CN111695737B - LSTM neural network-based group target traveling trend prediction method - Google Patents
LSTM neural network-based group target traveling trend prediction method Download PDFInfo
- Publication number
- CN111695737B CN111695737B CN202010542008.6A CN202010542008A CN111695737B CN 111695737 B CN111695737 B CN 111695737B CN 202010542008 A CN202010542008 A CN 202010542008A CN 111695737 B CN111695737 B CN 111695737B
- Authority
- CN
- China
- Prior art keywords
- target
- group
- prediction
- track
- sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 39
- 238000000034 method Methods 0.000 title claims abstract description 29
- 230000015654 memory Effects 0.000 claims abstract description 22
- 230000000694 effects Effects 0.000 claims abstract description 16
- 238000004891 communication Methods 0.000 claims abstract description 11
- 238000012549 training Methods 0.000 claims description 14
- 230000007787 long-term memory Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 7
- 238000004458 analytical method Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000005065 mining Methods 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000002474 experimental method Methods 0.000 claims description 4
- 230000009467 reduction Effects 0.000 claims description 4
- 238000000547 structure data Methods 0.000 claims description 4
- 230000005284 excitation Effects 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 3
- 230000001105 regulatory effect Effects 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 2
- 230000006403 short-term memory Effects 0.000 claims description 2
- 238000000638 solvent extraction Methods 0.000 claims description 2
- 239000013589 supplement Substances 0.000 claims description 2
- 238000012795 verification Methods 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 abstract description 7
- 230000002452 interceptive effect Effects 0.000 abstract description 7
- 230000000007 visual effect Effects 0.000 abstract description 2
- 230000000295 complement effect Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 3
- 230000033001 locomotion Effects 0.000 description 3
- 230000001537 neural effect Effects 0.000 description 3
- 230000001419 dependent effect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 125000004122 cyclic group Chemical group 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000005484 gravity Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
- 238000007794 visualization technique Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G06Q50/40—
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Game Theory and Decision Science (AREA)
- Tourism & Hospitality (AREA)
- Development Economics (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- General Business, Economics & Management (AREA)
- Probability & Statistics with Applications (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a group target advancing trend prediction method based on an LSTM neural network, which utilizes the relative spatial relation among group targets to model interactive group targets in an actual scene, and establishes the LSTM long-short-time memory neural network to perform track position prediction calculation. And carrying out dependence prediction according to the contact communication among the single targets in the group, the cooperative relationship, the single target model characteristics and the activity properties of the group targets, eliminating unreasonable factors, obtaining the minimum prediction space of the travelling track trend of the group targets, and carrying out visual expression on the final prediction result. Compared with the traditional single target track prediction, the method solves the problem of the prediction of the complex group target advancing trend with certain correlation and interactivity, can comprehensively and comprehensively calculate and predict by fully utilizing the information such as the target historical track, the target association relation, the target model type, the target volume characteristics, the target activity properties and the like, and provides the prediction result of the group target advancing trend as accurately as possible.
Description
Technical Field
The invention relates to a group target traveling trend prediction method based on an LSTM neural network.
Background
At present, the prediction technology of the target advancing trend is widely researched and applied in the fields of unmanned driving, traffic flow, pedestrian track and the like at home and abroad. The method has important significance for accurately predicting the advancing trend and the movement track of the target as soon as possible, grasping the future position of the target and excavating the behavior intention of the target. However, predicting overall traveling trend for a group of group targets with complex association relationships such as contact communication and collaboration still faces a number of problems: 1) The existing prediction technology is used for predicting a single type of target track, but lacks research on predicting the overall advancing trend of multiple types of interactive group targets. 2) Insufficient use of historical track information of a target and insufficient consideration of the dynamic nature of the environment in which the target is located in prediction cause the prediction accuracy to be excessively dependent on the current state of the target; 3) The weight influence of the inherent characteristics of the model, the volume and the like of the single target in the whole travel trend prediction of the target group is ignored. 4) There is a lack of consideration in the impact of target activity on prediction space. 5) There is no effective way to calculate and express the prediction result of the travelling trend of the group targets.
Disclosure of Invention
The invention aims to: aiming at the defects of the prior art, the invention provides a group target advancing trend prediction method based on an LSTM neural network. On the basis, the dependence prediction is carried out according to the information in the aspects of contact communication among single targets, cooperative relation, single target model characteristics, activity properties of the group targets and the like in the group, unreasonable factors are eliminated, the minimum prediction space of the travelling track trend of the group targets is obtained through analysis, and the final prediction result is visually expressed based on the geographical information background.
The invention discloses a group target traveling trend prediction method based on an LSTM neural network, which comprises the following steps:
step 1, preprocessing the track points of each single target in the group;
step 2, establishing an LSTM long short-time memory neural network prediction model, and predicting and calculating the trend of each single target track in the group by taking each single target track sequence in the group as input;
step 3, calculating to obtain a minimum prediction space about the overall advancing trend of the group target;
step 4, calculating a confidence interval of the overall advancing trend of the group according to the single target track prediction result and the minimum prediction space in the group;
and 5, visually designing and displaying a group target traveling trend prediction result based on the two-dimensional geographic information background.
The step 1 comprises the following steps: for a given set of unmanned car groups g= { G 1 ,g 2 ,g 4 ,g 5 ,…g m -wherein m represents the total number of single targets in the group of unmanned cars and m e N * ，N * G is a positive integer m A sequence of track points representing the mth target in the unmanned vehicle group is first extracted for each individual target in the collection G in a given time period t 1 ,t 2 ]The historical track information in the database forms sequence type structure data which is ordered according to time, wherein t 1 Represents the starting time point, t 2 Represents the termination time point, and t 1 <t 2 The method comprises the steps of carrying out a first treatment on the surface of the Secondly, fusing and verifying relevant suspected points according to a track point source mode and a track point similarity relation to obtain a history track point sequence; and finally, interpolating and complementing the historical track point sequence according to the step length requirement of the predicted input data sequence to obtain the density requirement of the predicted input data sequence.
In step 1, the fusion and verification of the related suspected points are performed according to the track point source mode and the track point similarity relationship to obtain a history track point sequence, which comprises the following steps: in the aspect of source mode, establishing a sampling priority rule of a track point source mode, and for one track point of a given single target, reporting the track point in the source mode with the highest sampling priority; if the source modes have the same priority, calculating the average longitude and latitude of the reporting points of all the source modes with the same priority as the track point of the final acquisition.
In step 1, a third-order bezier curve completion is performed on the historical track point sequences to obtain a set of historical track point sequences R= { p 1 (x 1 ,y 1 ),p 2 (x 2 ,y 2 ),p 3 (x 3 ,y 3 ),…p n (x n ,y n ) n.epsilon.N }, where n. * ，p n Represents the nth history trace point, (x) n ,y n ) Is p n Is set according to the predicted input data sequence requirement, and is required to be p n And p is as follows n+1 Interpolation is carried out between the two points, and the original point p is aimed at n And p n+1 Determining two control points c n And d n And c n At p n Before d n At p n Thereafter, c n The coordinates are (x) n +α(x n+1 –x n-1 ),y n +α(y n+1 -y n-1 ))，d n The coordinates are (x) n+1 -β(x n+2 –x n ),y n+1 -β(y n+2 -y n ) Where α, β are constants, then are located at p n And p is as follows n+1 The parameter equation of the third-order Bezier curve is as follows:
wherein s is p n And p is as follows n+1 And (3) the length proportion designated when the track points are interpolated, and each single-target historical track interpolation point complement calculation is performed according to the track prediction step length requirement designated by people.
The step 2 specifically comprises the following steps:
step 2-1, preprocessing each single-target historical track in the group according to the input requirement of the LSTM long-time memory neural network: for a given single target in a group, it is held for a period of time t 1 ,t 2 ]The set of historical track points within is labeled H [t1,t2] ＝{h 1 (x 1 ,y 1 ),h 2 (x 2 ,y 2 ),h 3 (x 3 ,y 3 ),…h n (x n ,y n ) n.epsilon.N }, where n. * ，h n Representing a given single target historical track point set H [t1,t2] N-th track point of (x) n ,y n ) Is h n Coordinates of (c);
for set H [t1,t2] The elements in the data set are subjected to dimension reduction processing to finally form two serialized data sets H lon ＝{x 1 ,x 2 ,x 3 ,…x n Sum H lat ＝{y 1 ,y 2 ,y 3 ,…y n }；
For input sequence H lon And H lat Equal subset division, namely mini-batch operation, is performed, and the time period [ t ] is ensured in the three-order Bezier interpolation completion operation of the historical track 1 ,t 2 ]The inner trace point ensures a certain density ([ t) 1 ,t 2 ]The density of the inner track points is ensured by setting the s value in the third-order Bessel interpolation complement), and the equal amount subset partitioning parametersThat is, input of ++into LSTM long-term memory neural network is regulated each time>Training and adjusting parameters by the samples;
setting a sequence length steps=3;
finally determining the two-dimensional array length of the batch_size stepsAs a set of trainingTraining data quantity;
step 2-2, constructing an LSTM circulating neural network, which comprises an input layer, an LSTM long-short time memory neural network layer and an output layer;
the length of the input data amount in the input layer is determined by the batch_size and steps together, and the input layer has three input elements C t-1 、H t-1 And X t ，C t-1 Is positioned on the main memory line and is used for memorizing the early state in the LSTM cycle, H t-1 Is the predicted result of the input feature vector at the last time point in the training sequence (the training sequence is the sequence formed by the results obtained by each training), X t Is the input feature vector at time t on the training sequence;
in the process of constructing the LSTM long-short-term memory neural network layer, dividing the processing process of the LSTM long-term memory neural network layer into a forgetting stage, a selective memory stage and an output stage, respectively designing corresponding threshold controllers, namely a forgetting gate, an input gate and an output gate, realizing the thresholds by using a sigmoid function, and selecting tan as an excitation length compensating function; wherein the forgetting gate is mainly used for determining which part of the output of the last unit module is to be reserved and forgotten, and is input C by long-term memory at t-1 t-1 Determining a designated forgetting factor; the input door supplements the new attribute information corresponding to the information discarded in the forget door found in the unit; and the output gate outputs a final result according to the unit states determined by the forget gate and the input gate in sequence.
Mapping the output of a plurality of neurons into a (0, 1) interval by using a softmax function through dimension mapping in an output layer (reference document: P hao.jli.incrustation allowed Learning the Hierarchical Softmax Function for Neural Language Models.[ C ]. Aaai.2017.), and obtaining probability distribution of a prediction result as a basis of final selection;
the output layer outputs the neuron state C at the time t in the prediction sequence t Is used for simultaneously outputting the prediction result H t ，H t Is one ofVector of dimensionsThe vector dimension is->Output vector H t ＝(h 1 ,h 2 ,h 3 ,…h m ) Wherein h is m Represents the predicted value of the mth point, and mεN * Then output vector H t The probability distribution of each value of (2) is expressed as: /> Wherein S (h) i ) Representative vector H t I is a positive integer, i=1, 2, … m; finally, the result set S (h i ) Selecting the final prediction result with the maximum probability value;
the step 3 comprises the following steps: for a given target group G, mining the contact communication relation among the single targets in the target group G, identifying the activity property of each single target, constructing an interactive characteristic index system for expressing the single targets, calculating interactive influence indexes, and combining the independent track prediction results of each single target to perform relevant linkage analysis and dependence prediction to obtain the minimum prediction space about the overall advancing trend of the group targets. The method comprises the steps of carrying out single-target association relation mining and frequent item set calculation through an FPTree algorithm, cutting the frequent item set smaller than 2, and cutting an active area according to the activity property of each single target, wherein the probability that the pedestrian directly spans a building area is low if the activity property of the pedestrian is determined, so that the building area can be cut. The target association mining and frequent item set calculation are calculated through an FPTree algorithm, wherein the FPTree is the prior art and is quoted from: i oukid.FPTree: A Hybird SCM-DRAM Persistent and Concurrent B-Tree for Storage Class Memory [ C ]. International Conference on Management of Data.2016.
Step 4 comprises: selecting t distribution to calculate prediction result confidence interval, and extracting points of longitude prediction results in the minimum prediction space range in z single targets as sampling samplesThis is marked as { x } 1 ,x 2 ,..x z Z e N * The average value of the sampling samples is recorded asFor t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating to obtain a confidence interval with the longitude confidence level of 95% in the traveling trend point position of the target group asLabeling the latitude sample set corresponding to the longitude prediction results of the extracted z single targets as { y } 1 ,y 2 ,..y z The sample { y } 1 ,y 2 ,..y z The mean value of } is denoted ∈ ->For t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating a confidence interval of which the latitude confidence level is 95% in the advancing trend point position of the target group as +.>
The step 5 comprises the following steps: based on the two-dimensional geographic information background, the group target traveling trend prediction junction is visually displayed in a mode of combining a circle and a sector.
The beneficial effects are that: compared with the prior art, the invention has the following advantages:
1) Compared with the conventional method for predicting the track of a single target in a single environment, the method provided by the invention provides a prediction technology for overall travel trend of a group target with multiple types and interaction characteristics, and solves the problem of overall state transition prediction of the group target under the characteristic condition.
2) In the aspect of the prediction method, not only the long-term history track information of the predicted objects is fully utilized, but also the interactive characteristic information such as contact communication, cooperative cooperation, activity property and the like among the predicted objects are fully utilized, and the long-term memory neural network prediction method and the interactive characteristic inference prediction are combined to form a multi-factor dependent prediction method, so that the prediction space of the motion track is reduced, the prediction result is more focused, and the overall advancing trend of the group targets is represented as far as possible.
3) In the aspect of prediction result processing and visualization, a single target track prediction result set in a group is used as a sampling sample, the confidence upper limit and the confidence lower limit of the prediction result are calculated, and a visualization method of the overall advancing trend of the group target by combining circles and sectors is designed by combining a geographic information background, so that the effect is visual and concise, and the perception and understanding efficiency of a decision maker on the prediction result is improved.
Drawings
The foregoing and/or other advantages of the invention will become more apparent from the following detailed description of the invention when taken in conjunction with the accompanying drawings and detailed description.
Fig. 1 is a schematic flow chart of the present invention.
FIG. 2 is a schematic diagram of the structure of the LSTM long-short-term memory neural network constructed by the invention.
FIG. 3 is a schematic view of clipping of the minimum prediction space in the present invention.
FIG. 4 is a diagram of a cluster target trend prediction result interval according to the present invention.
Detailed Description
The invention aims to provide a group target advancing trend prediction method based on an LSTM neural network, which models complex and abstract interactive group targets in an actual scene by utilizing relative spatial relations such as relative positions, relative orientations, relative volumes, relative speed and the like among the group targets, preprocesses single target long-term historical track points in the group, and establishes the LSTM long-term memory neural network to perform track position prediction calculation. On the basis, the dependence prediction is carried out according to the information in the aspects of contact communication among single targets, cooperative relation, single target model characteristics, activity properties of the group targets and the like in the group, unreasonable factors are eliminated, the minimum prediction space of the travelling track trend of the group targets is obtained through analysis, and the final prediction result is visually expressed based on the geographical information background. The invention mainly comprises the following steps:
step 1, preprocessing the track points of each single target in the group. Firstly, denoising an original track point; secondly, fusing and verifying a plurality of similar association points to obtain enhancement points; finally, interpolation complement operation is carried out on the track point sequence so as to enhance the track point density in a certain time span.
And 2, establishing an LSTM long and short-time memory neural network prediction model, and predicting and calculating the track trend of each single target in the group by taking each single target track sequence in the group target as input.
And 3, calculating a contact communication relation among the single targets in the group, identifying the activity property of each single target, combing the interaction characteristics among the single targets, and carrying out relevant linkage analysis and dependence prediction by combining the independent track prediction results of each single target to obtain the minimum prediction space about the overall advancing trend of the group targets.
Step 4, calculating a confidence interval of the overall advancing trend of the group according to the single target track prediction result and the minimum prediction space in the group;
and 5, visually designing and displaying the group target traveling trend prediction junction based on the two-dimensional geographic information background.
Examples
In the embodiment, the group target traveling trend prediction method based on the LSTM neural network is provided, and in an experiment for predicting the traveling trend of the unmanned automobile group, the method can ensure that the actual track of the unmanned automobile group falls in a prediction interval and calculate out the minimum traveling trend interval of the group. As shown in fig. 1, the specific implementation steps are as follows:
step 1, for a given target group set g= { G 1 ,g 2 ,g 4 ,g 5 ,…g m -where m e N * Extracting each single target in the collection G in a given time period [ t ] 1 ,t 2 ](t 1 <t 2 ) The historical track information in the track points form sequence type structure data which are ordered according to time, and the sequence type structure data are processed according to the track point source mode and the track point similarity relationAnd fusing and verifying relevant suspected points to obtain a history track point sequence with higher accuracy, and then carrying out third-order Bezier curve completion on the history track point sequence according to the step length requirement of the predicted input data sequence to obtain the density requirement of the predicted input data sequence. For example, for a set of sequences of trajectory points r= { p 1 (x 1 ,y 1 ),p 2 (x 2 ,y 2 ),p 3 (x 3 ,y 3 ),…p n (x n ,y n ) n.epsilon.N }, where n. * According to the predicted input data sequence requirement, the data is needed to be in p n And p is as follows n+1 Interpolation is carried out between the two points, and the original point p is aimed at n And p n+1 Determining two control points c n And d n And c n At p n Before d n At p n Thereafter, c n The coordinates are (x) n +α(x n+1 –x n-1 ),y n +α(y n+1 -y n-1 ))，d n The coordinates are (x) n+1 -β(x n+2 –x n ),y n+1 -β(y n+2 -y n ) Where α, β are constants, typically take a value of 0.25. From this, it can be deduced that p is located n And p is as follows n+1 The parameter equation of the third-order Bezier curve is as follows:
wherein s is p n And p is as follows n+1 The length proportion appointed when the track points are interpolated is taken as the track midpoint between two points, namely s=0.5 in the experiment, so that each single-target historical track interpolation complementary point is calculated according to the track prediction step length requirement.
Step 2, constructing a long-short-time memory cyclic neural network with unit states, performing dimension reduction decomposition on single-target historical track point data in a group according to longitude and latitude, and generating two sequence structures for a given single-target historical track H to represent the single-target historical track H, wherein the two sequence structures are marked as H lon And H lat And by H lon And H lat Input to LSTM long short term memory neural networkThe predictive calculation of the trajectory H is performed as follows:
and 2-1, preprocessing each single-target historical track in the group according to the input requirement of the LSTM long-time memory neural network. For a given single target in a group, it is held for a period of time t 1 ,t 2 ](t 1 <t 2 ) The set of historical track points within is labeled H [t1,t2] ＝{h 1 (x 1 ,y 1 ),h 2 (x 2 ,y 2 ),h 3 (x 3 ,y 3 ),…h n (x n ,y n ) n.epsilon.N }, where n. * The method comprises the steps of carrying out a first treatment on the surface of the First, for set H [t1,t2] The elements in the data set are subjected to dimension reduction processing to finally form two serialized data sets H lon ＝{x 1 ,x 2 ,x 3 ,…x n Sum H lat ＝{y 1 ,y 2 ,y 3 ,…y n n.epsilon.N }, where n. * The method comprises the steps of carrying out a first treatment on the surface of the Second, for the input sequence H lon And H lat Equal subset division, namely mini-batch operation, is performed, and the time period [ t ] is ensured in the three-order Bezier interpolation completion operation of the historical track 1 ,t 2 ](t 1 <t 2 ) The inner track points ensure a certain density, and in order to improve training efficiency, the equal sub-set dividing parameters are usedThat is, input of ++into LSTM long-term memory neural network is regulated each time>The training adjustment parameters are performed on each sample. Third, since the LSTM network is a sequence of multiple identical neural network units stacked together, three consecutive points in the target trajectory can represent the motion trend of the target on the segment of line, so the sequence length steps=3 is specified here. To sum up, the two-dimensional array length of batch_size steps is finally determined>As a set of training data volumes.
And 2-2, constructing an LSTM circulating neural network, wherein the LSTM circulating neural network comprises an input layer, an LSTM long-short time memory neural network layer and an output layer. The length of the amount of input data in the input layer construction is critical and is determined by both the batch_size and steps in 1). As shown in FIG. 2, the input layer has three input elements, C t-1 Is positioned on the main memory line and is used for memorizing the early state in the LSTM cycle, H t-1 Is the prediction result of the input characteristic vector at the last time point in the training sequence, X t The feature vector is input at time t on the training sequence. The input layer combines the earlier memory state and the predicted result, and maintains two transmission intermediate states, namely C t-1 As representative neuronal states and in H t-1 The represented hidden layer state is the basis for the neural network to complete long-term memory and is suitable for the prediction of longer sequence data. In the process of constructing the LSTM neural network layer, the processing process of the LSTM long-short-term memory neural network layer is divided into a forgetting stage, a selective memory stage and an output stage, and corresponding threshold controllers, namely a forgetting gate, an input gate and an output gate, are designed in each stage so as to ensure that information selectively passes through and enter the next link. Each threshold is implemented as a sigmoid function and tan h is selected as the excitation complement function. And mapping the outputs of a plurality of neurons into a (0, 1) interval by using a softmax function through dimension mapping in an output layer, and obtaining the probability distribution of a prediction result as a basis of final selection. In this experiment, the output layer outputs the neuron state C at time t in the predicted sequence t Is used for simultaneously outputting the prediction result H t ,H t Is one ofVector of dimension, note vector dimension +.>Output vector H t ＝(h 1 ,h 2 ,h 3 ,…h m ) Wherein m is N * Then output vector H t The probability distribution of each value of (c) can be expressed as: />Where i is a positive integer, i=1, 2, … m. Finally, the result set S (h i ) And selecting the final prediction result with the maximum probability value.
And 3, for a given target group G, mining the contact communication relation among the single targets in the group G, identifying the activity property of each single target, and carrying out relevant linkage analysis and dependence prediction by combining the independent track prediction results of each single target to obtain the minimum prediction space about the overall advancing trend of the group target. For a given target group set G, a frequent subset of the set of single-target inter-target communication relationships is identified through a group algorithm, as shown in fig. 3, a set E, F, K, H, P is a prediction area divided according to the single-target communication relationships and the single-target track prediction points in the group, wherein a set E is a frequent 2-item set, F is a frequent 1-item set, and setting frequent subset items smaller than 2 in a test can be regarded as a low probability area, so that the E, F area can be inferred to be cut. Meanwhile, the activity quality of the group decides that the region D is a low probability event, and then the region D is clipped in the travel trend range. Finally, by using the center of gravity of the frequent subset area as the center of a circle and determining the intersection points p1 and p2 of the two rays and the circular edge, the arc line p1p2 determines the predicted minimum range of the group travel.
And 4, calculating a confidence interval of the overall group advancing trend under the constraint of the minimum predicting space according to the single target track predicting result and the minimum predicting space of the advancing trend of the group targets in the group. Since the number of single targets in the group is small, t distribution is selected to calculate the confidence interval of the prediction result, and assuming that m single targets exist in the group, the point of the longitude prediction result in the minimum prediction space range in the z single targets is extracted as a sampling sample and marked as { x } 1 ,x 2 ,..x z Z e N * The average value of the sampling samples is recorded asFor t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating to obtain a confidence interval with the longitude confidence level of 95% in the traveling trend point position of the target group asLabeling the latitude sample set corresponding to the longitude prediction results of the extracted z single targets as { y } 1 ,y 2 ,..y z The sample { y } 1 ,y 2 ,..y z The mean value of } is denoted ∈ ->For t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating a confidence interval of which the latitude confidence level is 95% in the advancing trend point position of the target group as +.>
And 5, visually displaying the group target traveling trend prediction junction in a mode of combining a circle and a sector on the basis of a two-dimensional geographic information background. As shown in fig. 4, since the possible range of the next walking trend of the group is 0 ° to 360 °, the range of the final traveling trend radian of the group can be visually expressed in a circle, and the final traveling trend radian range of the group is obtained through the comprehensive calculation of the steps 1,2, 3 and 4, so that a sector, such as a white sector in the figure, can be jointly determined by the radian and the gray circle in fig. 4 to characterize the traveling trend range of the group G.
The invention provides a group target traveling trend prediction method based on an LSTM neural network, and the method and the way for realizing the technical scheme are numerous, the above description is only a preferred embodiment of the invention, and it should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the invention, and the improvements and modifications should be regarded as the protection scope of the invention. The components not explicitly described in this embodiment can be implemented by using the prior art.
Claims (1)
1. The group target traveling trend prediction method based on the LSTM neural network is characterized by being used for an experiment of unmanned automobile group traveling trend prediction, and capable of ensuring that the actual track of the unmanned automobile group falls in a prediction interval and calculating out the minimum traveling trend interval of the group, and comprising the following steps:
step 1, preprocessing the track points of each single target in an unmanned automobile group;
step 2, establishing an LSTM long and short-term memory neural network prediction model, and predicting and calculating the trend of each single target track in the unmanned automobile group by taking each single target track sequence in the unmanned automobile group as input;
step 3, calculating to obtain a minimum prediction space about the overall advancing trend of the group target;
step 4, calculating a confidence interval of the overall advancing trend of the unmanned automobile group according to the single target track prediction result and the minimum prediction space in the unmanned automobile group;
step 5, based on the two-dimensional geographic information background, visually designing and displaying a group target traveling trend prediction result;
the step 1 comprises the following steps: for a given set of unmanned car groups g= { G 1 ,g 2 ,g 4 ,g 5 ,…g m -wherein m represents the total number of single targets in the group of unmanned cars and m e N * ，N * G is a positive integer m A sequence of track points representing the mth target in the unmanned vehicle group is first extracted for each individual target in the collection G in a given time period t 1 ,t 2 ]The historical track information in the database forms sequence type structure data which is ordered according to time, wherein t 1 Represents the starting time point, t 2 Represents the termination time point, and t 1 <t 2 The method comprises the steps of carrying out a first treatment on the surface of the Secondly, fusing and verifying relevant suspected points according to a track point source mode and a track point similarity relation to obtain a historical track point sequence, and finally, interpolating and complementing the historical track point sequence according to the step length requirement of the predicted input data sequence to obtain a density requirement meeting the predicted input data sequence;
in step 1, the fusion and verification of the related suspected points are performed according to the track point source mode and the track point similarity relationship to obtain a history track point sequence, which comprises the following steps: in the aspect of source mode, establishing a sampling priority rule of a track point source mode, and for one track point of a given single target, reporting the track point in the source mode with the highest sampling priority; if the source mode priority is equal, calculating the average longitude and latitude of all the reporting points of the source modes with equal priority as the track point of the final acquisition;
in step 1, a third-order bezier curve completion is performed on the historical track point sequences to obtain a set of historical track point sequences R= { p 1 (x 1 ,y 1 ),p 2 (x 2 ,y 2 ),p 3 (x 3 ,y 3 ),…p n (x n ,y n ) n.epsilon.N }, where n. * ，p n Represents the nth history trace point, (x) n ,y n ) Is p n Is set according to the predicted input data sequence requirement, and is required to be p n And p is as follows n+1 Interpolation is carried out between the two points, and the original point p is aimed at n And p n+1 Determining two control points c n And d n And c n At p n Before d n At p n Thereafter, c n The coordinates are (x) n +α(x n+1 –x n-1 ),y n +α(y n+1 -y n-1 ))，d n The coordinates are (x) n+1 -β(x n+2 –x n ),y n+1 -β(y n+2 -y n ) Where α, β are constants, then are located at p n And p is as follows n+1 The parameter equation of the third-order Bezier curve is as follows:
wherein s is p n And p is as follows n+1 The length proportion appointed during interpolation track points is calculated according to the appointed track prediction step length requirement;
the step 2 specifically comprises the following steps:
step 2-1, preprocessing each single-target historical track in the group according to the input requirement of the LSTM long-time memory neural network: for a given single target in a group, it is held for a period of time t 1 ,t 2 ]The set of historical track points within is labeled H [t1,t2] ＝{h 1 (x 1 ,y 1 ),h 2 (x 2 ,y 2 ),h 3 (x 3 ,y 3 ),…h n (x n ,y n ) n.epsilon.N }, where n. * ，h n Representing a given single target historical track point set H [t1,t2] N-th track point of (x) n ,y n ) Is h n Coordinates of (c);
for set H [t1,t2] The elements in the data set are subjected to dimension reduction processing to finally form two serialized data sets H lon ＝{x 1 ,x 2 ,x 3 ,…x n Sum H lat ＝{y 1 ,y 2 ,y 3 ,…y n }；
For input sequence H lon And H lat Equal subset division, namely mini-batch operation, is performed, and the time period [ t ] is ensured in the three-order Bezier interpolation completion operation of the historical track 1 ,t 2 ]Internal track point guaranteed density, equal subset partitioning parametersThat is, input of ++into LSTM long-term memory neural network is regulated each time>Training and adjusting parameters by the samples;
setting a sequence length steps=3;
finally determining the two-dimensional array length of the batch_size stepsAs a set of training data volumes;
step 2-2, constructing an LSTM circulating neural network, which comprises an input layer, an LSTM long-short time memory neural network layer and an output layer;
the length of the input data amount in the input layer is determined by the batch_size and steps together, and the input layer has three input elements C t-1 、H t-1 And X t ，C t-1 Is positioned on the main memory line and is used for memorizing the early state in the LSTM cycle, H t-1 Is the prediction result of the input characteristic vector at the last time point in the training sequence, X t Is the input feature vector at time t on the training sequence;
in the process of constructing the LSTM long-short-term memory neural network layer, dividing the processing process of the LSTM long-term memory neural network layer into a forgetting stage, a selective memory stage and an output stage, respectively designing corresponding threshold controllers, namely a forgetting gate, an input gate and an output gate, realizing the thresholds by using a sigmoid function, and selecting tan as an excitation length compensating function; wherein the forget gate is used for determining which part of the output of the last unit module is to be reserved and forgotten, and is input C by long-term memory at t-1 t-1 Determining a designated forgetting factor; the input door supplements the new attribute information corresponding to the information discarded in the forget door found in the unit; the output gate outputs a final result according to the unit states determined by the forget gate and the input gate in sequence;
mapping the output of a plurality of neurons into a (0, 1) interval by using a softmax function through dimension mapping in an output layer, and obtaining probability distribution of a prediction result as a basis of final selection;
the output layer outputs the neuron state C at the time t in the prediction sequence t Is used for simultaneously outputting the prediction result H t ，H t Is one ofVector of dimension, note vector dimension +.>Output vector H t ＝(h 1 ,h 2 ,h 3 ,…h m ) Wherein h is m Represents the predicted value of the mth point, and m∈N * Then output vector H t The probability distribution of each value of (2) is expressed as: /> Wherein S (h) i ) Representative vector H t I is a positive integer, i=1, 2, … m; finally, the result set S (h i ) Selecting the final prediction result with the maximum probability value;
the step 3 comprises the following steps: for a given target group G, mining contact communication relations among single targets in the target group G, identifying activity properties of each single target, carrying out relevant linkage analysis and dependency prediction by combining independent track prediction results of each single target to obtain a minimum prediction space about the overall advancing trend of the group target, wherein the FPTree algorithm is used for mining single target association relations and calculating frequent item sets, cutting the frequent item sets smaller than 2, and then cutting an activity area according to the activity properties of each single target;
step 4 comprises: selecting t distribution to calculate prediction result confidence interval, extracting points of longitude prediction results in the minimum prediction space range in z single targets as sampling samples, and marking the points as { x } 1 ,x 2 ,..x z Z e N * The average value of the sampling samples is recorded asFor t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating to obtain a confidence interval with the longitude confidence level of 95% in the traveling trend point position of the target group asLabeling the latitude sample set corresponding to the longitude prediction results of the extracted z single targets as { y } 1 ,y 2 ,..y z The sample { y } 1 ,y 2 ,..y z The mean value of } is denoted ∈ ->For t distribution, adopting a confidence level of 95% and looking up a table by using the degree of freedom as z-1 to obtain a t value marked as theta, and calculating a confidence interval of which the latitude confidence level is 95% in the advancing trend point position of the target group as +.>
The step 5 comprises the following steps: based on the two-dimensional geographic information background, the group target traveling trend prediction junction is visually displayed in a mode of combining a circle and a sector.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010542008.6A CN111695737B (en) | 2020-06-15 | 2020-06-15 | LSTM neural network-based group target traveling trend prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010542008.6A CN111695737B (en) | 2020-06-15 | 2020-06-15 | LSTM neural network-based group target traveling trend prediction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111695737A CN111695737A (en) | 2020-09-22 |
| CN111695737B true CN111695737B (en) | 2023-08-11 |
Family
ID=72480907
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010542008.6A Active CN111695737B (en) | 2020-06-15 | 2020-06-15 | LSTM neural network-based group target traveling trend prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111695737B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113068131B (en) * | 2021-06-01 | 2021-09-10 | 浙江非线数联科技股份有限公司 | Method, device, equipment and storage medium for predicting user movement mode and track |
| CN114039744B (en) * | 2021-09-29 | 2024-02-27 | 中孚信息股份有限公司 | Abnormal behavior prediction method and system based on user feature labels |
| CN114446046A (en) * | 2021-12-20 | 2022-05-06 | 上海智能网联汽车技术中心有限公司 | LSTM model-based weak traffic participant track prediction method |
| CN114124260B (en) * | 2022-01-11 | 2023-12-12 | 中国人民解放军国防科技大学 | Spectrum prediction method, device, medium and equipment based on composite 2D-LSTM network |
| CN114792320A (en) * | 2022-06-23 | 2022-07-26 | 中国科学院自动化研究所 | Trajectory prediction method, trajectory prediction device and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
| CN110262486A (en) * | 2019-06-11 | 2019-09-20 | 北京三快在线科技有限公司 | A kind of unmanned equipment moving control method and device |
| CN110796093A (en) * | 2019-10-30 | 2020-02-14 | 上海眼控科技股份有限公司 | Target tracking method and device, computer equipment and storage medium |
| CN111046919A (en) * | 2019-11-21 | 2020-04-21 | 南京航空航天大学 | Peripheral dynamic vehicle track prediction system and method integrating behavior intents |
-
2020
- 2020-06-15 CN CN202010542008.6A patent/CN111695737B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110135314A (en) * | 2019-05-07 | 2019-08-16 | 电子科技大学 | A kind of multi-object tracking method based on depth Trajectory prediction |
| CN110262486A (en) * | 2019-06-11 | 2019-09-20 | 北京三快在线科技有限公司 | A kind of unmanned equipment moving control method and device |
| CN110796093A (en) * | 2019-10-30 | 2020-02-14 | 上海眼控科技股份有限公司 | Target tracking method and device, computer equipment and storage medium |
| CN111046919A (en) * | 2019-11-21 | 2020-04-21 | 南京航空航天大学 | Peripheral dynamic vehicle track prediction system and method integrating behavior intents |
Non-Patent Citations (1)
| Title |
|---|
| 孙亚.基于车辆轨迹分析的交通场景感知与预测.《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》.2020,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111695737A (en) | 2020-09-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111695737B (en) | LSTM neural network-based group target traveling trend prediction method | |
| Xing et al. | Personalized vehicle trajectory prediction based on joint time-series modeling for connected vehicles | |
| Xing et al. | Energy oriented driving behavior analysis and personalized prediction of vehicle states with joint time series modeling | |
| CN111260118B (en) | Vehicle networking traffic flow prediction method based on quantum particle swarm optimization strategy | |
| CN112347993B (en) | Expressway vehicle behavior and track prediction method based on vehicle-unmanned aerial vehicle cooperation | |
| Scheel et al. | Situation assessment for planning lane changes: Combining recurrent models and prediction | |
| Wirthmüller et al. | Predicting the time until a vehicle changes the lane using LSTM-based recurrent neural networks | |
| CN113780665B (en) | Private car stay position prediction method and system based on enhanced recurrent neural network | |
| Gaikwad | Vehicle Velocity Prediction Using Artificial Neural Networks and Effect of Real-World Signals on Prediction Window | |
| CN112633602B (en) | Traffic congestion index prediction method and device based on GIS map information | |
| Liao et al. | Online prediction of lane change with a hierarchical learning-based approach | |
| Song et al. | Personalized car-following control based on a hybrid of reinforcement learning and supervised learning | |
| CN115923833A (en) | Personifying decision control method and device for vehicle, vehicle and storage medium | |
| Zhao et al. | Interaction-aware personalized trajectory prediction for traffic participant based on interactive multiple model | |
| Fang et al. | Atten-GAN: pedestrian trajectory prediction with gan based on attention mechanism | |
| Thamizhazhagan et al. | AI Based Traffic Flow Prediction Model for Connected and Autonomous Electric Vehicles. | |
| CN113479187A (en) | Layered different-step-length energy management method for plug-in hybrid electric vehicle | |
| Zou et al. | Predicting vehicle lane-changing behavior with awareness of surrounding vehicles using LSTM network | |
| CN116968721A (en) | Predictive energy management method, system and storage medium for hybrid electric vehicle | |
| CN113762464B (en) | Train operation reference curve dynamic generation method based on learning | |
| Wen et al. | Modeling human driver behaviors when following autonomous vehicles: An inverse reinforcement learning approach | |
| CN115495661A (en) | Self-adaptive interest point recommendation method based on long-term and short-term preference of user | |
| Jain et al. | A machine learning pipeline for fuel-economical driving model | |
| Bagdatli et al. | Investigating lane-changing moves of vehicles departing from signalized junction | |
| CN113032688B (en) | Method for predicting access position of social network user at given future time |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |