Summary of the invention
The purpose of the present invention is to provide a kind of vehicle target Trajectory Tracking System, the system can realize move vehicle with
Track.
To achieve the goals above, the technical solution of use is the present invention:
A kind of vehicle target Trajectory Tracking System includes data collection system and data control system, data acquisition system
Uniting, there have to be multiple, and each data collection system includes interchanger, industrial personal computer, license plate camera and millimetre-wave radar, license plate camera
Industrial personal computer is connected by data line, millimetre-wave radar is connected by signal wire with interchanger, and interchanger passes through signal wire and industry control
Machine is connected;
Data control system includes data server and data display instrument, and industrial personal computer passes through public network data converter and data
Server is connected, and data display instrument is connected on data server.
Preferably, the data collection system further includes multiple positioning supports;License plate in each data collection system is taken the photograph
As head and millimetre-wave radar have multiple, interchanger and industrial personal computer in each data collection system are provided with one, each
A license plate camera and a millimetre-wave radar is arranged in the upper adaptation for positioning support.
Preferably, the industrial personal computer include data memory module, data analysis reject module, vehicle tracking computing module and
Data transmission blocks, data memory module connect data analysis by data line and reject module, vehicle tracking computing module sum number
According to sending module.
Another object of the present invention is to provide a kind of Vehicle tracing methods.
To achieve the goals above, the technical solution of use is the present invention:
A kind of Vehicle tracing method, specifically includes the following steps:
Step 1 controls data server and data collection system starting by data display instrument, after industrial personal computer starting,
The local and remote configuration file in data memory module is loaded, module is rejected in initialization data analysis, vehicle tracking calculates mould
Block and data transmission blocks;Data in local and remote configuration file enter analysis and reject module and vehicle tracking calculating mould
In block;
Step 2, millimetre-wave radar work, millimetre-wave radar emit electromagnetic wave and are irradiated to mobile and static items
And its echo is received, thus to obtain the distance of items to electromagnetic emission point, speed, position and RCS energy value etc. about items
The items information of position;
The items information monitored is transferred in interchanger by step 3, millimetre-wave radar;Interchanger is by the object after exchange
Item information is transferred to data analysis and rejects in module, and the items information of interchanger transmission at this time is initial data, and initial data is every
Secondary one frame of reception;
Step 4, data analysis reject module and carry out preliminary judgement to items information;
If items information is imperfect, items information is returned in millimetre-wave radar, and repeats step 2 after continued access
Receive the initial data of millimetre-wave radar;
If items information is complete, subsequent step is carried out;
Step 5, data analysis reject module and parse original items information, and data analysis is rejected module and passed through
Noise eliminating method rejects non-targeted items information, and obtains clean items information data;
Step 6, data analysis reject module and clean items information data are input in vehicle tracking computing module,
Vehicle tracking computing module obtains the track following data of corresponding items according to RNN vehicle target track following algorithm;
Track following data are transferred to data transmission blocks, data transmission blocks by step 7, vehicle tracking computing module
Track following result is transmitted in data server and in public network;
Step 8, after the data of a frame vehicle are parsed, return step five, industrial personal computer is using udp protocol after continued access
Receive the initial data of millimetre-wave radar monitoring.
Preferably, the noise eliminating method includes the following steps:
The first step rejects all ambient noises unrelated with items information itself and interferes data;
Second step is analyzed due to data and rejects in module and threshold value can be arranged according to RCS energy value size property, according to RCS
The threshold value of energy value rejects the noise targets unrelated with RCS energy value;
Third step, analyzes to reject in module threshold value is arranged according to distance, speed and position characteristic due to data, according to away from
From, the threshold value of speed and position, reject with apart from the unrelated noise targets of, speed and position.
Preferably, the RNN vehicle target track following algorithm includes the following steps:
The first step is established the collected initial data of millimetre-wave radar on the basis of polar coordinate system, both the two of object
Dimension space position indicates with radial distance and angle, for simplicity after mathematical computations, we by the information of polar coordinate system,
Two-dimensional space rectangular coordinate system is converted to, is indicated with (x, y) value;
Second step, by the conversion of coordinate system, when previous target data format is that (a (time of origin), x, y, s are (radial
Speed), r (RCS value)), with reference to the limiting value of millimetre-wave radar, i.e. maximum probe radial distance, maximum probe angle, maximum spy
Radial velocity, maximum RCS value are surveyed, data are become into the numerical value in (0,1) section, so as to subsequent processing;
Third step after data processing, passes through complete frame after data cleansing for incoming one group every time, thunder used in us
It is 32 target informations up to a frame, it may be assumed that
xi=(ai xi yi si ri)
After being often passed to one group of complete frame, data will carry out returning Recognition with Recurrent Neural Network program, return Recognition with Recurrent Neural Network
After the completion of program, corresponding target vehicle track following data are obtained;
4th step, RNN vehicle target track following algorithm are divided into multiple each RNN units, for each RNN unit binding one
A evaluator;
If RNN is continued working, evaluator meeting sustained activation, it can't be more than artificially to be arranged most that scoring, which can continue to increase,
Big value;
If RNN does not work for a long time, the continuous decrement that scores can then initialize this RNN when being reduced to minimum value in operation
Unit makes it that new target be waited to occur.
The beneficial effects of the present invention are:
The invention of this work utilizes the accurate target tracking characteristic of millimetre-wave radar, merges the identity of camera Car license recognition vehicle
Information realizes the accurate perception of vehicle driving order.Used technology is rarely reported in industry at present, is had preferable theoretical
Innovative and engineering practicability.The invention is by carrying out real-time dynamic monitoring to vehicle running track, to formulate traffic control plan
Reliable data supporting is slightly provided, the management of partial informationization existing for traffic management department blind area is effectively filled up.Meanwhile the invention
To the differentiation of the traffic flow parameters such as the magnitude of traffic flow, speed, queue length and the traffic abnormal incidents such as exception parking and traffic accident
Automatic detection provides condition, has a extensive future, and cost performance is higher, has biggish market popularization value.
The present invention is directed to identify road vehicle mesh based on RNN method by analyzing millimetre-wave radar detection data
Driving trace is marked, the accuracy that vehicle running state differentiates is improved, and propose a kind of vehicle target Trajectory Tracking System, realizes vehicle
The sequencing of track, is uncivil driving behavior management, and the traffic controls such as traffic flow modes monitoring provide real-time, accurate data branch
Support.
The present invention analyzes radar data using RNN in real time, using between artificial intelligence mode recognition detection target
Association, realize being continuously tracked for vehicle target, and propose a kind of traffic sequencing system based on millimetre-wave radar, carry out traffic
The acquisition of information reduces system cost.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawing:
Fig. 1 is combined first, and a kind of vehicle target Trajectory Tracking System includes data collection system and data control system
System.Data collection system has multiple, and each data collection system includes interchanger, industrial personal computer, license plate camera and millimeter wave thunder
It reaches, license plate camera connects industrial personal computer by data line, and millimetre-wave radar is connected by signal wire with interchanger, and interchanger passes through
Signal wire is connected with industrial personal computer.
Data control system includes data server and data display instrument, and industrial personal computer passes through public network data converter and data
Server is connected, and data display instrument is connected on data server.
Data collection system further includes multiple positioning supports;License plate camera and millimeter wave in each data collection system
Radar has multiple, and interchanger and industrial personal computer in each data collection system are provided with one, each positioning support it is upper
Adaptation one license plate camera of setting and a millimetre-wave radar.
Industrial personal computer includes that data memory module, data analysis rejecting module, vehicle tracking computing module and data send mould
Block, data memory module connect data analysis by data line and reject module, vehicle tracking computing module and data transmission blocks.
License plate camera in the present invention, is used cooperatively with millimetre-wave radar.When requiring to look up target vehicle, pass through vehicle
Board camera finds target vehicle, then in data display instrument, finds the motion track of target vehicle.
In conjunction with Fig. 2, a kind of Vehicle tracing method, specifically includes the following steps:
Step 1 controls data server and data collection system starting by data display instrument.After industrial personal computer starting,
Load the local and remote configuration file in data memory module.Module is rejected in initialization data analysis, vehicle tracking calculates mould
Block and data transmission blocks.Data in local and remote configuration file enter analysis and reject module and vehicle tracking calculating mould
In block.
Step 2, millimetre-wave radar work, millimetre-wave radar emit electromagnetic wave and are irradiated to mobile and static items
And its echo is received, thus to obtain the distance of items to electromagnetic emission point, speed, position and RCS energy value etc. about items
The items information of position.
The items information monitored is transferred in interchanger by step 3, millimetre-wave radar;Interchanger is assisted by TCP/IP
Items information after exchange is transferred to data analysis and rejected in module by view, and the items information of interchanger transmission at this time is original number
According to initial data only receives a frame every time.
Step 4, data analysis reject module and carry out preliminary judgement to items information;
If items information is imperfect, items information is returned in millimetre-wave radar, and repeats step 2 after continued access
Receive the initial data of millimetre-wave radar.Initial data is 264 16 binary datas, wherein the format of a frame is as shown in figure 3, industry control
Machine tests to initial data according to data byte length characteristic.If data packet the first two byte is 0x1973 and length meets
264 bytes then think that data are complete and carry out data parsing to it, and next frame data are continued to if imperfect.
If items information is complete, subsequent step is carried out.
Step 5, data analysis reject module and parse original items information;
In conjunction with Fig. 4, the data after parsing include head_indent and head_sequence, and head_indent is identity
Mark shows that the data are the data that radar is sent if the value is 0x1973.Head_sequence is frame number, and radar is every
A frame data are issued, which can add one.Above-mentioned data are parsed according to udp protocol, and are arranged the matrix of 32X4, every a line
Data format be [dn,vn,an,rn], which is subjected to noise targets rejecting.
Data analysis rejects module and is rejected non-targeted items information by noise eliminating method, and obtains clean object
Item information data;
Step 6, data analysis reject module and clean items information data are input in vehicle tracking computing module,
Vehicle tracking computing module obtains the track following data of corresponding items according to RNN vehicle target track following algorithm.
Radar data calculating is calculated by RNN vehicle target track following algorithm.RNN vehicle target track following is calculated
Method is a kind of algorithm for being continuously tracked using millimetre-wave radar and being calculated under multi-target condition.It is based on traditional circulation nerve
Network algorithm makes algorithm according to last moment state and this moment by addition linear regression unit and the short-term memory factor
Input calculates the unsupervised segmentation of this moment input, and updates the relevant parameter of linear regression unit, to realize continuous
Radar target is continuously tracked in variation.
Track following data are transferred to data transmission blocks, data transmission blocks by step 7, vehicle tracking computing module
Track following result is transmitted in data server and in public network;
Step 8, after the data of a frame vehicle are parsed, return step five, industrial personal computer is using udp protocol after continued access
Receive the initial data of millimetre-wave radar monitoring.
Noise eliminating method includes the following steps:
The first step rejects all ambient noises unrelated with items information itself and interferes data.
Second step is analyzed due to data and rejects in module and threshold value can be arranged according to RCS energy value size property, according to RCS
The threshold value of energy value rejects the noise targets unrelated with RCS energy value.
Third step, analyzes to reject in module threshold value is arranged according to distance, speed and position characteristic due to data, according to away from
From, the threshold value of speed and position, reject with apart from the unrelated noise targets of, speed and position.
RNN vehicle target track following algorithm includes the following steps:
The first step, the setting of data format.The collected initial data of millimetre-wave radar is established in polar coordinate system
On the basis of.Both the two-dimensional spatial location of object was indicated with radial distance and angle, for simplicity after mathematical computations, we
By the information of polar coordinate system, two-dimensional space rectangular coordinate system is converted to, is indicated with (x, y) value;
Second step, by the conversion of coordinate system, when previous target data format is that (a (time of origin), x, y, s are (radial
Speed), r (RCS value)), with reference to the limiting value of millimetre-wave radar, i.e. maximum probe radial distance, maximum probe angle, maximum spy
Radial velocity, maximum RCS value are surveyed, data are become into the numerical value in (0,1) section, so as to subsequent processing;
Third step carries out Recognition with Recurrent Neural Network unit.After data processing, it is passed to one group every time by complete after data cleansing
Whole frame, one frame of radar used in us is 32 target informations, it may be assumed that
xi=(ai xi yi si ri)
After being often passed to one group of complete frame, data will carry out returning Recognition with Recurrent Neural Network program, return Recognition with Recurrent Neural Network
After the completion of program, corresponding target vehicle track following data are obtained.In conjunction with the recurrence Recognition with Recurrent Neural Network in Fig. 5, recurrence is followed
Ring neural network includes following content:
A. variable-definition:
xtRepresent the input of the training sample in sequence index t.
htRepresent the hidden state of the model in sequence index t.htBy xtAnd ht-1It codetermines, t=T is last rope
Draw.
otRepresent the output of the model in sequence index t.otThe only hidden state h current by modeltIt determines.
EtRepresent the loss function of the model in sequence index t.
ytRepresent the true output of the training sample sequence in sequence index t.
WxhIt is that N*K weight matrix connects K input unit to N number of implicit layer unit.
WhhIt is N*N weight matrix, connects N number of implicit layer unit from moment t-1 to moment t.
WhyIt is the weight matrix of L*N, is connected to N number of implicit layer unit to L output layer unit.
ut=Wxh*xt+Whh*ht-1+ b is the latent vector of N*1 hidden layer.
ot=Why*ht+ c is the latent vector of L*1 output layer.
B, c are biasing coefficient, and γ is learning rate, f, g activation primitive, generally use tanh function and sigmoid function.
B. input layer:
xtInput pretreated rear millimetre-wave radar data.
C. hidden layer:
Hidden layer calculates:
ht=f (Wxh*xt+Whh*ht-1+b)
D. output layer:
Output layer calculates:
ot=Why*ht+c
Then complete propagated forward are as follows:
E. loss function:
Pass through loss function E(t), such as MSE loss function, we can loss with quantitative model in current location, i.e.,
And ytGap.
For RNN, since we have loss function, final loss E in each position of sequence are as follows:
F. backpropagation:
The more new strategy of weight W:
The o of the defined nucleotide sequence index position ttGradient are as follows:
The h of the defined nucleotide sequence index position ttGradient are as follows:
WhyGradient value calculate:
The gradient value of c calculates:
Wxh,Whh, the gradient calculation expression of b:
Then it follows that
4th step sets evaluator, and it is each that RNN vehicle target track following algorithm, which is divided into multiple each RNN units,
RNN unit binds an evaluator.
If RNN is continued working, evaluator meeting sustained activation, it can't be more than artificially to be arranged most that scoring, which can continue to increase,
Big value.
If RNN does not work for a long time, the continuous decrement that scores can then initialize this RNN when being reduced to minimum value in operation
Unit makes it that new target be waited to occur.
The data of evaluator include the following:
A. parameter
Scoret is present score.
Range is cycle length.
Step is step-length.
statetFor the operating state of the RNN unit of current bindings.0 indicates that this time circulation does not update, and 1 indicates this time to recycle
It has updated.
l(statet,scorett) it is evaluation function, scorett+1=l (statet,scoret), i.e. subsequent time is commented
Divide and is determined by currently scoring with RNN unit operating status.
B. process
It is linear function that evaluation function, which is arranged, in we, and it is 0 point that base value, which had both been arranged, and period maximum value is range,
Work as statetWhen=1:
scoret+!=min (scoret+step,range)
Work as statetWhen=1:
scoret+1=max (scoret+step,0)
Work as scoret+1When≤0: bound RNN unit resetting waits next target for needing to track.
In conjunction with Fig. 6 to Fig. 9, parameter is as follows in the overall flow in RNN algorithm:
A. global parameter
RNN_count initializes the number that RNN returns device.
B. process initializes
After starting algorithm, by the way that global parameter RNN_count is arranged, our test platforms use 40, both one
1/4 redundancy is added in frame on the basis of most target numbers 32.
C.RNN circulation
The circulation of RNN is the circulation carried out according to system time.No matter both can all be recycled either with or without data input, and
And evaluator can be calculated with the time, and corresponding movement is taken to the RNN that it is bound.
D. data are incoming
When incoming data are not classified in systems, new classification needs are created.When incoming data have point in systems
When class, the classification that current goal should belong to first is judged in current class, if it is determined that belonging to a certain classification, then current goal is included into
The classification needs one new classification of creation if being not belonging to current any classification.
E.RNN updates
RNN is updated according to backpropagation process, we have been illustrated above using gradient descent algorithm.
F. circulation output
RNN output is the vector of One-Hot coding, is both labelled with classification of the input in current RNN tracking target.
The specific number for tracking target by finding RNN later, determines the number of input target.
G. the effect of evaluator
When RNN unit for a long time without effect when, it is necessary to RNN unit is discharged, to continue new target of classifying.
The result obtained after calculating is the matrix of n*5, and each row format is [dn,vn,an,rn,IDn], after result is arranged
Be transmitted to data transmission blocks, saved result to data center by data transmission blocks, and by data-pushing to other service
Platform.
In data center and other service platforms, radar data sending module follows the pub-sub interface specification in ZMQ.
Industrial personal computer is as the end pub, other service platforms are as the end sub.Other service platforms can receive data after subscribing to the industrial personal computer
Push Service.After data transmission blocks receiving locus tracking result, result is saved to data center, and by data-pushing to its
His service platform.
Certainly, the above description is not a limitation of the present invention, and the present invention is also not limited to the example above, this technology neck
The variations, modifications, additions or substitutions that the technical staff in domain is made within the essential scope of the present invention also should belong to of the invention
Protection scope.