.
The present invention has also carried out parallel processing to Hough transformation itself while multizone is parallel, can be more fullyUsing computing resource, real-time is improved.Compared with prior art, more inputs be can handle in interval at the same timePoint mark, can also accomplish to handle in real time under strong clutter environment.
Detailed description of the invention
Fig. 1 parameter space cumulant matrix structure chart.
Fig. 2 submatrix parallel processing schematic diagram.
Specific embodiment
The parallel track initiation method of region rasterizing Multiple feature association based on Hough transformation is realized in accordance with the following steps.
(1) region division: according to the maximum speed v of expectation targetmaxAnd the minimum in scan period Ts computation partition regionDistance DivDis=10*vmax* Ts takes DivDis if calculated DivDis is less than preset minimum range span M=M.It is DivDis+ δ that search coverage in radar horizon R, which is divided into X span,dOverlapped zonule, δdIt isThe distance of overlapping region,
(2) it determines the parameter in each region: determining reference cycle number T firstagingWith peak extraction thresholding Thr, if havingPrior information knows the power of the clutter in investigative range, is arranged by prior information；If no, according to the region T within 40kmagingRegion T within=8, Thr=7,40-80kmagingRegion T other than=6, Thr=5,80kmaging=4, Thr=3 are in this wayMode value.Then according to taking fixed TagingAnd Thr, in conjunction with radar resolution, offline by the way of Monte Carlo simulation,Calculate the quantized interval Δ θ and Δ ρ in each region.Δ θ and Δ ρ, which takes, is capable of forming 95% hope target effectivelyThe minimum quantization interval of accumulation.Here so-called formation effectively accumulation refers to TagingThe Targets Dots in a period can be in parameter spaceThere is Thr or more ballot in one or more quantifying units in cumulant matrix.
(3) parameter setting and initialization: according to the parameter in each region of parameter setting that step 2 determines.According to Δ θ by parameterSpatial parameter θ ∈ (0 °, 180 °) is divided intoSection (for simplicity, 180) the quantized interval Δ θ of setting should be able to divide exactly,According to Δ ρ by parameter space parameter ρ ∈ (rmin,rmax) be divided intoSection, establishes parameter space cumulant matrixThe content that A and each of which elements A (i, j) are included is as shown in Figure 1, total throwing comprising the quantifying unitPoll TotalVoteNum and a length are equal to reference cycle number TagingVote information fifo queueEach element Q of vote information queuekComprising corresponding history votes HisVoteNum (k),Point mark array d (k)={ d (k, g) }G=1,2 ..., HisVoteNum (k).Total votes of quantifying unit are equal to quantifying unit ballot letterThe sum of queue all elements votes are ceased, i.e.,So far, initial workIt is fully completed, step below is alternately performed in each sector in sequence.
(4) Hough transformation ballot accumulation: all inputs in note current period region are DOT={ (x (doti), y(doti))}Doti=1,2 ..., DotNum(the point mark sum that DotNum is current period).For arbitrary point mark (x (doti), y(doti)), in conjunction with each θj=j* Δ θ, j=1,2 ..., N can be calculatedAccording to the i that each is calculated with its corresponding j, A (i, j) and A (i-1, j), A (i+1, j) these three quantifying units are enabledTail of the queue element Q in total votes TotalVoteNum and vote information queueTagingBallot value HisVoteNum(Taging) add one, and the mark is recorded in QTagingPoint mark information array in.
(5) parameter space cumulant matrix is traversed: whenever sector triggering carrys out interim, traversal parameter space cumulant matrix.ForEach quantifying unit regards the unit if its ballot value is greater than Thr as local peaking.The quantization list come out for peak extractionMember generates candidate track.The falseness in candidate track is rejected plus the method for observation error matrix using straight line fitting, will be rejectedResult afterwards is regarded as the quasi- track of current period, and there are in quasi- track group.For the step for, can be real using mode shown in Fig. 2Existing N number of submatrix parallel processing, N are determined by computing resource.By taking powerpc board as an example, using semaphore combination message queueMode realizes the synchronously control of parallel task.The task of each parallel processing waits the same semaphore, by semflush letterIt counts while activating all N number of parallel tasks, realize the synchronization of task；It is past to disappear after each task completes the processing of itselfIt ceases and sends a completion message in queue, continue waiting for semaphore；After all parallel tasks after the completion of message queue will receive NItem completes message, carries out subsequent serial arithmetic again later.
(6) track duplicate removal and sector update: after the completion of parameter space cumulant matrix all traversal, repeating in quasi- track groupTrack leave out, in quasi- track group remaining all tracks as initial results export.All initial results outputs of current periodLater, quasi- track group is emptied.For the either element A (i, j) in parameter space accumulation matrix A, by its vote information queueHead of the queue element Q1Team out updates total votes TotalVoteNum=TotalVoteNum-HisVoteNum (1).Sequentially by Q2…QTagingIt is assigned to Q1…QTaging-1, by tail of the queue element QTagingBallot value zero setting, point mark array empty.