Summary of the invention
The present invention provides a kind of recognition methods of illegal operation vehicle and system and computer-readable storage medium, withSolve existing the technical issues of can not accurately identifying to illegal operation vehicle.
According to an aspect of the present invention, a kind of recognition methods of illegal operation vehicle is provided comprising following steps:
S1: acquisition vehicle data is to establish number plate of vehicle database, vehicle running track database and bayonet longitude and latitude degreeAccording to library；
S2: the vehicle data of acquisition is cleaned；
S3: deep learning is carried out based on the cleaned vehicle data and establishes intelligent recognition model；And
S4: all vehicles are identified using intelligent recognition model.
Further, step S2 is further included steps of
S21: time entanglement in vehicle running track database and duplicate data are cleaned；
S22: the illegal vehicle in use type that is identified of selection, according to being identified illegal vehicle in use type-collectionThe vehicle running track data of respective type vehicle, and legal commerial vehicle number plate set and non-is obtained according to vehicles operation classificationCommerial vehicle number plate set；And
S23: classification exploratory analysis is carried out to dispose serious distortion data to the vehicle running track data of extraction.
Further, the mode that deep learning is carried out in step S3 is one-dimensional time locus deep learning, two-dimensional time railMark deep learning, the study of two-dimension time-space track depth, two-dimentional thermodynamic chart deep learning, three-dimensional space-time geometric locus figure deep learningWith one of three-dimensional thermodynamic chart deep learning or a variety of.
Further, the mode that deep learning is carried out in step S3 is one-dimensional time locus deep learning, and step S3 is specificThe following steps are included:
S31a: selection sample, including positive sample and negative sample；
S32a: vehicle running track data are sorted by number plate of vehicle and chronological order, are obtained every by building bayonet pairSequence is numbered by the bayonet track sets of a number plate from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1,By the time △ t of adjacent bayoneti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
；
S33a: setting bayonet is to transit time granularity, the overlength △ t for rejecting night's rest, closing a business, using normal state pointCloth asks △ t to fall in the section t in ± 3 σ of u, and section t is also all value interval of the bayonet to transit time granularity, setting cardMouth is λ to transit time granularity, and λ is a time slice, λ ∈ t；
S34a: one-dimensional time running track matrix modeling is carried out；
S35a: normalized；And
S36a: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S34a specifically:
The time running track of each number plate vehicle is processed into one-dimensional matrix Rm needed for convolutional neural networks, is taken one-dimensionalLength the length=T*24*60/ λ, T of time running track indicate that the date range number of days of the sample of selection, λ indicate bayonet pairTransit time granularity, j* λ indicate the total time of j time slice, if certain vehicle appears in some bayonet in j-th of segment,Then:
Rm [j]=1, otherwise, Rm [j]=0,
Firstly, being initialized to Rm, Rm [j]=0, j ∈ (0, length)
Then, its t0 at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicleTo time △ tft0=tf-t0 used in f-th of bayonet, then:
Rm [△ tft0/ λ]=1
According to this, the time running track matrix Rm of each number plate is obtained；Positive sample matrix Pm and negative sample are also obtained simultaneouslyThis matrix N m.
Further, the mode of deep learning is carried out in step S3 as the study of two-dimensional time track depth, step S3 is specificThe following steps are included:
S31b: selection sample, including positive sample and negative sample；
S32b: vehicle running track data are sorted by number plate of vehicle and chronological order, are obtained every by building bayonet pairSequence is numbered by the bayonet track sets of a number plate from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1,By the time △ t of adjacent bayoneti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
；
S33b: setting bayonet obtains time granularity, rejecting night's rest, the overlength △ t to close a business using normal distribution△ t is taken to fall in the section t in ± 3 σ of u, section t is also all value interval of the bayonet to transit time granularity, sets bayonet pairTransit time granularity is λ, and λ is a time slice, then λ ∈ t；
S34b: the modeling of two-dimensional time running track matrix is carried out；
S35b: normalized；And
S36b: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S34b is specially
Two-dimensional matrix R needed for the time running track of each number plate vehicle is processed into convolutional neural networksm, listShow number of days, col width Column_width=T, T are date range, and row indicates transit time, row_height=24*60/ λ, λ tableShow bayonet to transit time granularity, then to RmIt is initialized:
Rm[m] [n-1]=0, m ∈ (0,24*60/ λ), n ∈ (1, T), max (m) * λ=24*60
It on the date for passing through f-th of bayonet according to each number plate vehicle, calculates it and comes n-th day in (1, T), selected from sampleT at the beginning of selecting0To time △ t used in f-th of bayonetft0=tf-t0, then m=△ tft0/ λ-(n-1) * 24*60/ λ, that:
Rm[△tft0/ λ-(n-1) * 24*60/ λ] [n-1]=1
According to this, the time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm。
Further, the mode of deep learning is carried out in step S3 as the study of two-dimension time-space track depth, step S3 is specificThe following steps are included:
S31c: selection sample, including positive sample and negative sample；
S32c: vehicle running track data are sorted by number plate of vehicle and chronological order, are obtained every by building bayonet pairSequence is numbered by the bayonet track sets of a number plate from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1,By the time △ t of adjacent bayoneti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
；
S33c: setting bayonet obtains time granularity, rejecting night's rest, the overlength △ t to close a business using normal distribution△ t is taken to fall in the section t in ± 3 σ of u, section t is also all value interval of the bayonet to transit time granularity, sets bayonet pairTransit time granularity is λ, and λ is a time slice, then λ ∈ t；
S34c: two-dimentional bayonet space matrix modeling is carried out；
S35c: the modeling of two-dimension time-space running track matrix is carried out；
S36c: normalized；And
S37c: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S34c is specially
According to bayonet longitude and latitude data, square lattice of the region division that will operate at side length L dimension size, according to longitude and latitudeThe number of squares in direction is spent, bayonet number is numbered each grid, obtains bayonet space matrix SPACEm, SPACEm[u][v]=grid number, the affiliated grid of bayonet is spatial position representated by bayonet.
Further, step S35c is specially
Two-dimensional matrix R needed for the time running track of each number plate vehicle is processed into convolutional neural networksm, listShow number of days, col width Column_width=T, T are date range, and row indicates transit time, row_height=24*60/ λ, λ tableShow bayonet to transit time granularity, then to RmIt is initialized:
Rm[m] [n-1]=0, m ∈ (0,24*60/ λ), n ∈ (1, T), max (m) * λ=24*60
It on the date for passing through f-th of bayonet according to each number plate vehicle, calculates it and comes n-th day in (1, T), selected from sampleT at the beginning of selecting0To time △ t used in f-th of bayonetft0=tf-t0, then m=△ tft0/ λ-(n-1) * 24*60/ λ, passes throughThe number of f-th of bayonet searches bayonet space matrix SPACEm, obtain corresponding space number SPACEm[u] [v], then:
Rm[△tft0/ λ-(n-1) * 24*60/ λ] [n-1]=SPACEm[u][v]
According to this, the space-time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm。
Further, the mode that deep learning is carried out in step S3 is two-dimentional thermodynamic chart deep learning, and step S3 is specifically wrappedInclude following steps:
S31d: selection sample, including positive sample and negative sample；
S32d: forming prospect thermodynamic chart based on vehicle running track data, and vehicle running track data are pressed number plate of vehicleIt sorts with chronological order, vehicle is labeled on map in the form of thermodynamic chart the number of bayonet, to the last oneThen a bayonet goes out background base map, form prospect thermodynamic chart；
S33d: two-dimentional bayonet space matrix modeling is carried out；
S34d: normalized；And
S35d: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S33d is specially
According to bayonet longitude and latitude data, square lattice of the region division that will operate at side length L dimension size, according to longitude and latitudeThe number of squares in direction is spent, bayonet number is numbered each grid, obtains bayonet space matrix SPACEm, bayonet institute possessive caseSon is spatial position representated by bayonet, finally, the heating power value filling of square corresponding in thermodynamic chart is obtained bayonet spatial momentBattle array SPACEm[u] [v] obtains the thermodynamic chart running track matrix R-SPACE of each number plate according to thism；Positive sample is also obtained simultaneouslyThis matrix P-SPACEmWith negative sample matrix N-SPACEm。
Further, the mode that deep learning is carried out in step S3 is three-dimensional space-time geometric locus figure deep learning, stepS3 specifically includes the following steps:
S31e: selection sample, including positive sample and negative sample；
S32e: vehicle running track data are sorted by number plate of vehicle and chronological order, are obtained every by building bayonet pairSequence is numbered by the bayonet track sets of a number plate from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1,By the time △ t of adjacent bayoneti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
；
S33e: setting bayonet obtains time granularity, rejecting night's rest, the overlength △ t to close a business using normal distribution△ t is taken to fall in the section t in ± 3 σ of u, section t is also all value interval of the bayonet to transit time granularity, sets bayonet pairTransit time granularity is λ, and λ is a time slice, then λ ∈ t；
S34e: the modeling of three-dimensional time running track matrix is carried out；
S35e: normalized；And
S36e: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S34e is specially
According to bayonet longitude and latitude data, square lattice of the region division that will operate at side length L dimension size, according to longitude and latitudeThe number of squares in direction is spent, bayonet number is numbered each grid, obtains bayonet space matrix SPACEm, SPACEm[u][v]=grid number, the affiliated grid of bayonet are spatial position representated by bayonet, and port number is one-dimensional time running track sequenceMaximum value, channel=T*24*60/ λ, λ indicate bayonet to transit time granularity, and T indicates the number of days of the sample of selection；It is rightThree-dimensional matrice is initialized:
Channel-SPACEm [u] [v] [c]=0, c ∈ (0, channel)
Its t at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicle0To fTime △ t used in a bayonetft0=tf-t0, then c=△ tft0/ λ directly determines SPACE according to card number numberm[u] [v], thenHave:
Channel-SPACEm[u][v][△tft0/ λ]=SPACEm[u][v]
If will be by adjacent bayonet with being directly connected to, one spiralling three-dimensional space-time geometric locus of formation；
According to this, the three-dimensional space-time geometric locus figure R of each number plate is obtainedm；Positive sample P is also obtained simultaneouslymAnd negative sampleNm。
Further, the mode that deep learning is carried out in step S3 is three-dimensional thermodynamic chart deep learning, and step S3 is specifically wrappedInclude following steps:
S31f: selection sample, including positive sample and negative sample；
S32f: vehicle running track data are sorted by number plate of vehicle and chronological order, are obtained every by building bayonet pairSequence is numbered by the bayonet track sets of a number plate from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1,By the time △ t of adjacent bayoneti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
；
S33f: setting bayonet obtains time granularity, rejecting night's rest, the overlength △ t to close a business using normal distribution△ t is taken to fall in the section t in ± 3 σ of u, section t is also all value interval of the bayonet to transit time granularity, sets bayonet pairTransit time granularity is λ, and λ is a time slice, then λ ∈ t；
S34f: three-dimensional thermodynamic chart modeling is carried out；
S35f: normalized；And
S36f: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
Further, step S34f is specially
According to bayonet longitude and latitude data, square lattice of the region division that will operate at side length L dimension size, according to longitude and latitudeThe number of squares in direction is spent, bayonet number is numbered each grid, obtains bayonet space matrix SPACEm, bayonet institute possessive caseSon is spatial position representated by bayonet, and port number is one-dimensional time running track sequence maximum value, channel=T*24*60/ λ, λ indicate bayonet to transit time granularity, and T indicates the number of days of the sample of selection；Three-dimensional matrice is initialized:
Channel-SPACEm [u] [v] [c]=0, c ∈ (0, channel)
Vehicle running track data are sorted by number plate of vehicle and chronological order, by vehicle by the number of bayonet withThe form of thermodynamic chart will obtain the heating power of a chpn gradual change using λ as unit stratification drawing or the cumulative drafting of layering in this wayScheme, or be layered the thermodynamic chart of variation as unit of λ, to the last a bayonet, then go out the figure that breaks off the base, forms prospect thermodynamic chart；
Its t at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicle0TofIt is aTime △ t used in bayonetft0=tf-t0, then have c=△ tft0/λ；
The heating power value for being layered corresponding square in thermodynamic chart is inserted into three-dimensional thermodynamic chart Matrix C hannel-SPACEm [u] [v][c]；
According to this, the three-dimensional thermodynamic chart R-SPACE of each number plate is obtainedm；Positive sample P-SPACE is also obtained simultaneouslymWith negative sampleThis N-SPACEm。
Further, the vehicle running track data of specific time period can be extracted in step S22, and the period is selected with specific aim identificationThe vehicle of illegal operation.
It further, further include obtaining the video camera data of law enforcement point to establish law enforcement video database in step S1；
The recognition methods of illegal operation vehicle is further comprising the steps of:
S5: obtaining illegal operation number plate of vehicle based on recognition result, and from law enforcement video database it is corresponding transfer it is illegalThe video evidence or picture evidence of operation.
The present invention also provides a kind of identifying system of illegal operation vehicle, which includes
Data acquisition module, for acquiring vehicle data to establish number plate of vehicle database, vehicle running track databaseWith bayonet longitude and latitude database；
Data cleansing module, for being cleaned to the vehicle data of acquisition；
Deep learning module, for carrying out deep learning based on the cleaned vehicle running track database to establish intelligenceIdentification model simultaneously identifies all vehicles using the intelligent recognition model.
The present invention also provides a kind of computer-readable storage mediums, are used to store progress illegal operation vehicle identificationComputer program, the computer program execute following steps when running on computers:
S1: acquisition vehicle data is to establish number plate of vehicle database, vehicle running track database and bayonet longitude and latitude degreeAccording to library；
S2: the vehicle data of acquisition is cleaned；
S3: deep learning is carried out based on the cleaned vehicle data and establishes intelligent recognition model；And
S4: all vehicles are identified using intelligent recognition model.
The invention has the following advantages:
The recognition methods of illegal operation vehicle of the invention establishes number plate of vehicle data by acquisition vehicle dataLibrary, vehicle running track database and bayonet longitude and latitude database.And collected vehicle data is cleaned, cleaning processIn can be according to extract the vehicle running track data of corresponding type of vehicle by selection illegal operation type of vehicle, thus realNow accurately and accurately the illegal operation vehicle of different vehicle type is quickly identified；It can also be according to illegal operation vehicleThe periodically strong feature of operating slot correspondingly only extracts specific time period when extracting vehicle running track data, is selected with identificationThe vehicle of period illegal operation.And classification based training is carried out using deep learning convolutional neural networks, to realize efficiently, preciselyIllegal operation vehicle identification.
In addition, it is different according to the dimension of input data, there are one-dimensional time locus deep learning, two-dimensional time track depthIt practises, the study of two-dimension time-space track depth, two-dimentional thermodynamic chart deep learning, three-dimensional space-time geometric locus figure deep learning, Three Dimensional ThermalTry hard to a variety of deep learning modes of deep learning, can be enabled simultaneously from different dimensions, each recognition result is complementary to one another, and is mutually helpedCard, accuracy of identification are high.
Other than objects, features and advantages described above, there are also other objects, features and advantages by the present invention.Below with reference to figure, the present invention is described in further detail.
Specific embodiment
The embodiment of the present invention is described in detail below in conjunction with attached drawing, but the present invention can be limited by following andThe multitude of different ways of covering is implemented.
As shown in Figure 1, the preferred embodiment of the present invention provides a kind of recognition methods of illegal operation vehicle, may be implementedIllegal operation vehicle is carried out efficiently, accurately identify, the recognition methods of the illegal operation vehicle the following steps are included:
Step S1: acquisition vehicle data is to establish number plate of vehicle database, vehicle running track database and bayonet longitude and latitudeSpend database；
Step S2: the vehicle data of acquisition is cleaned；
Step S3: deep learning is carried out based on the cleaned vehicle data and establishes intelligent recognition model；
Step S4: all vehicles are identified using intelligent recognition model；And
Step S5: illegal operation number plate of vehicle is obtained based on recognition result, and correspondence is transferred from law enforcement video databaseThe video evidence or picture evidence of illegal operation.
It is appreciated that the step S5 can be omitted.
It is appreciated that in the step S1, number plate of vehicle data can be obtained from special bus pipe data system, it is describedNumber plate of vehicle data include at least number plate of vehicle, type of vehicle, operation type etc., and based on the number plate of vehicle data of acquisitionEstablish number plate of vehicle database；From traffic block port data system, vehicle bayonet data are obtained, the vehicle bayonet data are at leastVehicle is established including number plate of vehicle, bayonet number, by time and bayonet photo etc., and based on the vehicle bayonet data of acquisitionRunning track database；From traffic block port data system, the geographic position data of bayonet, the bayonet geographical location are obtainedData include at least bayonet number, bayonet belonging positions region and bayonet longitude and latitude etc., and based on the bayonet data of acquisitionEstablish bayonet longitude and latitude database.Preferably, in the step S1 also law enforcement can be acquired by the video camera of law enforcement pointThe video camera data of point are to establish law enforcement video database.Preferably, can also be by non-bayonet in the step S1Data are acquired to supplement the data volume in vehicle running track database.
It is appreciated that as shown in Fig. 2, the step S2 specifically includes the following steps:
S21: time entanglement, duplicate data in vehicle running track database are cleaned；
S22: the illegal vehicle in use type that is identified of selection, according to being identified illegal vehicle in use type-collectionThe vehicle running track data of respective type vehicle, and legal commerial vehicle number plate set and non-is obtained according to vehicles operation classificationCommerial vehicle number plate set；And
S23: classification exploratory analysis is carried out to dispose serious distortion data to the vehicle running track data of extraction.
Time entanglement, duplicate data can be cleaned in the step S21, it is ensured that collected vehicle bayonet numberAccording to accuracy.
In the step S22, illegal vehicle in use type includes car, minibus, commercial vehicle or car etc., can be withBy the illegal vehicle in use type for selecting to be identified, such as car is only selected, so as to extract all carsVehicle running track data, and by the running track image watermarking of other type of vehicle such as minibus, commercial vehicle and car riseCome, and legal commerial vehicle number plate set A and non-commerial vehicle number plate set B is obtained according to vehicles operation classification.It is appreciated thatVehicles operation classification includes non-commerial vehicle and commerial vehicle.The present invention can be extracted by selection illegal operation type of vehicleThe vehicle running track data of corresponding type of vehicle, may be implemented the illegal operation vehicle accurately and accurately to different vehicle typeIt is quickly identified.It is further appreciated that as it is further preferred that periodically strong according to the illegal operation vehicles operation periodFeature, some in festivals or holidays, correspondingly only took out when extracting vehicle running track data at morning peak, evening peak, night by someSpecific time period is taken, to identify the vehicle for selecting period illegal operation.
In the step S23, classification exploratory analysis is carried out to the vehicle running track data extracted in step S22,Bayonet data required for classification exploratory analysis are carried out to include at least: statistics is average daily cross bayonet number, when cross bayonet number, averageBayonet number, working day and weekend averagely mistake bayonet number are crossed, and applies normal distribution model, data in ± 3 σ of u are intercepted, to washSerious distortion data can also directly select the vehicle running track data within the scope of certain percentage according to statistical characteristics, withWash serious distortion data.The present invention, can be with by carrying out classification exploratory analysis to the vehicle running track data of extractionThe data for washing serious distortion further ensure the accuracy of identification.
It is appreciated that in the step S3, the mode for carrying out deep learning has one-dimensional time locus deep learning, two dimensionTime locus deep learning, the study of two-dimension time-space track depth, two-dimentional thermodynamic chart deep learning, three-dimensional space-time geometric locus figure are deepDegree study and three-dimensional thermodynamic chart deep learning.The present invention can carry out deep learning based on multidimensional running track, run to vehicleTrack data carries out Conceptual Modeling, carries out classification based training using deep learning convolutional neural networks, to realize efficiently, accuratelyIdentify illegal operation vehicle.
As shown in figure 3, the step S3 includes following step when using one-dimensional time locus deep learning in step S3It is rapid:
S31a: selection sample, including positive sample and negative sample；
S32a: building bayonet pair；
S33a: setting bayonet is to transit time granularity；
S34a: one-dimensional time running track matrix modeling is carried out；
S35a: normalized；And
S36a: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that in the step S31a, from by being selected in step S2 the cleaned vehicle running track databaseThe vehicle running track data R in certain date range T days is selected, the positive sample P and negative sample of same number plate quantity x are selected in RThis N, i.e., legal commerial vehicle running track data and non-commerial vehicle running track data, while obtaining positive sample number plate setPxWith negative sample number plate set Nx。
It is appreciated that vehicle running track data R is pressed number plate of vehicle and chronological order in the step S32aSequence, obtain the bayonet track sets of each number plate, bayonet track sets be numbered from i=1, have it is adjacent by bayonetForm bayonet pair: kiki+1, by the time △ t of adjacent bayoneti+1=ti+1-ti, building vehicle bayonet to transit time data,Format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
It is appreciated that in the step S33a, t >=4 hour overlength △ t, △ rejecting night's rest, closing a business, applicationNormal distribution asks △ t to fall in the section t in ± 3 σ of u, and section t is also all value interval of the bayonet to transit time granularity, ifDetermining bayonet is λ to transit time granularity, and λ is a time slice, then λ ∈ t.
It is appreciated that one-dimensional time running track matrix is carried out to the sample of selection and is modeled in the step S34a, it willThe time running track of each number plate vehicle is processed into one-dimensional matrix R needed for convolutional neural networksm, take one-dimensional time operation railLength length=T*24*60/ λ, the j* λ of mark indicates the total time of j time slice, if certain vehicle occurs in j-th of segmentIn some bayonet, then:
Rm[j]=1, otherwise, Rm[j]=0,
Firstly, to RmIt is initialized, Rm[j]=0, j ∈ (0, length)
Then, its t0 at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicleTo time △ t used in f-th of bayonetft0=tf-t0, then:
Rm[△tft0/ λ]=1
According to this, the time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm。
It is appreciated that the time running track matrix R of each number plate in the step S35am, positive sample matrix PmWithNegative sample matrix NmIt is normalized, to simplify data processing amount, to ensure efficient identification.
It is appreciated that in the step S36a, by the corresponding one-dimensional time running track of positive sample P and responsible sample NThe positive sample matrix P of matrixmWith negative sample matrix NmConvolutional neural networks are inputted, carry out feature learning, and utilize softmax pointsClass device classifies to the feature that convolutional network learns, and obtains the Vehicular intelligent identification model PN based on one-dimensional deep learningAI,Then, it using verify data, constantly adjustment T, x, λ, convolutional neural networks parameter, is trained, obtains best intelligent recognition mouldType PNAI.The structure of the convolutional neural networks is 4 Conv2d convolutional layers, 4 pond max-pooling layers, 2 Dense completeArticulamentum and output layer.
For example, from vehicle running track data R in T=30 days is selected in the cleaned vehicle running track database, in RThe positive sample P and negative sample N of the middle same number plate quantity x=20000 platform of selection, i.e. taxi running track data and family-sized carRunning track data, while obtaining positive sample number plate set PxWith negative sample number plate set Nx。
Vehicle running track data are sorted by number plate of vehicle and chronological order again, obtain the bayonet rail of each number plateSequence is numbered by mark sequence from i=1, then by it is adjacent by bayonet form bayonet pair: kiki+1, by adjacent bayonetTime △ ti+1=ti+1-ti, vehicle bayonet is constructed to transit time data, and format is as follows:
| License plate number | Bayonet is to kiki+1 | Transit time △ ti+1 |
△ t is asked to fall in the area in ± 3 σ of u using normal distribution in t≤4 hour overlength △ rejecting night's rest again, closing a businessBetween t (3,71), section t is also all value interval of the bayonet to transit time granularity, set bayonet to transit time granularityFor λ, then λ ∈ t, initially takes λ=5 minute.
Then one-dimensional time running track matrix modeling is carried out, the time running track of each number plate vehicle is handled coiledOne-dimensional matrix R needed for product neural networkm, the length length=30*24*60/5=8640 of one-dimensional time running track is taken,(i-1) * λ indicates vehicle by the time used in i-th of bayonet, to RmIt is initialized:
Rm[i]=0, i ∈ (1,8640)
Its t at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicle0To fTime t used in a bayonetft0=tf-t0, then:
Rm[tft0/ 5+1]=1
According to this, the time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm；
Again to Rm、Pm、NmNormalized；
Then by the positive sample matrix P of the corresponding one-dimensional time running track matrix of positive sample P and responsible sample NmWith it is negativeSample matrix NmConvolutional neural networks are inputted, carry out feature learning, and arrive to convolutional network study using softmax classifierFeature classify, obtain the Vehicular intelligent identification model PN based on one-dimensional deep learningAI, then, using verify data, noDisconnected adjustment T, x, λ, convolutional neural networks parameter, are trained, obtain best intelligent recognition model PNAI。
It is appreciated that as shown in figure 4, when in step S3 use two-dimensional time track depth study when, the step S3 packetInclude following steps:
S31b: selection sample, including positive sample and negative sample；
S32b: building bayonet pair；
S33b: setting bayonet is to time granularity；
S34b: the modeling of two-dimensional time running track matrix is carried out；
S35b: normalized；And
S36b: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that the step S31b is identical as S31a, S32b is identical as S32a, and S33b is identical as S33a, stepS36b is identical as step S36a.
It is appreciated that the modeling of two-dimensional time running track matrix is carried out to the sample of selection in the step S34b, it willTwo-dimensional matrix R needed for the time running track of each number plate vehicle is processed into convolutional neural networksm, column expression number of days, columnWide Column_width=T, row indicate transit time, row_height=24*60/ λ, then to RmIt is initialized:
Rm[m] [n-1]=0, m ∈ (0,24*60/ λ), n ∈ (1, T), max (m) * λ=24*60；
It on the date for passing through f-th of bayonet according to each number plate vehicle, calculates it and comes n-th day in (1, T), selected from sampleT at the beginning of selecting0To time △ t used in f-th of bayonetft0=tf-t0, then m=△ tft0/ λ-(n-1) * 24*60/ λ, that:
Rm[△tft0/ λ-(n-1) * 24*60/ λ] [n-1]=1；
According to this, the time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm。
It is appreciated that the time running track matrix R of each number plate in the step S35bm, positive sample matrix PmWithNegative sample matrix NmIt is normalized, to simplify data processing amount, to ensure efficient identification.
It is appreciated that as shown in figure 5, when in step S3 use two-dimension time-space track depth study when, the step S3 packetInclude following steps:
S31c: selection sample, including positive sample and negative sample；
S32c: building bayonet pair；
S33c: setting bayonet is to time granularity；
S34c: continue two-dimentional bayonet space matrix modeling；
S35c: the modeling of two-dimension time-space running track matrix is carried out；
S36c: normalized；And
S37c: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that the step S31c is identical with S31a, S32c is identical as S32a, and S33c is identical as S33a, S36c andS35a is identical, and S37c is identical as S36a.
It is further appreciated that in the step S34c, according to bayonet longitude and latitude data, by operation region division at L dimensionThe square lattice of size, according to the number of squares in longitude and latitude direction, bayonet number is numbered each grid, obtains bayonetSpace matrix SPACEm, SPACEm[u] [v]=grid number, the affiliated grid of bayonet is spatial position representated by bayonet.
In the step S35c, the modeling of two-dimension time-space running track matrix is carried out to the sample of selection, by each number plateTwo-dimensional matrix R needed for the time running track of vehicle is processed into convolutional neural networksm, column expression number of days, col width Column_Width=T, row indicate transit time, row_height=24*60/ λ, then to RmIt is initialized:
Rm[m] [n-1]=0, m ∈ (0,24*60/ λ), n ∈ (1, T), max (m) * λ=24*60
It on the date for passing through f-th of bayonet according to each number plate vehicle, calculates it and comes n-th day in (1, T), selected from sampleT at the beginning of selecting0To time △ t used in f-th of bayonetft0=tf-t0, then m=△ tft0/ λ-(n-1) * 24*60/ λ, passes throughThe number of f-th of bayonet searches bayonet space matrix SPACEm, obtain corresponding space number SPACEm[u] [v], then:
Rm[△tft0/ λ-(n-1) * 24*60/ λ] [n-1]=SPACEm[u][v]
According to this, the space-time running track matrix R of each number plate is obtainedm；Positive sample matrix P is also obtained simultaneouslymAnd negative sampleMatrix Nm。
It is appreciated that as shown in fig. 6, the step S3 includes when in step S3 using two-dimentional thermodynamic chart deep learningFollowing steps:
S31d: selection sample, including positive sample and negative sample；
S32d: prospect thermodynamic chart is formed based on vehicle running track data；
S33d: two-dimentional bayonet space matrix modeling is carried out；
S34d: normalized；And
S35d: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that the step S31d is identical as S31a, S34d is identical as S35a, and S35d is identical as S36a.
In the step S32d, vehicle running track data are sorted by number plate of vehicle and chronological order, by vehicleIt is labeled on map in the form of thermodynamic chart the number of bayonet, to the last a bayonet, then goes out background base map,Formation prospect thermodynamic chart.
In the step S33d, two-dimentional bayonet space matrix modeling is carried out to the sample of selection, according to bayonet longitude and latitudeData, square lattice of the region division that will operate at side length L dimension size, according to the number of squares in longitude and latitude direction, bayonet is compiledNumber, each grid is numbered, bayonet space matrix SPACE is obtainedm, the affiliated grid of bayonet is space representated by bayonetPosition, finally, the heating power value filling of square corresponding in thermodynamic chart is obtained bayonet space matrix SPACEm[u][v].According to this, it obtainsObtain the thermodynamic chart running track matrix R-SPACE of each number platem；Positive sample matrix P-SPACE is also obtained simultaneouslymAnd negative sampleMatrix N-SPACEm。
As shown in fig. 7, when the mode for carrying out deep learning in step S3 is three-dimensional space-time geometric locus figure deep learning, instituteState step S3 specifically includes the following steps:
S31e: selection sample, including positive sample and negative sample；
S32e: building bayonet pair；
S33e: setting bayonet is to time granularity；
S34e: three-dimensional time running track matrix modeling；
S35e: normalized；And
S36e: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that the step S31e is identical with S31a, S32e is identical as S32a, and S33e is identical as S33a, S35e andS35a is identical, and S36e is identical as S36a.
It is appreciated that the modeling of three-dimensional space-time geometric locus figure is carried out to the sample of selection in the step S34e, according toBayonet longitude and latitude data, square lattice of the region division that will operate at side length L dimension size, the grid according to longitude and latitude directionNumber, bayonet number, is numbered each grid, obtains bayonet space matrix SPACEm, SPACEm[u] [v]=grid number,The affiliated grid of bayonet is spatial position representated by bayonet.Port number is one-dimensional time running track sequence maximum value,Channel=T*24*60/ λ；Three-dimensional matrice is initialized:
Channel-SPACEm[u] [v] [c]=0, c ∈ (0, channel)
Its t at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicle0To fTime △ t used in a bayonetft0=tf-t0, then c=△ tft0/ λ directly determines SPACE according to card number numberm[u] [v], thenHave:
Channel-SPACEm[u][v][△tft0/ λ]=SPACEm[u][v]
If will be by adjacent bayonet with being directly connected to, one spiralling three-dimensional space-time geometric locus of formation；
According to this, the three-dimensional space-time geometric locus figure R of each number plate is obtainedm；Positive sample P is also obtained simultaneouslymAnd negative sampleNm。
As shown in figure 8, when the mode for carrying out deep learning in the step S3 is three-dimensional thermodynamic chart deep learning, it is describedStep S3 specifically includes the following steps:
S31f: selection sample, including positive sample and negative sample；
S32f: building bayonet pair；
S33f: setting bayonet is to time granularity；
S34f: three-dimensional thermodynamic chart modeling is carried out；
S35f: normalized；And
S36f: deep learning is carried out to establish intelligent recognition model using convolutional neural networks.
It is appreciated that the step S31f is identical with S31a, S32f is identical as S32a, and S33f is identical as S33a, S35f andS35a is identical, and S36f is identical as S36a.
It is appreciated that three-dimensional thermodynamic chart modeling is carried out to the sample of selection, according to bayonet longitude and latitude in the step S34fDegree evidence, square lattice of the region division that will operate at side length L dimension size, according to the number of squares in longitude and latitude direction, bayonetNumber, is numbered each grid, obtains bayonet space matrix SPACEm, the affiliated grid of bayonet is sky representated by bayonetBetween position.Port number is one-dimensional time running track sequence maximum value, channel=T*24*60/ λ；Three-dimensional matrice is carried out justBeginningization:
Channel-SPACEm[u] [v] [c]=0, c ∈ (0, channel)
Vehicle running track data are sorted by number plate of vehicle and chronological order, by vehicle by the number of bayonet withThe form of thermodynamic chart will obtain the heating power of a chpn gradual change using λ as unit stratification drawing or the cumulative drafting of layering in this wayScheme, or be layered the thermodynamic chart of variation as unit of λ, to the last a bayonet, then go out the figure that breaks off the base, forms prospect thermodynamic chart.
Its t at the beginning of the samples selection is calculated by the time of f-th of bayonet according to each number plate vehicle0To fTime △ t used in a bayonetft0=tf-t0, then have c=△ tft0/λ；
The heating power value for being layered corresponding square in thermodynamic chart is inserted into three-dimensional thermodynamic chart Matrix C hannel-SPACEm[u][v][c]。
According to this, the three-dimensional thermodynamic chart R-SPACE of each number plate is obtainedm；Positive sample P-SPACE is also obtained simultaneouslymWith negative sampleThis N-SPACEm。
It is appreciated that in the step S3, the characteristics of according to convolutional neural networks, since the information for obtaining sample is detailedDegree is different, therefore, three-dimensional space-time geometric locus figure deep learning, three-dimensional thermodynamic chart deep learning recognition accuracy be higher thanThe recognition accuracy of the study of two-dimensional time track depth, the study of two-dimension time-space track depth and two-dimentional thermodynamic chart deep learning, twoDimension time locus deep learning, two-dimension time-space track depth learn, the recognition accuracy of two-dimentional thermodynamic chart deep learning is higher thanThe recognition accuracy of one-dimensional time locus deep learning.When being identified, can be enabled simultaneously from different dimensions, various identificationsAs a result it is complementary to one another, mutually proves, check layer by layer, ensure that the accuracy rate of identification.
It is appreciated that identifying, being sought to all sample datas using intelligent recognition model in the step S4Vehicle set A ', non-commerial vehicle set B ' are transported, since there are illegal operation vehicle, i.e., non-commerial vehicle participates in commerial vehicleOperation, then: A '>A, B '<B, then A '-A is then illegal operation number plate of vehicle set.A ∪ B ≈ A ' ∪ B ', when discrimination isWhen 100%, there is A+B=A '+B '.A ' ∩ (A ' ∪ B '-A ∪ B), this is the number plate of vehicle of strange land vehicle local Hacking Run.ThisInvention can carry out intelligent recognition to illegal operation vehicle by intelligent recognition model, quickly and accurately check out illegal operationVehicle.
It is appreciated that obtaining the vehicle of illegal operation vehicle based on the recognition result in step S4 in the step S5Number plate, and the corresponding video evidence or picture evidence for transferring illegal operation from law enforcement video database, or directly pass throughThe video camera at law enforcement point scene obtains the video evidence or picture evidence of illegal operation.
The recognition methods of illegal operation vehicle of the invention is built respectively by carrying out various dimensions to vehicle running track dataMould is learnt and is trained using convolutional neural networks, and the intelligent recognition model under each dimensional model is obtained, and obtains corresponding modelUnder illegal operation vehicle set, be complementary to one another, mutually prove, it is ensured that the precision of identification, especially combine video data acquiring,It is strong to improve the legitimacy and efficiency for checking and detaining illegal operation vehicle, beautify vehicles operation environment, ensures people's life and property peaceEntirely.Meanwhile using convolutional neural networks, the error introduced in explicit feature extraction and threshold value setting is avoided, and implicitlyLearnt from training data；Convolutional neural networks have unique superiority, especially multidimensional defeated in terms of image procossingThe image of incoming vector can directly input the complexity that network this feature avoids data reconstruction in feature extraction and assorting processDegree.
As shown in figure 9, another embodiment of the present invention also provides a kind of identifying system of illegal operation vehicle, may be implementedIllegal operation vehicle efficiently, accurately identify, the identification side of illegal operation vehicle as described above is preferably applied toMethod.The identifying system of the illegal operation vehicle includes for acquiring vehicle data to establish number plate of vehicle database, vehicle fortuneThe data acquisition module 12 of row track database and bayonet longitude and latitude database, for what is cleaned to the vehicle data of acquisitionData cleansing module 13, controller 14 and for based on the cleaned vehicle running track database carry out deep learning to buildVertical intelligent recognition model and the deep learning module 15 that all vehicles are identified using the intelligent recognition model, data acquisitionModule 12, data cleansing module 13, deep learning module 15 are connect with controller 14, and data cleansing module 13 is gone back and dataAcquisition module 12 connects.
It is appreciated that preferably, the identifying system of the illegal operation vehicle further includes bayonet collector 11, it is describedBayonet collector 11 is connect with data acquisition module 12, and the data acquisition module 12 obtains vehicle by bayonet collector 11Bayonet data are to establish vehicle running track database.
Preferably, the identifying system of the illegal operation vehicle further includes the video camera number for acquiring law enforcement pointAccording to video acquisition device 17, video acquisition device 17 connect with data acquisition module 12, and the data acquisition module 12 passes throughVideo acquisition device 17 come obtain law enforcement point video camera data to establish law enforcement video database.It is appreciated that the viewFrequency acquisition device 17 is video camera.
As shown in Figure 10, the data acquisition module 12 includes for acquiring number plate of vehicle data to establish number plate of vehicle numberThe bayonet data of vehicle running track database are established according to the number plate acquisition unit 121 in library, for acquiring vehicle bayonet dataAcquisition unit 122 and geographic position data for acquiring bayonet are to establish the bayonet longitude and latitude acquisition of bayonet longitude and latitude databaseUnit 123；Number plate acquisition unit 121, bayonet data acquisition unit 122 and bayonet longitude and latitude acquisition unit 123 are and controller14 connections.The number plate acquisition unit 121 can obtain number plate of vehicle data, the vehicle from special bus pipe data systemNumber plate data include at least number plate of vehicle, type of vehicle, operation type etc., and are established based on the number plate of vehicle data of acquisitionNumber plate of vehicle database.The bayonet data acquisition unit 122 can obtain vehicle bayonet number from traffic block port data systemAccording to or the bayonet data acquisition unit 122 connect with bayonet collector 11, vehicle is directly obtained by bayonet collector 11Bayonet data, the vehicle bayonet data include at least number plate of vehicle, bayonet number, by time and bayonet photo etc., andVehicle running track database is established based on the vehicle bayonet data of acquisition.The bayonet longitude and latitude acquisition unit 123 canTo obtain the geographic position data of each bayonet from traffic block port data system, the bayonet geographic position data is included at leastBayonet number, bayonet belonging positions region and bayonet longitude and latitude etc., and bayonet is established based on the vehicle bayonet data of acquisitionLongitude and latitude database.Preferably, the data acquisition module 12 further includes the video camera data for acquiring law enforcement pointTo establish the video data acquiring unit 124 of law enforcement video database, video data acquiring unit 124 and controller 14 and videoAcquisition device 17 connects.Preferably, the data acquisition module 12 further include for non-bayonet data be acquired withTo the non-bayonet data acquisition unit 125 that the data volume in vehicle running track database is supplemented, non-bayonet data acquisitionUnit 125 is connect with controller 14.
As shown in figure 11, the data cleansing module 13 include for time entanglement in vehicle running track database,Data cleansing unit 131 that duplicate data are cleaned, for selecting the illegal vehicle type identified, according to quiltThe vehicle running track data of identification illegal vehicle type-collection respective type vehicle are simultaneously legal according to the acquisition of vehicles operation classificationThe data selection unit 132 of commerial vehicle number plate set and non-commerial vehicle number plate set and for running rail to the vehicle of extractionMark data carry out classification exploratory analysis to dispose the data truncation unit 133 of serious distortion data, data cleansing unit131, data selection unit 132 and data truncation unit 133 are connect with controller 14.It is appreciated that the data selection is singleMember 132 is also connect with data cleansing unit 131 and data truncation unit 133.
It is appreciated that the data cleansing module 13 cleans time entanglement, duplicate data, it is ensured that collectedThe accuracy of vehicle bayonet data.The data selection unit 132 selects the illegal vehicle in use type identified, exampleAs only selected car, so as to extract the vehicle running track data of all cars, and by minibus, commercial vehicle and carRunning track image watermarking etc. other type of vehicle gets up, and obtains legal commerial vehicle number plate collection according to vehicles operation classificationClose A and non-commerial vehicle number plate set B.Preferably, in order to which to legal commerial vehicle, self clone's deck is illegal to carry outThe behavior of operation, can be by data selection unit 132 by legal commerial vehicle track database random pair, same number plate vehicleTrack data is double, thus dedicated for identifying Clonal illegal operation vehicle.It is further appreciated that as further preferred, according to the feature that the illegal operation vehicles operation period is periodically strong, some is in festivals or holidays, and some is at morning peak, evening peak, nightBetween, specific time period is correspondingly only extracted when extracting vehicle running track data by data selection unit 132, when selecting with identificationThe vehicle of section illegal operation.The vehicle running track number that the data truncation unit 133 can extract data selection unit 132According to classification exploratory analysis is carried out, carry out bayonet data required for classification exploratory analysis and include at least: average daily cross of statistics is blockedMouthful number, when cross bayonet number, be averaged bayonet number, working day and weekend averagely cross bayonet number, and apply normal distribution model, sectionData in ± 3 σ of u are taken, to wash serious distortion data, certain percentage range can also be directly selected according to statistical characteristicsInterior vehicle running track data, to wash serious distortion data.The present invention passes through the vehicle running track data to extractionClassification exploratory analysis is carried out, can understand the data for falling serious distortion, further ensures the accuracy of identification.
As shown in figure 12, the deep learning module 15 includes for selecting from the cleaned vehicle running track databaseSelect the sample selecting unit 151 of sample, the data modeling unit 152 for carrying out modeling processing to sample, for utilizing convolutionNeural network carries out deep learning to the sample after modeling to establish the deep learning unit 153 of intelligent recognition model and for benefitThe recognition unit 154 of illegal operation vehicle, sample selecting unit 151, data modeling unit are identified with intelligent recognition model152, deep learning unit 153 and recognition unit 154 are connect with controller 14.It is appreciated that the data modeling unit 152It is connect respectively with sample selecting unit 151 and deep learning unit 153, the deep learning unit 153 and recognition unit 154 connectIt connects.It is appreciated that the mode for carrying out deep learning in the deep learning unit 153 has one-dimensional time locus deep learning, twoTie up time locus deep learning, the study of two-dimension time-space track depth, two-dimentional thermodynamic chart deep learning, three-dimensional space-time geometric locus figureDeep learning and three-dimensional thermodynamic chart deep learning.The structure for the convolutional neural networks for including in the deep learning unit 153 is 4A Conv2d convolutional layer, 4 pond max-pooling layers, the full articulamentum of 2 Dense and output layer.Preferably, the depthUnit 153 uses softmax classifier to classify to obtain intelligent recognition the feature that convolutional neural networks learnModel.
Preferably, the identifying system of the illegal operation vehicle further includes corresponding for being transferred based on recognition resultThe evidence collection module 16 of illegal operation video evidence or picture evidence, evidence collection module 16 and controller 14 and data acquireModule 12 connects.Feedback information connects to controller 14, controller 14 after deep learning module 15 identifies illegal operation vehicleIt receives control 16 executive evidence of evidence collection module after the feedback information and collects instruction, evidence collection module 16 is from data acquisition moduleThe number plate of illegal operation vehicle is extracted in the number plate of vehicle database of block 12, and is accordingly transferred from law enforcement video databaseThe illegal operation vehicle carries out the video evidence and/or picture evidence of illegal operation, and alarm is retransmited after having collected evidenceInformation is to controller 14.It is appreciated that being used as a kind of deformation, the evidence collection module 16 can also be with video acquisition device 17The video evidence and/or picture evidence of illegal operation are directly collected in connection by video acquisition device 17.
The controller 14 can be by sending acquisition data command to data acquisition module 12 to control data acquisition moduleBlock 12 carries out vehicle data collection, and data acquisition module 12 is sent completely information after acquiring and feeds back to controller 14, and waitsThe data acquisition instructions of device 14 to be controlled next time.Feedback is completed in the acquisition that controller 14 receives the feedback of data acquisition module 12Transmission data cleansing, which is instructed, after information is adopted to data cleansing module 13 with controlling data cleansing module 13 to data acquisition module 12The vehicle data collected is cleaned.Data cleansing module 13 is sent completely information after cleaning and feeds back to controller 14 and waitThe data cleansing instruction of device 14 to be controlled next time, controller 14 receive feedback information transmission deep learning and instruct to depthStudy module 15, deep learning module 15 execute deep learning instruction to identify to illegal operation vehicle, identify completionInformation is sent completely afterwards to feed back to controller 14 and wait the deep learning instruction of controller 14 next time.Controller 14 receivesIt sends evidence-gathering after the information that deep learning module 15 is fed back to instruct to evidence-gathering module 16, evidence-gathering module 16 executesEvidence-gathering instruct and extracted from the number plate of vehicle database of data acquisition module 12 illegal operation vehicle number plate and fromVideo evidence and/or picture evidence that the illegal operation vehicle carries out illegal operation are accordingly transferred in law enforcement video database,After evidence-gathering is completed, evidence-gathering module 16 sends a warning message to controller 14.
The identifying system of illegal operation vehicle of the invention is built respectively by carrying out various dimensions to vehicle running track dataMould is learnt and is trained using convolutional neural networks, and the intelligent recognition model under each dimensional model is obtained, and obtains corresponding modelUnder illegal operation vehicle set, be complementary to one another, mutually prove, it is ensured that the precision of identification, especially combine video data acquiring,It is strong to improve the legitimacy and efficiency for checking and detaining illegal operation vehicle, beautify vehicles operation environment, ensures people's life and property peaceEntirely.Meanwhile using convolutional neural networks, the error introduced in explicit feature extraction and threshold value setting is avoided, and implicitlyLearnt from training data；Convolutional neural networks have unique superiority, especially multidimensional defeated in terms of image procossingThe image of incoming vector can directly input the complexity that network this feature avoids data reconstruction in feature extraction and assorting processDegree.
Another embodiment of the present invention also provides a kind of computer-readable storage medium, is used to store and illegally be soughtThe computer program of vehicle identification is transported, which executes following steps when running on computers:
S1: acquisition vehicle data is to establish number plate of vehicle database, vehicle running track database and bayonet longitude and latitude degreeAccording to library；
S2: the vehicle data of acquisition is cleaned；
S3: deep learning is carried out based on the cleaned vehicle data and establishes intelligent recognition model；And
S4: all vehicles are identified using intelligent recognition model.
The form of general computer-readable medium includes: floppy disk (floppy disk), flexible disc (flexibleDisk), hard disk, tape, it is any its with magnetic medium, CD-ROM, remaining any optical medium, punched card (punchCards), paper tape (paper tape), remaining any physical medium of pattern with hole, random access memory (RAM),Programmable read only memory (PROM), erasable programmable read-only memory (EPROM), the read-only storage of quick flashing erasable programmableDevice (FLASH-EPROM), remaining any memory chip or cassette or it is any remaining can allow computer read medium.InstructionIt can further be sent or receive by a transmission medium.This term of transmission medium may include any tangible or invisible medium,It, which can be used to store, encodes or carries, is used to the instruction that executes to machine, and include digital or analog communication signal or its withPromote the intangible medium of the communication of above-metioned instruction.Transmission medium includes coaxial cable, copper wire and optical fiber, and it comprises be used to passThe conducting wire of the bus of a defeated computer data signal.
The foregoing is only a preferred embodiment of the present invention, is not intended to restrict the invention, for the skill of this fieldFor art personnel, the invention may be variously modified and varied.All within the spirits and principles of the present invention, made any to repairChange, equivalent replacement, improvement etc., should all be included in the protection scope of the present invention.