CN114997297B - Target movement intention reasoning method and system based on multi-level region division - Google Patents
Target movement intention reasoning method and system based on multi-level region division Download PDFInfo
- Publication number
- CN114997297B CN114997297B CN202210582031.7A CN202210582031A CN114997297B CN 114997297 B CN114997297 B CN 114997297B CN 202210582031 A CN202210582031 A CN 202210582031A CN 114997297 B CN114997297 B CN 114997297B
- Authority
- CN
- China
- Prior art keywords
- target
- level
- motion
- intention
- region
- 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
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000012549 training Methods 0.000 claims abstract description 101
- 239000011159 matrix material Substances 0.000 claims abstract description 62
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 33
- 238000002372 labelling Methods 0.000 claims abstract description 15
- 230000008569 process Effects 0.000 claims description 30
- 230000008859 change Effects 0.000 claims description 7
- 230000001131 transforming effect Effects 0.000 claims description 6
- 238000007796 conventional method Methods 0.000 abstract 1
- 238000013528 artificial neural network Methods 0.000 description 9
- 210000004027 cell Anatomy 0.000 description 8
- 230000006870 function Effects 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000012706 support-vector machine Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Classifications
-
- 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
-
- 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
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A30/00—Adapting or protecting infrastructure or their operation
- Y02A30/60—Planning or developing urban green infrastructure
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Feedback Control In General (AREA)
- Image Analysis (AREA)
Abstract
A target movement intention reasoning method and system based on multi-level region division relates to the technical field of movement intention reasoning of moving targets, and aims to solve the problem that the prior knowledge of the target movement intention is lacking in the conventional method so as to not infer the target movement intention. The method comprises the steps of dividing urban environment into multiple levels and multiple areas to form a movement intention set of a moving target; labeling the acquired urban motion trajectories of a plurality of moving targets in a motion intention set to construct a training data set; discretizing the training data set to construct a feature map matrix; inputting the characteristic map matrix into a multi-level target movement intention reasoning model based on a convolutional neural network for training; labeling the urban motion trail of the moving target to be inferred in the motion intention set, discretizing, inputting the labeled urban motion trail into a trained inference model, and obtaining the probability that the moving target goes to each subarea in each level of areas. The method and the device can be applied to the reasoning problem that the destination of the moving target is unknown.
Description
Technical Field
The invention relates to the technical field of target movement intention reasoning, in particular to a target movement intention reasoning method and system based on multistage region division.
Background
The target movement intention reasoning technology mainly comprises two main methods, namely an intention reasoning method based on a generative model and an intention reasoning method based on a discriminant model. In the generated model, typical algorithms include an intention reasoning method based on a Bayesian theory and an intention reasoning method based on a hidden Markov model; in the discriminant model, typical algorithms include an intention reasoning method based on a support vector machine and an intention reasoning method based on a deep neural network. In the intention inference method based on the bayesian theory, it is necessary to first establish a likelihood probability model between the target motion behavior and the target motion intention, and then iteratively infer the motion intention of the target according to the observed target motion state. Although the iterative reasoning mode has a clear reasoning architecture, the problem that the target motion trail cannot be completely used exists. In the intention reasoning method based on the hidden Markov model, modeling is carried out on the movement behavior of the target by using two random processes, wherein one random process is hidden and unobservable, the other random process is observable, in the intention reasoning process, firstly, corresponding hidden Markov models are respectively built for all possible intentions of the target, then, the occurrence probability of the observed movement state of the target is calculated according to the built models, and finally, the movement intention of the target corresponding to the model with the largest output probability is adopted as a final output result. In the intention reasoning method based on the support vector machine, the object movement intention reasoning problem is regarded as a classification problem, the observed object movement state is segmented by searching a hyperplane with the largest segmentation interval, and then the movement intention label corresponding to the observed object movement state is regarded as the deduced object movement intention. The intent inference method based on the hidden Markov model and the support vector machine is difficult to process the high-dimensional state, so that the complex high-dimensional urban environment information is difficult to process. The intention reasoning method based on the deep neural network is to establish an end-to-end intention reasoning network, directly process the original information of the target motion trail and the urban environment, establish a mapping relation between the target motion state and the target motion intention through layer-by-layer coding of the deep neural network, and directly obtain the motion intention of the target from the target motion state by the intention reasoning network after training. On the premise of sufficient training data, the intention reasoning method based on the deep neural network has better intention reasoning performance compared with other methods.
In the above types of intent inference methods, it is assumed that the set of target motion intents is known, i.e., that the set of possible destinations of the target is known. However, for moving targets in urban environments, their possible set of motion intents is generally not available in advance.
Disclosure of Invention
In order to solve the problem that the prior knowledge of the target movement intention set is lacking in the existing method, the invention provides a target movement intention reasoning method and system based on multi-level region division, and the movement intention set of the target is constructed by adopting a mode of multi-level multi-region division on the environment where the target is positioned, so that the movement intention of the target can be deduced under the condition that the target movement intention (namely destination position) set is unknown.
According to an aspect of the present invention, there is provided a target movement intention inference method based on multi-level region division, the method comprising the steps of:
dividing the urban environment into multiple levels of regions, wherein each divided level of regions form a movement intention set of a moving target;
acquiring a plurality of moving target city motion trajectories, and labeling the trajectories in the motion intention set to construct a training data set;
Discretizing the training data set to construct a feature map matrix; the characteristic map matrix is used for representing the motion state characteristics of the moving target related to the urban environment;
Inputting the characteristic map matrix into a multi-stage target movement intention inference model based on a convolutional neural network for training, and obtaining a trained multi-stage target movement intention inference model;
Labeling the moving object urban movement track to be inferred in the movement intention set, discretizing, inputting the labeled moving object urban movement track into a trained multi-stage object movement intention inference model, and obtaining the probability that the moving object goes to each subarea in each stage of areas.
Further, the specific process of discretizing the training data set to construct the feature map matrix is as follows: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
Further, the feature map matrices correspond to assigned grid maps at a plurality of moments, and n1=0.2, n2=0.6 and n3=0.4 are set to be used as matricesCharacteristic map matrix representing time t,/>The definition is as follows:
in the method, in the process of the invention, Representing matrix/>, at time tElement values of the kth row and the first column; c kl denotes a grid cell located in the kth row and the first column; c (B acc) represents the set of grid elements occupied by all accessible areas of the building; c (B inacc) represents the set of grid elements occupied by all inaccessible areas of the building; /(I)Indicating the position/>, where the target is located, at time tAn occupied grid cell; t inf denotes an inference cycle of the movement intention of the target, that is, the movement intention of the target is inferred once every time period T inf according to a change in the movement state of the target.
Further, the multi-stage target movement intention inference model based on the convolutional neural network is established as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is the feature matrix corresponding to the region Q i in the feature map matrix, and the output is the motion intention of the moving target in the region Q i, namely the moving target goes to each sub-region/>, to which the region Q i belongsIs expressed as:
wherein: p (Q i) represents each sub-region to which the moving target is directed to the region Q i Probability of (2); /(I)Representing a feature map matrix corresponding to the region Q i; /(I)Representing an i-th level target movement intention inference network/>Is a parameter of (a).
Further, the convolution neural network-based multi-stage target movement intention inference model determines an ith-stage target movement intention inference network by optimizing the following loss function in the training processParameter/>
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion trail of the m-th moving object city in the training data set before leaving the i-th level region Q i is in the i-th level sub-regionIf yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
Further, the specific process of discretizing the motion trail of the moving target city to be inferred comprises the following steps: in the grid map, the grid unit with the attribute of being capable of entering the building is assigned as N1, the grid unit with the attribute of being incapable of entering the building is assigned as N2, each position point of the movement track of the moving target city to be inferred is obtained in real time, the grid unit with each position point is assigned as N3, and therefore the grid map corresponding to the assigned grid map at different moments is updated in real time and is used as a characteristic map matrix.
According to another aspect of the present invention, there is provided a target movement intention inference system based on multi-level region division, the system comprising:
The motion intention set acquisition module is configured to divide the urban environment into multiple levels and multiple areas, and each divided level of subareas form a motion intention set of the moving target;
the training data acquisition module is configured to acquire a plurality of moving target city movement tracks, and mark the tracks in the movement intention set so as to construct a training data set;
a feature map acquisition module configured to discretize the training data set to construct a feature map matrix; the characteristic map matrix is used for representing the motion state characteristics of the moving target related to the urban environment;
The intention inference model training module is configured to input the characteristic map matrix into a multi-level target movement intention inference model based on a convolutional neural network for training, so as to obtain a trained multi-level target movement intention inference model;
The motion intention reasoning module is configured to mark the motion trail of the moving target city to be inferred in the motion intention set, discretize the motion trail, input the motion trail into a trained multi-stage target motion intention reasoning model, and acquire the probability that the moving target goes to each subarea in each stage of areas.
Further, the specific process of discretizing the training data set in the feature map obtaining module to construct the feature map matrix is as follows: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
Further, the multi-stage target movement intention inference model based on the convolutional neural network in the intention inference model training module is established as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is the feature matrix corresponding to the region Q i in the feature map matrix, and the output is the motion intention of the moving target in the region Q i, namely the moving target goes to each sub-region/>, to which the region Q i belongsIs expressed as:
wherein: p (Q i) represents each sub-region to which the moving target is directed to the region Q i Probability of (2); /(I)Representing a feature map matrix corresponding to the region Q i; /(I)Representing an i-th level target movement intention inference network/>Is a parameter of (a).
Further, the multi-stage target movement intention inference model based on the convolutional neural network in the intention inference model training module determines an ith stage target movement intention inference network by optimizing the following loss function in the training processParameter/>
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion trail of the m-th moving object city in the training data set before leaving the i-th level region Q i is in the i-th level sub-regionIf yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
The beneficial technical effects of the invention are as follows:
According to the invention, when the target motion intention set is unknown, the motion intention of the target can be represented through the constructed multi-level target motion area, and the reasoning of the target motion intention can be realized under the condition of lacking the priori knowledge of the target motion intention set; the multistage target movement intention inference network established based on the convolutional neural network can characterize the mapping relation between the target movement state and the target movement intention, extract effective characteristic information from the fused target movement track and urban environment information, and timely and accurately infer the movement intention of the target through the observed target movement state after training the intention inference network.
Drawings
The invention may be better understood by reference to the following description taken in conjunction with the accompanying drawings, which are included to provide a further illustration of the preferred embodiments of the invention and to explain the principles and advantages of the invention, together with the detailed description below.
Fig. 1 is a flowchart of a target movement intention reasoning method based on multi-level region division according to an embodiment of the present invention.
Fig. 2 is an exemplary diagram of multi-level multi-region division of a urban environment in an embodiment of the invention.
FIG. 3 is an exemplary diagram of discretizing the state of motion of a target and urban environments in an embodiment of the invention.
FIG. 4 is an exemplary diagram of a two-level target motion intent inference network built based on convolutional neural networks in an embodiment of the present invention.
Fig. 5 is a schematic diagram of a loss value change curve of a two-stage target movement intention inference network training process in an embodiment of the invention.
Fig. 6 is an exemplary diagram of an inference process for two-stage motion intention of a moving object in a city in an embodiment of the present invention.
Fig. 7 is a schematic structural diagram of a target motion intention inference system based on multi-level region division according to an embodiment of the present invention.
Detailed Description
In order that those skilled in the art will better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. It is apparent that the described embodiments or examples are only implementations or examples of a part of the invention, not all. All other embodiments or examples, which may be made by one of ordinary skill in the art without undue burden, are intended to be within the scope of the present invention based on the embodiments or examples herein.
The invention provides a target movement intention reasoning method and system based on multi-level region division, wherein the target movement intention refers to a destination position of a target movement track. According to the method, firstly, the urban environment where the target is located is divided into multiple levels and multiple areas according to the characteristics of the urban environment where the target is located and the position where the target is located, and each level of areas after division is used for representing a possible movement intention set of the target; then discretizing the observed target motion trail and the urban environment to obtain a characteristic state after the target motion trail and the urban environment are fused; and finally, establishing a multi-stage target movement intention inference network based on a convolutional neural network, training parameters of the multi-stage target movement intention inference network by using the collected target movement locus data set, and deducing the intention of the target to all stages of subareas according to the observed target movement locus and urban environment information by using the intention inference network after training.
The embodiment of the invention provides a target movement intention reasoning method based on multi-level region division, which comprises the following steps as shown in fig. 1:
dividing the urban environment into multiple levels of regions, wherein each divided level of regions form a movement intention set of a moving target;
Acquiring a plurality of moving target city motion trajectories, and labeling the trajectories in a motion intention set to construct a training data set;
Discretizing the training data set to construct a feature map matrix; the feature map matrix is used for representing the motion state features of the moving target related to the urban environment;
Inputting the characteristic map matrix into a multi-level target movement intention inference model based on a convolutional neural network for training to obtain a trained multi-level target movement intention inference model;
Labeling the urban motion trail of the moving target to be inferred in the motion intention set, discretizing, inputting the labeled urban motion trail into a trained multi-stage target motion intention inference model, and obtaining the probability that the moving target goes to each subarea in each stage of areas.
In this embodiment, optionally, the specific process of discretizing the training data set to construct the feature map matrix is: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
In this embodiment, optionally, a plurality of feature map matrices are set to correspond to the assigned grid map at a plurality of moments, where n1=0.2, n2=0.6, n3=0.4, and the matrix is usedCharacteristic map matrix representing time t,/>The definition is as follows:
in the method, in the process of the invention, Representing matrix/>, at time tElement values of the kth row and the first column; c kl denotes a grid cell located in the kth row and the first column; c (B acc) represents the set of grid elements occupied by all accessible areas of the building; c (B inacc) represents the set of grid elements occupied by all inaccessible areas of the building; /(I)Indicating the position/>, where the target is located, at time tAn occupied grid cell; t inf denotes an inference cycle of the movement intention of the target, that is, the movement intention of the target is inferred once every time period T inf according to a change in the movement state of the target.
In this embodiment, optionally, a multi-level target motion intention inference model based on a convolutional neural network is built as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is the feature matrix corresponding to the region Q i in the feature map matrix, and the output is the motion intention of the moving target in the region Q i, namely the moving target goes to each sub-region/>, to which the region Q i belongsIs expressed as:
wherein: p (Q i) represents each sub-region to which the moving target is directed to the region Q i Probability of (2); /(I)Representing a feature map matrix corresponding to the region Q i; /(I)Representing an i-th level target movement intention inference network/>Is a parameter of (a).
In this embodiment, optionally, the multi-level target motion intention inference model based on convolutional neural network determines the ith level of target motion intention inference network by optimizing the following loss function during trainingParameter/>
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion track of the m-th moving target city in the training data set before leaving the i-th level region Q i is in the i-th level sub-region Q i k, if yes, Y i (m, k) =1, otherwise Y i (m, k) =0; Representing inference of network/>, using i-th level target motion intent The motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
In this embodiment, optionally, the specific process of discretizing the motion trail of the moving target city to be inferred includes: in the grid map, the grid unit with the attribute of being capable of entering the building is assigned as N1, the grid unit with the attribute of being incapable of entering the building is assigned as N2, each position point of the movement track of the moving target city to be inferred is obtained in real time, the grid unit with each position point is assigned as N3, and therefore the grid map corresponding to the assigned grid map at different moments is updated in real time to be used as the characteristic map matrix.
The invention further provides a target movement intention reasoning method based on multi-level region division, which comprises the following steps:
Step one: dividing the urban environment into multiple levels and multiple areas according to the urban environment where the target is and the current position of the target, wherein each level of divided subareas form a movement intention set of the target;
According to the embodiment of the invention, for the urban environment omega where the moving target is located, the position of the target relative to the urban environment is determined And the structural distribution characteristics of the urban environment, dividing the urban environment omega into n-level areas Q 1,Q2,…,Qn with decreasing occupied area; simultaneously decomposing the divided regions of each level Q i (i=1, 2, …, n) into m i mutually non-overlapping subregions/>In the multi-level region after the urban environment Ω is divided, the relationship between the multi-level regions can be expressed as:
wherein: Representing the jth sub-region in the partitioned ith-1 level environment region, i.e./>
Through the above stepwise multi-region division of the urban environment where the target is located, the set of motion intents Q of the target in the urban environment can be expressed as the following formula:
Wherein: q represents a set of constructed multi-level multi-region urban moving object motion intents, wherein the different sub-regions to which each level of regions Q i (i=1, 2, …, n) belong Representing all the movement intents of the target at the current level.
The above multi-level multi-region division of the urban environment where the target is located converts the reasoning about the target movement intention into the reasoning about the intention of each level of subregion of the target going to the division, namely the reasoning about the multi-level target movement intention means that the target is deduced to each level of subregion according to the observation state thetaSports intention/>
Fig. 2 shows a division process of the urban environment Ω where the target is located, which divides each level of area into 4 sub-areas of the same size. The sub-regions may be different in size.
Step two: discretizing the observed target motion trail and urban environment, and fusing the two information to construct a characteristic state for target motion intention reasoning;
According to the embodiment of the invention, first, the urban environment Ω where the moving object is located is discretized, and decomposed into C X×CY grid units with equal area, where C X and C Y respectively represent the number of grid units in the X-axis direction and the Y-axis direction.
Then, defining a characteristic map matrix integrating the target motion trail and the urban environmentFor representing object motion state characteristics related to urban environment,/>The definition is as follows:
wherein: Representing the motion state characteristic matrix/>, of the target at the moment t Element values of the kth row and the first column; c kl represents grid cells located in the kth row and the first column in the discretized urban environment; c (B acc) represents the set of grid elements occupied by all accessible areas of the building; c (B inacc) represents the set of grid elements occupied by all inaccessible areas of the building; /(I)Indicating the position/>, where the target is located, at time tAn occupied grid cell; t inf represents an inference cycle of the motion intention of the constructed multi-level target, that is, the motion intention of the target is inferred once every time period T inf according to the change of the motion state of the target.
Fig. 3 shows a process of discretizing a target motion trajectory and an urban environment. Wherein an urban environment area with an actual area of 600m×500m is discretized into a discretized urban map with 60×50 grid cells, i.e. the area size of each grid cell is 10m×10m, c X＝60,CY =50. The time intervals between the adjacent target intention inference positions shown in the figure are all 40s, i.e., the target movement intention inference period T inf =40 s. It can be seen from the graph that after the target moves 40s, the movement state of the target is obviously changed, that is, the change of the position of the target is obvious, so that the movement intention of the target can be inferred through the change trend of the movement state of the target.
Step three: aiming at the multi-level target movement intention set established in the first step, respectively establishing target movement intention reasoning networks corresponding to all levels of areas;
According to the embodiment of the invention, aiming at each level of areas Q i(Qi epsilon Q, i=1, 2, … and n which are established in the step one and represent the target movement intention, corresponding target movement intention reasoning networks are established based on convolutional neural networks respectively Wherein the network/>, is inferred for the intent of the i-th level region Q i The input feature state of (a) is feature matrix/>Feature matrix/>, corresponding to middle region Q i The output is the movement intention of the target in the ith stage area Q i, namely the target goes to each subarea/>, which the area Q i belongs toThe probability size of (2) can be expressed as:
Wherein: p (Q i) represents each sub-region to which the target-going region Q i belongs Probability of (2), i.eRepresenting an i-th level target movement intention inference network/>Is a parameter of (a).
Fig. 4 shows a two-level target movement intention inference network established for the first-level region and the second-level region divided in fig. 1. In the figure, the first-stage target movement intention reasoning network and the second-stage target movement intention reasoning network are similar in structure, and the difference is that the first-stage target movement intention reasoning networkIs input into the feature matrix corresponding to the first level region Q 1 While the second level target movement intention reasoning network/>Input of (a) is a feature matrix/>, corresponding to the second-stage region Q 2
The network is inferred by the first-level target movement intentionTo illustrate the structure of the network. The first-stage target movement intention inference network is composed of 5 layers of neural networks, wherein the first two layers of neural networks are two-dimensional convolution neural networks, and the feature matrix/>, through convolution operation, is formedThe method comprises the steps of extracting key characteristic information from a first layer of convolutional neural network, wherein the first layer of convolutional neural network is provided with 4 two-dimensional convolutional kernels, the size of the convolutional kernels is (2, 2) and the convolutional sliding step length is 1, and the second layer of convolutional neural network is also provided with 4 two-dimensional convolutional kernels, the size of the convolutional kernels is (2, 2) and the convolutional sliding step length is 2. The three-layer neural network is a fully-connected network with the number of neurons being 100, 100 and 4 respectively, and is used for further processing the characteristic information extracted by the two-layer convolutional neural network to finally obtain the target which goes to each subarea/>Probability size/>In the built two-stage target movement intention reasoning network structure, the activation functions of the convolutional neural network and the first two layers of fully-connected networks are all ReLU, the activation function of the output layer is Softmax, and the purpose is to limit the value range of the output value of the network between (0, 1) so that the value range accords with the representation range of the probability value.
Step four: determining a multi-level target movement intention inference network established in the third stepParameter/>
According to an embodiment of the invention, the network structure is determined by optimizing the following loss functionParameter/>The following formula is shown:
Wherein: n D represents the number of the target motion trail in the training data; m m represents the number of position points of the mth motion trail in the training data; y i (m, k) is a flag bit indicating whether the last position point of the mth motion trajectory in the training data set before leaving the ith stage region Q i is in the ith stage sub-region If yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe inferred movement track of the mth item mark in the training data set goes to the ith sub-area/>, at the jth position pointΛ is a positive coefficient.
Step five: and (3) utilizing the trained multi-stage target movement intention inference network to infer the movement intention of the target, namely obtaining the probability that the target goes to each sub-region in each stage of region according to the observed target movement state, wherein the probability is shown in the following formula:
the following experiment is further adopted to verify the technical effect of the invention.
The correctness and rationality of the invention are verified by adopting a digital simulation mode. A virtual urban environment is first constructed in the Python environment, as shown in the first level of area in fig. 2. Wherein there are 3 kinds of buildings, the inaccessible building represents a building into which the target cannot enter, the accessible building represents a building into which the target can enter, the target building represents a possible destination position of the target, the target moves from a starting point to a true destination position thereof, the speed of the target outside the building is 2m/s, and the speed inside the building is 1m/s. The reasoning of the target movement intention is to infer each level of subareas where the real destination position of the target is located according to the observed target movement track and the urban environment. The simulation test software environment of the invention is Windows 10+Python3.7, and the hardware environment is AMD Ryzen 5 3550H CPU+16.0GB RAM.
Experiments first, training the two-stage target movement intention inference network established in fig. 4, and a loss value curve in the training process is shown in fig. 5. As can be seen from the graph, the training process goes through 500 training periods in total, and in the early training period, that is, when the training period is less than 200, the loss value of the first-stage intention inference network and the loss value of the second-stage intention inference network decrease faster along with the increase of the training period, which indicates that the network is quickly learning parameters; in the later training period, namely when the training period is more than 200, the loss value of the first-stage intention inference network and the loss value of the second-stage intention inference network are gradually reduced along with the increase of the training period, so that the training process of the network is gradually converged; at the end of training, i.e. when the training period is greater than 400, the loss value of the first-stage intent inference network is substantially unchanged from the loss value of the second-stage intent inference network, indicating that the network training process has substantially converged. The training process above illustrates that the two-stage target movement intention inference network established by the invention can learn stable network parameters through training data.
The present invention then verifies the effectiveness of the method of the present invention by once reasoning about the two-level target motion intent represented by the first-level region and the second-level region established in fig. 2. The reasoning process is shown in fig. 6. From the figure, when t=100deg.S, the target is in the sub-region of the first level regionThe target deduced by the observed target track goes to each subarea/>, of the first-stage areaProbability size of (2) that the target is going to each sub-region of the second level regionAs shown in FIG. 6 (a), the first-level sub-region most likely to be targeted at this time is known as/>, from the reasoning resultThe secondary subregion to which the target is most likely to go is/>As the target continues to move, when t=200s, the target continues to be in the sub-region/>, of the first level regionThe first-level sub-region to which the target is most likely to go, deduced by the observed target motion trail, is/>The secondary subregion to which the target is most likely to go is/>When t=300 s, the target just enters the first level subregion/>In this case, the first level subregion to which the inferred target is most likely to go is/>Due to the target being in the region/>The motion trail in the method is less, so that the target in the area/>, cannot be deduced exactlyIs a motion intention in the middle. When t=400 s, the target is in the sub-region/>, of the first order regionThe first level sub-region to which the inferred target is most likely to go is/>The secondary subregion to which the target is most likely to go is region/>The sub-region/>Since the target real destination location is located in the first level sub-region/>The sub-region/>Therefore, the area where the target real destination is correctly inferred at this time. When t=467s, the target reached its true destination location, confirming the reasoning results at t=400 s.
According to the reasoning process, when the target destination position set is unknown, the method can reasonably represent the target movement intention set, and the movement intention of the target at different moments is deduced through the observed target track, and the area where the target real destination position is located is deduced before the target reaches the real destination position. According to the method, the inference of the movement intention of the urban moving target can be realized, and a new technical thought is provided for the implementation mode of the inference of the movement intention when the position set of the destination of the moving target is unknown.
Another embodiment of the present invention provides a target movement intention inference system based on multi-level region division, as shown in fig. 7, the system comprising:
A movement intention set acquisition module 10 configured to divide the urban environment into multiple levels of multiple regions, each level of divided sub-regions constituting a movement intention set of the moving object;
A training data acquisition module 20 configured to acquire a plurality of moving target city motion trajectories, labeling the trajectories in a set of motion intents to construct a training data set;
A feature map acquisition module 30 configured to discretize the training data set to construct a feature map matrix; the feature map matrix is used for representing the motion state features of the moving target related to the urban environment;
The intention inference model training module 40 is configured to input the feature map matrix into a multi-stage target movement intention inference model based on a convolutional neural network for training, so as to obtain a trained multi-stage target movement intention inference model;
The motion intention inference module 50 is configured to label the motion trail of the moving target city to be inferred in the motion intention set and discretize the motion trail, input the motion trail into the trained multi-stage target motion intention inference model, and acquire the probability that the moving target goes to each sub-region in each stage region.
In this embodiment, optionally, the specific process of discretizing the training data set in the feature map obtaining module 30 to construct the feature map matrix is as follows: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
In this embodiment, optionally, the multi-level target motion intention inference model based on the convolutional neural network in the intention inference model training module 40 is established as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is the feature matrix corresponding to the region Q i in the feature map matrix, and the output is the motion intention of the moving target in the region Q i, namely the moving target goes to each sub-region/>, to which the region Q i belongsIs expressed as:
wherein: p (Q i) represents each sub-region to which the moving target is directed to the region Q i Probability of (2); /(I)Representing a feature map matrix corresponding to the region Q i; /(I)Representing an i-th level target movement intention inference network/>Is a parameter of (a).
In this embodiment, optionally, the multi-stage target motion intention inference model based on convolutional neural network in the intention inference model training module 40 determines the i-th stage target motion intention inference network by optimizing the following loss function during trainingParameter/>
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion trail of the m-th moving object city in the training data set before leaving the i-th level region Q i is in the i-th level sub-regionIf yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
The function of the target motion intention inference system based on multi-level region division in this embodiment may be described by the aforementioned target motion intention inference method based on multi-level region division, so that details of this embodiment are not described, and reference may be made to the above method embodiments, which are not described herein.
While the invention has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of the above description, will appreciate that other embodiments are contemplated within the scope of the invention as described herein. The disclosure of the present invention is intended to be illustrative, but not limiting, of the scope of the invention, which is defined by the appended claims.
Claims (10)
1. The target movement intention reasoning method based on multi-level region division is characterized by comprising the following steps of:
dividing the urban environment into multiple levels of regions, wherein each divided level of regions form a movement intention set of a moving target;
acquiring a plurality of moving target city motion trajectories, and labeling the trajectories in the motion intention set to construct a training data set;
Discretizing the training data set to construct a feature map matrix; the characteristic map matrix is used for representing the motion state characteristics of the moving target related to the urban environment;
Inputting the characteristic map matrix into a multi-stage target movement intention inference model based on a convolutional neural network for training, and obtaining a trained multi-stage target movement intention inference model;
Labeling the moving object urban movement track to be inferred in the movement intention set, discretizing, inputting the labeled moving object urban movement track into a trained multi-stage object movement intention inference model, and obtaining the probability that the moving object goes to each subarea in each stage of areas.
2. The target motion intention reasoning method based on multi-level region division according to claim 1, wherein the specific process of discretizing the training data set to construct the feature map matrix is as follows: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
3. The method for reasoning the target movement intention based on multi-level regional division according to claim 2, wherein the feature map matrices correspond to grid maps assigned at a plurality of moments, and n1=0.2, n2=0.6, n3=0.4 are set by using the matrixCharacteristic map matrix representing time t,/>The definition is as follows:
in the method, in the process of the invention, Representing matrix/>, at time tElement values of the h-th row and the first column of the medium; c hl denotes a grid cell located in the h row and the first column; c (B acc) represents the set of grid elements occupied by all accessible areas of the building; c (B inacc) represents the set of grid elements occupied by all inaccessible areas of the building; /(I)Indicating the position/>, where the target is located, at time tAn occupied grid cell; t inf denotes an inference cycle of the movement intention of the target, that is, the movement intention of the target is inferred once every time period T inf according to a change in the movement state of the target.
4. The target motion intention inference method based on multi-level regional division according to claim 3, wherein the multi-level target motion intention inference model based on convolutional neural network is established as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is a feature matrix corresponding to the region Q i in the feature map matrix, the output is the motion intention of the moving target in the region Q i, that is, the probability that the moving target goes to each sub-region to which the region Q i belongs, expressed as:
Wherein: p (Q i) represents the probability that the moving target goes to each sub-region to which the region Q i belongs; Representing a feature map matrix corresponding to the region Q i; /(I) Representing an i-th level target movement intention inference network/>Is a parameter of (a).
5. The method for reasoning the target motion intention based on multi-level regional division as recited in claim 4, wherein the multi-level target motion intention reasoning model based on convolutional neural network determines the i-th level target motion intention reasoning network by optimizing the following loss function during trainingParameter/>
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion trail of the m-th moving object city in the training data set before leaving the i-th level region Q i is in the i-th level sub-regionIf yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
6. The target motion intention inference method based on multi-level regional division according to any one of claims 1 to 5, wherein the specific process of discretizing the motion trail of the moving target city to be inferred comprises: in the grid map, the grid unit with the attribute of being capable of entering the building is assigned as N1, the grid unit with the attribute of being incapable of entering the building is assigned as N2, each position point of the movement track of the moving target city to be inferred is obtained in real time, the grid unit with each position point is assigned as N3, and therefore the grid map corresponding to the assigned grid map at different moments is updated in real time and is used as a characteristic map matrix.
7. A target motion intention inference system based on multi-level region division, comprising:
The motion intention set acquisition module is configured to divide the urban environment into multiple levels and multiple areas, and each divided level of subareas form a motion intention set of the moving target;
the training data acquisition module is configured to acquire a plurality of moving target city movement tracks, and mark the tracks in the movement intention set so as to construct a training data set;
a feature map acquisition module configured to discretize the training data set to construct a feature map matrix; the characteristic map matrix is used for representing the motion state characteristics of the moving target related to the urban environment;
The intention inference model training module is configured to input the characteristic map matrix into a multi-level target movement intention inference model based on a convolutional neural network for training, so as to obtain a trained multi-level target movement intention inference model;
The motion intention reasoning module is configured to mark the motion trail of the moving target city to be inferred in the motion intention set, discretize the motion trail, input the motion trail into a trained multi-stage target motion intention reasoning model, and acquire the probability that the moving target goes to each subarea in each stage of areas.
8. The system for reasoning the target movement intention based on multi-level regional division according to claim 7, wherein the specific process of discretizing the training data set in the feature map acquisition module to construct the feature map matrix is as follows: transforming the motion intention set after labeling the motion trail into a grid map; in the grid map, assigning a grid unit with an attribute of being capable of entering a building as N1, assigning a grid unit with an attribute of being incapable of entering the building as N2, and assigning grid units with a plurality of position points of each moving target city motion track in a training data set as N3;0< N1<1,0< N2<1,0< N3<1, and N1, N2, N3 are all unequal; thereby obtaining a plurality of feature map matrices.
9. The multi-level regional division-based target motion intention inference system of claim 8, wherein the multi-level target motion intention inference model based on convolutional neural network in the intention inference model training module is built as follows:
For each level of region Q i for representing the target movement intention, respectively establishing an ith level of target movement intention inference network corresponding to each level of region Q i based on the convolutional neural network The input is a feature matrix corresponding to the region Q i in the feature map matrix, the output is the motion intention of the moving target in the region Q i, that is, the probability that the moving target goes to each sub-region to which the region Q i belongs, expressed as:
Wherein: p (Q i) represents the probability that the moving target goes to each sub-region to which the region Q i belongs; Representing a feature map matrix corresponding to the region Q i; /(I) Representing an i-th level target movement intention inference network/>Is a parameter of (a).
10. The multi-level regional division based target motion intent inference system of claim 9, wherein the multi-level target motion intent inference model based on convolutional neural network in the intent inference model training module determines the i-th level target motion intent inference network by optimizing the following loss function during trainingParameters of (2)
Wherein: n D represents the number of the motion trail of the moving target city in the training data set; m m represents the number of position points of the motion trail of the M-th moving target city in the training data set; y i (m, k) is a flag bit, which indicates whether the last position point of the motion trail of the m-th moving object city in the training data set before leaving the i-th level region Q i is in the i-th level sub-regionIf yes, Y i (m, k) =1, otherwise Y i (m, k) =0; /(I)Representing inference of network/>, using i-th level target motion intentThe motion trail of the m-th moving target city in the deduced training data set goes to the i-th level subarea/>, at the j-th position pointProbability of (2); lambda is a positive coefficient.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210582031.7A CN114997297B (en) | 2022-05-26 | 2022-05-26 | Target movement intention reasoning method and system based on multi-level region division |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210582031.7A CN114997297B (en) | 2022-05-26 | 2022-05-26 | Target movement intention reasoning method and system based on multi-level region division |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114997297A CN114997297A (en) | 2022-09-02 |
| CN114997297B true CN114997297B (en) | 2024-05-03 |
Family
ID=83028354
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210582031.7A Active CN114997297B (en) | 2022-05-26 | 2022-05-26 | Target movement intention reasoning method and system based on multi-level region division |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114997297B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109165344A (en) * | 2018-08-06 | 2019-01-08 | 百度在线网络技术（北京）有限公司 | Method and apparatus for pushed information |
| 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 |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN112797995A (en) * | 2020-12-17 | 2021-05-14 | 北京工业大学 | Vehicle emergency navigation method with space-time characteristic situation information |
| CN113435644A (en) * | 2021-06-25 | 2021-09-24 | 天津大学 | Emergency prediction method based on deep bidirectional long-short term memory neural network |
| WO2022022721A1 (en) * | 2020-07-31 | 2022-02-03 | 商汤集团有限公司 | Path prediction method and apparatus, device, storage medium, and program |
| CN114049602A (en) * | 2021-10-29 | 2022-02-15 | 哈尔滨工业大学 | Escape target tracking method and system based on intention reasoning |
| CN114067552A (en) * | 2021-11-08 | 2022-02-18 | 山东高速建设管理集团有限公司 | Pedestrian crossing track tracking and predicting method based on roadside laser radar |
| CN114283576A (en) * | 2020-09-28 | 2022-04-05 | 华为技术有限公司 | Vehicle intention prediction method and related device |
-
2022
- 2022-05-26 CN CN202210582031.7A patent/CN114997297B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109165344A (en) * | 2018-08-06 | 2019-01-08 | 百度在线网络技术（北京）有限公司 | Method and apparatus for pushed information |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device 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 |
| WO2022022721A1 (en) * | 2020-07-31 | 2022-02-03 | 商汤集团有限公司 | Path prediction method and apparatus, device, storage medium, and program |
| CN114283576A (en) * | 2020-09-28 | 2022-04-05 | 华为技术有限公司 | Vehicle intention prediction method and related device |
| CN112797995A (en) * | 2020-12-17 | 2021-05-14 | 北京工业大学 | Vehicle emergency navigation method with space-time characteristic situation information |
| CN113435644A (en) * | 2021-06-25 | 2021-09-24 | 天津大学 | Emergency prediction method based on deep bidirectional long-short term memory neural network |
| CN114049602A (en) * | 2021-10-29 | 2022-02-15 | 哈尔滨工业大学 | Escape target tracking method and system based on intention reasoning |
| CN114067552A (en) * | 2021-11-08 | 2022-02-18 | 山东高速建设管理集团有限公司 | Pedestrian crossing track tracking and predicting method based on roadside laser radar |
Non-Patent Citations (4)
| Title |
|---|
| "基于OKNN的目标战术意图识别方法";刘秋辉等;《现代防御技术》;20211231;第第49卷卷(第第03期期);全文 * |
| "基于全连接神经网络的空战目标威胁评估方法研究";翟翔宇;《中国优秀硕士电子期刊网》;20201115(第第11期期);全文 * |
| "机载相控阵雷达工作模式识别与行为意图预测技术研究";惠晓龙;《中国优秀硕士电子期刊网》;20220415(第第04期期);全文 * |
| 基于深度神经网络的空中目标作战意图识别;周旺旺;姚佩阳;张杰勇;王勋;魏帅;;航空学报;20180827(第11期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114997297A (en) | 2022-09-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109636049B (en) | Congestion index prediction method combining road network topological structure and semantic association | |
| CN112382082B (en) | Method and system for predicting traffic running state in congested area | |
| CN115270956B (en) | Continuous learning-based cross-equipment incremental bearing fault diagnosis method | |
| CN111241952A (en) | Reinforced learning reward self-learning method in discrete manufacturing scene | |
| CN112820105B (en) | Road network abnormal area processing method and system | |
| CN109840595B (en) | Knowledge tracking method based on group learning behavior characteristics | |
| CN111967271A (en) | Analysis result generation method, device, equipment and readable storage medium | |
| CN115048870A (en) | Target track identification method based on residual error network and attention mechanism | |
| Abdellah et al. | VANET traffic prediction using LSTM with deep neural network learning | |
| CN113159403A (en) | Method and device for predicting pedestrian track at intersection | |
| CN105955921B (en) | Robot Hierarchical reinforcement learning initial method based on automatic discovery abstract action | |
| CN114707641A (en) | Training method, device, equipment and medium for neural network model of double-view diagram | |
| CN110908384A (en) | Formation navigation method for distributed multi-robot collaborative unknown random maze | |
| Zheng et al. | Real‐time driving style classification based on short‐term observations | |
| CN114065003A (en) | Network structure searching method, system and medium oriented to super large searching space | |
| CN114997297B (en) | Target movement intention reasoning method and system based on multi-level region division | |
| CN112528556B (en) | Micro-electro-mechanical system design optimization method based on integrated model assisted social learning particle swarm algorithm | |
| CN117312770A (en) | Aero-engine residual life prediction method based on multi-level graph feature fusion | |
| CN113312442B (en) | Smart city electronic map generation method and system | |
| Tamura et al. | Recurrent type ANFIS using local search technique for time series prediction | |
| CN112613227B (en) | Model for predicting remaining service life of aero-engine based on hybrid machine learning | |
| CN115019342A (en) | Endangered animal target detection method based on class relation reasoning | |
| CN114881308A (en) | Internet vehicle speed prediction method based on meta-learning | |
| CN111546327A (en) | Method, apparatus and computer program for determining a motion or trajectory of a robot | |
| Rizki et al. | Application self-organizing maps and particle swarm optimization for clustering |
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 |