CN116383685A - Vehicle lane change detection method based on space-time interaction diagram attention network - Google Patents
Vehicle lane change detection method based on space-time interaction diagram attention network Download PDFInfo
- Publication number
- CN116383685A CN116383685A CN202310219674.XA CN202310219674A CN116383685A CN 116383685 A CN116383685 A CN 116383685A CN 202310219674 A CN202310219674 A CN 202310219674A CN 116383685 A CN116383685 A CN 116383685A
- Authority
- CN
- China
- Prior art keywords
- vehicle
- attention
- time
- lane change
- output
- 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.)
- Pending
Links
- 230000008859 change Effects 0.000 title claims abstract description 49
- 230000003993 interaction Effects 0.000 title claims abstract description 24
- 238000001514 detection method Methods 0.000 title claims abstract description 15
- 238000010586 diagram Methods 0.000 title claims abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 35
- 230000007246 mechanism Effects 0.000 claims abstract description 20
- 238000010276 construction Methods 0.000 claims abstract description 8
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 32
- 230000006399 behavior Effects 0.000 claims description 26
- 239000013598 vector Substances 0.000 claims description 24
- 239000010410 layer Substances 0.000 claims description 21
- 239000011159 matrix material Substances 0.000 claims description 15
- 238000013528 artificial neural network Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 238000009826 distribution Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000000306 recurrent effect Effects 0.000 claims description 6
- 238000004140 cleaning Methods 0.000 claims description 5
- 238000011156 evaluation Methods 0.000 claims description 4
- 238000001914 filtration Methods 0.000 claims description 4
- 230000007774 longterm Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 239000002356 single layer Substances 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 230000008034 disappearance Effects 0.000 claims description 2
- 238000002474 experimental method Methods 0.000 claims description 2
- 238000000605 extraction Methods 0.000 claims description 2
- 230000009467 reduction Effects 0.000 claims description 2
- 238000003860 storage Methods 0.000 claims description 2
- 230000002159 abnormal effect Effects 0.000 claims 1
- 238000005259 measurement Methods 0.000 claims 1
- 230000008447 perception Effects 0.000 claims 1
- 238000003064 k means clustering Methods 0.000 abstract description 4
- 238000000547 structure data Methods 0.000 abstract description 2
- 230000001133 acceleration Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 241000339782 Tomares Species 0.000 description 1
- 208000027418 Wounds and injury Diseases 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006378 damage Effects 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 208000014674 injury Diseases 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011273 social behavior Effects 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- 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
- 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
Abstract
A vehicle lane change detection method based on a space-time interaction diagram attention network comprises the steps of firstly preprocessing a vehicle track, and then quantifying the degrees of different driving styles of surrounding vehicles by introducing a social value guiding theory and a k-means clustering method. And (3) allocating different influence weights for the influence degrees of different driving styles on surrounding vehicles, and embedding the influence weights into the space interaction construction module. The GAT is used for operating the graph structure data in the space interaction construction module, and the characteristics of each graph node are calculated by focusing on adjacent nodes so as to capture the space dependence of different vehicles in the traffic scene. In the time series prediction module, then, using an attention mechanism based LSTM, the model can be made to capture key parts in the timing characteristics by introducing an attention mechanism. Finally, the Softmax function is used for obtaining the probability of different lane changing intentions of each vehicle, and the maximum value of the probability is used as a result of identifying the lane changing intentions of the vehicle. The invention improves the prediction precision of the lane change intention of the vehicle.
Description
Technical Field
The invention relates to the field of intelligent traffic system safety, in particular to a vehicle lane change detection method, which has important significance in building intelligent traffic and guaranteeing travel safety.
Background
With the acceleration of industrialization progress, the perfection of road transportation networks and the rapid development of automobile manufacturing industry, the average automobile possession rises year by year, and the traffic safety problem brought along with the average automobile possession also becomes the focus of social group attention. According to the statistics of world health organization, the annual road traffic death number reaches 135 ten thousand people, and the road traffic injury is now the main cause of teenager death. Where lane change behavior of the vehicle is one of the most basic high risk driving behaviors. The lane change behavior of the vehicle has a higher requirement for the judgment ability and the operation ability of the driver than the lane keeping behavior. In a traffic scene of man-machine mixing, if an automatically driven vehicle cannot correctly recognize the lane change intention of surrounding vehicles, traffic accidents are likely to occur. Therefore, how to accurately and rapidly identify the lane change behavior of surrounding driving vehicles has important significance for relieving road congestion, ensuring vehicle safety and improving vehicle travel efficiency.
In recent years, many researchers have proposed a series of models of different structures to study the lane-changing behavior of vehicles. Previous methods have primarily divided vehicle lane behavior by rule. The earliest lane-change rule model was the Gipps model, which defined the lane-change process as a series of rules, used a decision tree and output the results as binary classification results. The MOBIL model proposed by Kesting et al determines whether to change lanes by specifying utility and safety rules, and is suitable for following, freely changing and forced changing of various rules. In addition, there are minimum safe distance models that determine safe lane change conditions by specifying minimum safe distance rules between vehicles. However, this type of rule model has the disadvantage that the driving environment conditions under consideration are relatively simple and the influence of the driving behavior of the driver on the lane change is not excessively taken into account. Researchers have then employed machine learning methods, such as Tomar et al and Deng et al, to predict the lane change behavior of vehicles using Support Vector Machines (SVM) and Hidden Markov Models (HMM), respectively. These models take as input the position, speed or other factors of the vehicle and estimate the maneuver predictions as a function. However, the method has the defects of low precision, poor robustness and the like. With the rapid rise of artificial intelligence research and the development of various deep learning network models. Many researchers have begun to attempt to enhance the ability of models to predict vehicle lane changes using different deep learning methods. The mainstream deep learning approach is to use LSTM, such as Su et al, to predict the lane change intention of a vehicle by applying a surrounding-aware long-short-term memory (LSTM) algorithm that uses the past trajectory of the vehicle and the current state of the neighbors. However, such methods model the vehicle as a time dependent model based solely on its historical characteristics, and do not take into account how better interactions with surrounding vehicles are captured.
Disclosure of Invention
The invention provides a vehicle lane change detection method based on a space-time interaction diagram attention network, which aims to overcome the defects of the existing vehicle lane change detection model. According to the invention, the driving styles of different vehicles are analyzed based on the K-means clustering method and the social value guiding theory and embedded into the model, the space-time dependence of the vehicles is captured through the GAT and the LSTM based on the attention mechanism, and the lane change intention of surrounding vehicles is detected, so that the self-vehicle can make safer driving decisions.
The nature and characteristics of the graph structure enable the graph structure to more comprehensively and accurately express potential interaction relations existing between different adjacent vehicles. Furthermore, it is noted that the diversity of driver driving styles is also a key ring in influencing vehicle lane change predictions. The method reasonably solves the key factor of driving style affecting lane change decision, and brings the key factor into the analysis process before prediction, thus having important significance for improving safety and efficiency. Based on this, the invention proposes a new deep learning model called the space-time interaction diagram attention network model (EDS-STIGAT) of the embedded driving style. The model firstly adopts a K-means method and SVO theory to analyze driving styles of different vehicles, and distributes different influence weights for the vehicles according to the different driving styles. Secondly, the invention captures the spatial dependence of different vehicles by adopting GAT, which is an advanced graph neural network, wherein nodes can adaptively allocate different importance to neighbors according to an attention mechanism. The present invention then uses the LSTM based attention mechanism to model the time-series dependency between the sequence outputs from the spatial modules, enabling the model to adaptively extract the time-series dependency characteristics of the vehicle for prediction. The space-time dependence of the vehicle characteristics is captured through the model, so that the accuracy of vehicle lane change prediction is greatly improved.
The invention achieves the aim through the following technical scheme: a vehicle lane change detection method based on a space-time interaction diagram attention network comprises the following steps:
s1: preprocessing data in an original data set;
s2: performing feature extraction on the data processed in the step S1, and analyzing driving styles of different vehicles;
s3: embedding the data analyzed in the step S2 into a space interaction construction module, and extracting the space dependence among vehicles by using GAT;
s4: in the time sequence prediction module, adopting an LSTM (least squares) based on an attention mechanism to capture dynamic time sequence dependency in vehicle characteristic information, combining the space dependency among vehicles in the step S3 with the time sequence dependency information, and predicting lane change intention of surrounding vehicles;
s5: the prediction Precision of the model is embodied by adopting Accuracy (Accuracy), precision (Precision), recall ratio (Recall) and F1 fraction so as to analyze the influence of different model parameters on the result.
The step S1 specifically includes:
s1.1: preprocessing vehicle track data; and carrying out noise reduction pretreatment on the coordinates and the speed of the vehicle through a Savizkg-Golag filtering algorithm.
S1.2: cleaning the track data of the vehicle; including clearing the missing values and deleting the apparent outlier.
S1.3: selecting characteristic parameters; characteristic parameters of vehicle coordinates, speed and collision time are selected for the model.
The step S2 specifically includes the following steps:
s2.1: randomly dividing and gathering vehicle objects; randomly dividing a group of vehicle objects into k clusters, wherein the center of each cluster is represented by the average value of all objects in the corresponding class; where k averages are called centroids and each of the remaining objects are clustered around the nearest centroid.
S2.2: dynamically calculating a clustering center; re-calculating the center of the changed cluster according to the average value of the data points in each cluster, and re-distributing the data points to the nearest cluster after calculating the new centroid; through continuous iteration and reassignment of the clustering center until the assignment is stable, the clustering process is finished, and a clustering result is returned; the final output is a set of clusters with centroids that minimize the error function defined below:
wherein y is 1 ,y 2 ,...,y k For k clusters, μ (y i ) Is cluster y i Centroid of d (x, μ (y) i ) Represents object x and centroid μ (y) i ) Distance between them.
S2.3: selecting a clustering distance calculation method; the distance calculation method includes Euclidean distance and Markov distance. The Euclidean distance is used herein. If x= { x 1 ,x 2 ,...,x n Sum μ= { μ 1 ,μ 2 ,...,μ n The euclidean distance of x to μ is calculated as follows:
the step S3 specifically comprises the following steps:
s3.1: vehicle network topology construction and calculation of inter-vehicle attention coefficientsThe method comprises the steps of carrying out a first treatment on the surface of the First, all vehicles in a certain time period are taken as a node set V, wherein V i Representing the ith vehicle node, since the graph structure input to the GAT layer is a graph structure composed of all the vehicle nodes, the connection relationship of the nodes is different at different time intervals; based on this, it will further
Represented as node v in time slice t i . Then, the characteristics of the node at layer l are expressed as +.>
Where d (l) is expressed as the length of the node's feature in layer l. The attention factor that a node and its neighboring nodes are at time slice t can then be expressed as:
wherein the method comprises the steps of
Is a weight matrix of a shared linear transformation (F is the dimension, F' is the output dimension), and a (·) is a function of the correlation between the compute nodes and the nodes. In this context, a single-layer feedforward neural network is used and trained from the weight vector +.>
Parameterization, which can be expressed as:
wherein · T Representing a transpose operation, || is a join operation, and LeakyReLU is used as an activation function.
S3.2: normalizing the attention coefficient; in order to compare significantly the attention coefficients of different neighboring nodes, the obtained attention coefficients are normalized using a softmax function:
wherein the method comprises the steps of
Representing node v i A set of all neighboring nodes.
S3.3: calculating a new feature vector of the vehicle node; by weighted summation of the attention mechanisms, the nodes are calculated
New feature vector at time slice t:
wherein σ (·) is a nonlinear function. The above equation describes how a single graph attention layer works. In this context, two layers of schematic forces are employed, enhancing the perceived range between nodes. Then, the feature vectors updated by all nodes at the time t of the time slice are considered to be aggregated and used as the output of the GAT module.
S3.4: converting the output vector; to facilitate input to the timing prediction module, the output vector is converted into a feature vector
Whole output->
Representing all spatial features extracted from the input demand sequence by the spatial module can be noted as:
O t+1 ＝[β n |n＝t,t-1,t-2,...,t-L+1] (7)
the step S4 specifically includes the following steps:
s4.1: introducing a time sequence prediction module; in long-term sequence prediction, recurrent neural networks can suffer from gradient vanishing problems. LSTM is an improved recurrent neural network. The method introduces a storage unit and a gating mechanism, solves the problem of gradient disappearance in the traditional recurrent neural network, and can well capture long-term sequence data dependence. Therefore, the LSTM is used as a main framework to construct the time sequence prediction module. And attention mechanisms are introduced into the LSTM to weight the output of the LSTM unit under each time step, so that the effectiveness of the characteristics under each time step is enhanced and the loss of network characteristic information is avoided. There are three gates in the LSTM cell, including a forget gate, an input gate, and an output gate. The forget gate decides how much of the cell state was saved to the current time at the previous time and the input gate decides how much of the current input was saved to the cell state. And an output gate controls how much cell state is transferred to the current LSTM cell output. The specific formula is defined as follows:
i t ＝σ(W ii x t +b ii +W hi h t-1 +b hi ) (8)
f t ＝σ(W if x t +b if +W hf h t-1 +b hf ) (9)
g t ＝tanh(W ig x t +b ig +W hg h t-1 +b hg ) (10)
o t ＝σ(W io x t +b io +W ho h t-1 +b ho ) (11)
wherein f t ,g t ,o t Is input, forget, unit and output threshold value, W ii ,W if ,W ig ,W io ,W hi ,W hf ,W hg ,W ho Respectively, a weight matrix connected to three gates, σ is an activation function, tanh is a hyperbolic cut function,
representing the hadamard product.
S4.2: calculating a driving behavior weight value of the vehicle; constructing an output matrix h= [ H ] by LSTM 1 ,h 2 ,...,h l ]As input information of the Attention layer; wherein h is i The output of the nodes is implied for each time LSTM network. h is a i The concentration of the vehicle behavior is represented by a score function S, and the score is larger, the h is i The greater the contribution weight to the vehicle behavior characterization.
S(h,h i )＝w T tanh(Wh+Uh i +b) (14)
Wherein W, W, U is a weight matrix; b is the offset; tanh is a nonlinear activation function; h may be regarded as a behavior representation vector of one level higher than the vehicle state information.
S4.3: carrying out normalization calculation on the score S of each time step; obtaining a concentration probability distribution matrix A= [ a ] of each input distribution 1 ,a 2 ,...,a l ]Wherein:
the output of the Attention layer at the t time is:
s4.4: calculating the lane change probability of the vehicle; finally, introducing a Softmax function to obtain probabilities of different vehicle lane change behaviors, and taking the maximum value of the probabilities as a vehicle lane change identification result.
The step S5 specifically includes the following steps:
introducing evaluation indexes; accuracy (Accuracy), precision (Precision), recall (Recall), and F1 score were used as evaluation indicators for the experiment. The accuracy represents the proportion of the predicted correct data to all data, the accuracy represents the proportion of the real samples in the predicted result to all the predicted samples, the recall ratio represents the proportion of the model to correctly predict the samples of one class of the whole real samples of the class, and the F1 score is used as the measure of classification problems and is the harmonic average value of the accuracy and the recall ratio, the maximum value is 1, and the minimum value is 0. The parameters are defined as follows:
among them, those predicted to be positive are called True Positives (TP), and those predicted to be negative are called False Negatives (FN). In addition, a negative predicted to be positive is referred to as False Positive (FP), and a negative predicted to be negative is referred to as True Negative (TN).
According to the method, firstly, the vehicle track is preprocessed, and then the degree of different driving styles of surrounding vehicles is quantified by introducing a social value guiding theory and a k-means clustering method. Because different driving styles have different degrees of influence on surrounding vehicles, different influence weights are allocated to the surrounding vehicles, and the surrounding vehicles are embedded into the space interaction construction module. The GAT is used for operating the graph structure data in the space interaction construction module, and the characteristics of each graph node are calculated by focusing on adjacent nodes so as to capture the space dependence of different vehicles in the traffic scene. In the time series prediction module, then, using an attention mechanism based LSTM, the model can be made to capture key parts in the timing characteristics by introducing an attention mechanism. Finally, the Softmax function is used for obtaining the probability of different lane changing intentions of each vehicle, and the maximum value of the probability is used as a result of identifying the lane changing intentions of the vehicle. According to the invention, a new vehicle lane change intention detection method is provided, the embedded vehicle driving style is considered, the space-time dynamics is captured through the attention mechanism, the defect that the space-time dependency in the vehicle characteristic information cannot be dynamically obtained in the traditional vehicle lane change prediction method is overcome, and the prediction precision of the vehicle lane change intention is improved.
The invention has the advantages that:
(1) A vehicle lane change prediction framework combining driving style and attention mechanism is designed for the first time, deep interactive information among vehicles is excavated through a graph attention neural network, time sequence dependence of the vehicles is captured by adopting a long-period memory network based on the attention mechanism, and the problem of vehicle lane change prediction under a complex traffic scene is solved.
(2) By introducing social value orientation theory into the model, social behavior preference of different drivers is analyzed, so that the model can model and reveal interaction process among vehicles more intuitively, and finally, prediction accuracy of the model is effectively improved.
Drawings
FIG. 1 is an overall flow chart of the method of the present invention.
FIG. 2 is a general block diagram of a spatiotemporal interaction diagram attention network model of the method of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the following detailed description of the specific embodiments of the present invention will be provided.
The embodiment of the invention provides a vehicle lane change detection method based on a space-time interaction diagram attention network, wherein the system flow is shown in a figure 1, the model structure is shown in a figure 2, and the method comprises the following steps:
s1: the method comprises the following specific steps of cleaning an original public traffic data set to extract useful data:
s1.1: preprocessing vehicle track data; the german highway data set HighD is chosen here as an example, which contains the motion trajectories of 110500 vehicles observed on a 420-meter highway stretch at 6 different locations. The trajectory of each vehicle is automatically extracted, including vehicle type, size, and operation. Using most advanced computer vision algorithms, the positioning error is typically less than ten centimeters. Because the original data set has more attributes, the required attributes are extracted after column filtering as shown in the following table 1:
TABLE 1
| Numbering device | Name of the name | Annotating |
| 1 | VehicleId | Vehicle numbering |
| 2 | FrameId | Data frame number |
| 3 | LocalX | Acquiring X values of a region coordinate system |
| 4 | LocalY | Acquiring Y values of a region coordinate system |
| 5 | xVelocity | Longitudinal speed of vehicle in coordinate system |
| 6 | yVelocity | Lateral speed of vehicle in coordinate system |
| 7 | xAcceleration | Longitudinal acceleration of a vehicle in a coordinate system |
| 8 | yAcceleration | Lateral acceleration of a vehicle in a coordinate system |
| 9 | TTC | Time of collision of vehicle with surrounding vehicle |
S1.2: cleaning the track data of the vehicle; and after the required column information is extracted, cleaning the vehicle track data by adopting a Savizkg-Golag filtering algorithm, and deleting the missing value. To reduce the effect of noise in the raw data on the experimental results, some of the data in the raw data set will be deleted, including those that change significantly in a short time and those that lack items.
S1.3: selecting characteristic parameters; it is important to select the actual lane keeping data and lane change data, so that some standard behavior data can be selected according to the following method. For the lane keeping data, if the traveling vehicle at the lateral position is not coincident with the lane line, it is judged that the vehicle is traveling along the current lane. For lane change data, if the head of the vehicle passes the lane line and does not return to the current lane within a period of time, the behavior of the vehicle is determined to be lane change.
S2: aiming at the data processed in the step S1, the driving style of the vehicle is firstly analyzed based on a K-means clustering method and a social value guiding theory, and the specific steps are as follows:
s2.1: randomly dividing and gathering vehicle objects; a group of objects is herein randomly divided into k clusters, each cluster center being represented by an average of all objects in the corresponding class. Where k averages are called centroids and each of the remaining objects are clustered around the nearest centroid.
S2.2: dynamically calculating a clustering center; re-calculating the center of the changed cluster according to the average value of the data points in each cluster, and re-distributing the data points to the nearest cluster after calculating the new centroid; through continuous iteration and reassignment of the clustering center until the assignment is stable, the clustering process is finished, and a clustering result is returned; the final output is a set of clusters with centroids that minimize the error function defined below:
wherein y is 1 ,y 2 ,...,y k For k clusters, μ (y i ) Is cluster y i Centroid of d (x, μ (y) i ) Represents object x and centroid μ (y) i ) Distance between them.
S2.3: selecting a clustering distance calculation method; the distance calculation method includes Euclidean distance and Markov distance. The Euclidean distance is used herein. If x= { x 1 ,x 2 ,...,x n Sum μ= { μ 1 ,μ 2 ,...,μ n The euclidean distance of x to μ is calculated as follows:
s3: and (3) distributing different weights for the driving style of the vehicle obtained by the S2 analysis and embedding the weights into a space interaction construction module to obtain the space dependency relationship among different vehicles, wherein the steps are as follows:
s3.1: constructing a vehicle network topological graph and calculating the attention coefficient among vehicles; first, all vehicles in a certain time period are taken as a node set V, wherein V i Representing the ith vehicle node, since the graph structure input to the GAT layer is a graph structure composed of all the vehicle nodes, the connection relationship of the nodes is different at different time intervals; based on this, it will further
Represented as node v in time slice t i . Then, the characteristics of the node at layer l are expressed as +.>
Where d (l) is expressed as the length of the node's feature in layer l. The attention factor that a node and its neighboring nodes are at time slice t can then be expressed as:
wherein the method comprises the steps of
Is a weight matrix of a shared linear transformation (F is the dimension, F' is the output dimension), and a (·) is a function of the correlation between the compute nodes and the nodes. In this context, a single-layer feedforward neural network is used and trained from the weight vector +.>
Parameterization, which can be expressed as:
wherein · T Representing transpose operation, || is join operation, leakyReLU as activation function。
S3.2: normalizing the attention coefficient; in order to compare significantly the attention coefficients of different neighboring nodes, the obtained attention coefficients are normalized using a softmax function:
wherein the method comprises the steps of
Representing node v i A set of all neighboring nodes.
S3.3: calculating a new feature vector of the vehicle node; by weighted summation of the attention mechanisms, the nodes are calculated
New feature vector at time slice t:
wherein σ (·) is a nonlinear function. The above equation describes how a single graph attention layer works. In this context, two layers of schematic forces are employed, enhancing the perceived range between nodes. Then, the feature vectors updated by all nodes at the time t of the time slice are considered to be aggregated and used as the output of the GAT module.
S3.4: converting the output vector; to facilitate input to the timing prediction module, the output vector is converted into a feature vector
Whole output->
Representing all spatial features extracted from the input demand sequence by the spatial module can be noted as:
O t+1 ＝[β n |n＝t,t-1,t-2,...,t-L+1] (7)
s4: the time sequence dependency information of the vehicle is obtained through the time sequence prediction module, and the lane change intention of the vehicle is predicted by combining the space dependency characteristics captured by the S3, and the specific steps are as follows:
s4.1: introducing a time sequence prediction module; the LSTM is used as a main framework to construct a time sequence prediction module. And attention mechanisms are introduced into the LSTM to weight the output of the LSTM unit under each time step, so that the effectiveness of the characteristics under each time step is enhanced and the loss of network characteristic information is avoided. There are three gates in the LSTM cell, including a forget gate, an input gate, and an output gate. The forget gate decides how much of the cell state was saved to the current time at the previous time and the input gate decides how much of the current input was saved to the cell state. And an output gate controls how much cell state is transferred to the current LSTM cell output. The specific formula is defined as follows:
i t ＝σ(W ii x t +b ii +W hi h t-1 +b hi ) (8)
f t ＝σ(W if x t +b if +W hf h t-1 +b hf ) (9)
g t ＝tanh(W ig x t +b ig +W hg h t-1 +b hg ) (10)
o t ＝σ(W io x t +b io +W ho h t-1 +b ho ) (11)
wherein f t ,g t ,o t Is input, forget, unit and output doorValue, W ii ,W if ,W ig ,W io ,W hi ,W hf ,W hg ,W ho Respectively, a weight matrix connected to three gates, σ is an activation function, tanh is a hyperbolic cut function,
representing the hadamard product.
S4.2: calculating a driving behavior weight value of the vehicle; constructing an output matrix h= [ H ] by LSTM 1 ,h 2 ,...,h l ]As input information of the Attention layer; wherein h is i The output of the nodes is implied for each time LSTM network. h is a i The concentration of the vehicle behavior is represented by a score function S, and the score is larger, the h is i The greater the contribution weight to the vehicle behavior characterization.
S(h,h i )＝w T tanh(Wh+Uh i +b) (14)
Wherein W, W, U is a weight matrix; b is the offset; tanh is a nonlinear activation function; h may be regarded as a behavior representation vector of one level higher than the vehicle state information.
S4.3: carrying out normalization calculation on the score S of each time step; obtaining a concentration probability distribution matrix A= [ a ] of each input distribution 1 ,a 2 ,...,a l ]Wherein:
the output of the Attention layer at the t time is:
s4.4: calculating the lane change probability of the vehicle; finally, introducing a Softmax function to obtain probabilities of different vehicle lane change behaviors, and taking the maximum value of the probabilities as a vehicle lane change identification result.
The foregoing is considered as illustrative of the principles of the present invention, and has been described herein before with reference to the accompanying drawings, in which the invention is not limited to the specific embodiments shown.
Claims (6)
1. A vehicle lane change detection method based on a space-time interaction diagram attention network is characterized by comprising the following steps:
s1: preprocessing data in an original data set;
s2: performing feature extraction on the data processed in the step S1, and analyzing driving styles of different vehicles;
s3: embedding the data analyzed in the step S2 into a space interaction construction module, and extracting the space dependence among vehicles by using GAT;
s4: in the time sequence prediction module, adopting an LSTM (least squares) based on an attention mechanism to capture dynamic time sequence dependency in vehicle characteristic information, combining the space dependency among vehicles in the step S3 with the time sequence dependency information, and predicting lane change intention of surrounding vehicles;
s5: the prediction Precision of the model is embodied by adopting Accuracy (Accuracy), precision (Precision), recall ratio (Recall) and F1 fraction so as to analyze the influence of different model parameters on the result.
2. The vehicle lane change detection method based on the spatiotemporal interaction graph attention network as claimed in claim 1, wherein: the step S1 includes the steps of:
s1.1: preprocessing vehicle track data; carrying out noise reduction pretreatment on the coordinates and the speed of the vehicle through a Savizkg-Golag filtering algorithm;
s1.2: cleaning the track data of the vehicle; the method comprises the steps of clearing missing values and deleting obvious abnormal data;
s1.3: selecting characteristic parameters; characteristic parameters of vehicle coordinates, speed and collision time are selected for the model.
3. The vehicle lane change detection method based on the spatiotemporal interaction graph attention network as claimed in claim 1, wherein: step S2 comprises the steps of:
s2.1: randomly dividing and gathering vehicle objects; randomly dividing a group of vehicle objects into k clusters, wherein the center of each cluster is represented by the average value of all objects in the corresponding class; where k averages are called centroids, and each remaining object is clustered around the nearest centroid;
s2.2: dynamically calculating a clustering center; re-calculating the center of the changed cluster according to the average value of the data points in each cluster, and re-distributing the data points to the nearest cluster after calculating the new centroid; through continuous iteration and reassignment of the clustering center until the assignment is stable, the clustering process is finished, and a clustering result is returned; the final output is a set of clusters with centroids that minimize the error function defined below:
wherein y is 1 ,y 2 ,...,y k For k clusters, μ (y i ) Is cluster y i Centroid of d (x, μ (y) i ) Represents object x and centroid μ (y) i ) A distance therebetween;
s2.3: selecting a clustering distance calculation method; the distance calculation method comprises Euclidean distance based, markov distance based and the like; the Euclidean distance is used herein; if x= { x 1 ,x 2 ,...,x n Sum μ= { μ 1 ,μ 2 ,...,μ n The euclidean distance of x to μ is calculated as follows:
4. the vehicle lane change detection method based on the spatiotemporal interaction graph attention network as claimed in claim 1, wherein: the step S3 includes the steps of:
s3.1: vehicle network topology constructionCalculating the attention coefficient among vehicles; first, all vehicles in a certain time period are taken as a node set V, wherein V i Representing the ith vehicle node, since the graph structure input to the GAT layer is a graph structure composed of all the vehicle nodes, the connection relationship of the nodes is different at different time intervals; based on this, it will further
Represented as node v in time slice t i The method comprises the steps of carrying out a first treatment on the surface of the Then, the characteristics of the node at layer l are expressed as +.>
Where d (l) is represented as the length of the node's feature in layer l; the attention factor that a node and its neighboring nodes are at time slice t can then be expressed as:
wherein the method comprises the steps of
Is a weight matrix of a shared linear transformation (F is the dimension, F' is the output dimension), a (·) is a function of the correlation between the compute nodes and the nodes; in this context, a single-layer feedforward neural network is used and trained from the weight vector +.>
Parameterization, which can be expressed as:
wherein · T Representing a transpose operation, || is a join operation, and LeakyReLU is used as an activation function;
s3.2: normalizing the attention coefficient; in order to compare significantly the attention coefficients of different neighboring nodes, the obtained attention coefficients are normalized using a softmax function:
wherein the method comprises the steps of
Representing node v i A set of all neighboring nodes;
s3.3: calculating a new feature vector of the vehicle node; by weighted summation of the attention mechanisms, node v is calculated i New feature vector at time slice t:
wherein σ (·) is a nonlinear function, the above equation describes how the individual graph attention layers work; in the method, two graph annotation force layers are adopted, so that the perception range between nodes is enhanced; then, the feature vectors updated by all nodes at the moment of the time slice t are considered to be aggregated and used as the output of the GAT module;
s3.4: converting the output vector; to facilitate input to the timing prediction module, the output vector is converted into a feature vector
Whole output->
Representing all spatial features extracted from the input demand sequence by the spatial module can be noted as:
O t+1 ＝[β n |n＝t,t-1,t-2,...,t-L+1] (7)
5. the vehicle lane change detection method based on the spatiotemporal interaction graph attention network as claimed in claim 1, wherein: the step S4 includes the steps of:
s4.1: introducing a time sequence prediction module; in long-term sequence prediction, the recurrent neural network can have a gradient vanishing problem; LSTM is an improved recurrent neural network; the method introduces a storage unit and a gating mechanism, solves the problem of gradient disappearance in the traditional recurrent neural network, and can well capture long-term sequence data dependence; therefore, a time sequence prediction module is built by taking LSTM as a main framework; the attention mechanism is introduced into the LSTM to weight the output of the LSTM unit under each time step, so that the effectiveness of the characteristics under each time step is enhanced and the loss of network characteristic information is avoided; three gates are arranged in the LSTM unit, including a forget gate, an input gate and an output gate; the forget gate decides how much of the cell state was saved to the current time at the previous time, the input gate decides how much of the current input was saved to the cell state; and an output gate controls how much cell state is transferred to the current LSTM cell output; the specific formula is defined as follows:
i t ＝σ(W ii x t +b ii +W hi h t-1 +b hi ) (8)
f t ＝σ(W if x t +b if +W hf h t-1 +b hf ) (9)
g t ＝tanh(W ig x t +b ig +W hg h t-1 +b hg ) (10)
o t ＝σ(W io x t +b io +W ho h t-1 +b ho ) (11)
wherein f t ,g t ,o t Is input, forget, unit and output threshold value, W ii ,W if ,W ig ,W io ,W hi ,W hf ,W hg ,W ho Respectively, a weight matrix connected to three gates, σ is an activation function, tanh is a hyperbolic cut function,
represents the hadamard product;
s4.2: calculating a driving behavior weight value of the vehicle; constructing an output matrix h= [ H ] by LSTM 1 ,h 2 ,...,h l ]As input information of the Attention layer; wherein h is i Implicit node output for each time LSTM network; h is a i The attention proportion of the vehicle behavior is represented by a score function S, and the greater the score is, the greater the contribution weight of hi to the vehicle behavior representation is;
S(h,h i )＝w T tanh(Wh+Uh i +b) (14)
wherein W, W, U is a weight matrix; b is the offset; tanh is a nonlinear activation function; h may be regarded as a behavior representation vector of one level higher than the vehicle state information;
s4.3: carrying out normalization calculation on the score S of each time step; obtaining a concentration probability distribution matrix A= [ a ] of each input distribution 1 ,a 2 ,...,a l ]Wherein:
the output of the Attention layer at the t time is:
s4.4: calculating the lane change probability of the vehicle; finally, introducing a Softmax function to obtain probabilities of different vehicle lane change behaviors, and taking the maximum value of the probabilities as a vehicle lane change identification result.
6. The vehicle lane change detection method based on the spatiotemporal interaction graph attention network as claimed in claim 1, wherein: the step S5 includes the steps of:
introducing evaluation indexes; using Accuracy (Accuracy), precision (Precision), recall (Recall) and F1 score as evaluation indexes of the experiment; the accuracy represents the proportion of the predicted correct data to all data, the accuracy represents the proportion of the real samples in the predicted result to all the predicted samples, the recall ratio represents the proportion of the model to the samples of one class of the whole real samples of the class, the F1 fraction is used as the measurement of the classification problem, and is the harmonic average value of the accuracy and the recall ratio, the maximum value is 1, and the minimum value is 0; the parameters are defined as follows:
among them, those predicted to be positive are called True Positives (TP), while those predicted to be negative are called False Negatives (FN); in addition, a negative predicted to be positive is referred to as False Positive (FP), and a negative predicted to be negative is referred to as True Negative (TN).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310219674.XA CN116383685A (en) | 2023-03-09 | 2023-03-09 | Vehicle lane change detection method based on space-time interaction diagram attention network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310219674.XA CN116383685A (en) | 2023-03-09 | 2023-03-09 | Vehicle lane change detection method based on space-time interaction diagram attention network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116383685A true CN116383685A (en) | 2023-07-04 |
Family
ID=86964755
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310219674.XA Pending CN116383685A (en) | 2023-03-09 | 2023-03-09 | Vehicle lane change detection method based on space-time interaction diagram attention network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116383685A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116959260A (en) * | 2023-09-20 | 2023-10-27 | 东南大学 | Multi-vehicle driving behavior prediction method based on graph neural network |
| CN116985793A (en) * | 2023-09-26 | 2023-11-03 | 深圳市交投科技有限公司 | Automatic driving safety control system and method based on deep learning algorithm |
| CN117333847A (en) * | 2023-12-01 | 2024-01-02 | 山东科技大学 | Track prediction method and system based on vehicle behavior recognition |
| CN117709602A (en) * | 2024-02-05 | 2024-03-15 | 吉林大学 | Urban intelligent vehicle personification decision-making method based on social value orientation |
-
2023
- 2023-03-09 CN CN202310219674.XA patent/CN116383685A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116959260A (en) * | 2023-09-20 | 2023-10-27 | 东南大学 | Multi-vehicle driving behavior prediction method based on graph neural network |
| CN116959260B (en) * | 2023-09-20 | 2023-12-05 | 东南大学 | Multi-vehicle driving behavior prediction method based on graph neural network |
| CN116985793A (en) * | 2023-09-26 | 2023-11-03 | 深圳市交投科技有限公司 | Automatic driving safety control system and method based on deep learning algorithm |
| CN116985793B (en) * | 2023-09-26 | 2023-12-22 | 深圳市交投科技有限公司 | Automatic driving safety control system and method based on deep learning algorithm |
| CN117333847A (en) * | 2023-12-01 | 2024-01-02 | 山东科技大学 | Track prediction method and system based on vehicle behavior recognition |
| CN117333847B (en) * | 2023-12-01 | 2024-03-15 | 山东科技大学 | Track prediction method and system based on vehicle behavior recognition |
| CN117709602A (en) * | 2024-02-05 | 2024-03-15 | 吉林大学 | Urban intelligent vehicle personification decision-making method based on social value orientation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN116383685A (en) | Vehicle lane change detection method based on space-time interaction diagram attention network | |
| Ma et al. | Driving style recognition and comparisons among driving tasks based on driver behavior in the online car-hailing industry | |
| CN107958269B (en) | Driving risk degree prediction method based on hidden Markov model | |
| CN109697852B (en) | Urban road congestion degree prediction method based on time sequence traffic events | |
| CN109034264B (en) | CSP-CNN model for predicting severity of traffic accident and modeling method thereof | |
| CN110949398B (en) | Method for detecting abnormal driving behavior of first-vehicle drivers in vehicle formation driving | |
| CN107247956B (en) | Rapid target detection method based on grid judgment | |
| CN109191922B (en) | Large-scale four-dimensional track dynamic prediction method and device | |
| CN113401143B (en) | Individualized self-adaptive trajectory prediction method based on driving style and intention | |
| CN111460919B (en) | Monocular vision road target detection and distance estimation method based on improved YOLOv3 | |
| CN108346293B (en) | Real-time traffic flow short-time prediction method | |
| CN111242015B (en) | Method for predicting driving dangerous scene based on motion profile semantic graph | |
| CN113911129A (en) | Traffic vehicle intention identification method based on driving behavior generation mechanism | |
| CN112270355A (en) | Active safety prediction method based on big data technology and SAE-GRU | |
| CN109034020A (en) | A kind of community's Risk Monitoring and prevention method based on Internet of Things and deep learning | |
| CN112990545B (en) | Traffic safety state prediction method for expressway intersection area | |
| CN112734094B (en) | Intelligent city intelligent rail vehicle fault gene prediction method and system | |
| CN113159403B (en) | Intersection pedestrian track prediction method and device | |
| CN111907523B (en) | Vehicle following optimizing control method based on fuzzy reasoning | |
| CN107688556B (en) | Real-time travel time calculation method based on functional principal component analysis | |
| Xiong et al. | Vehicle driving risk prediction based on Markov chain model | |
| CN113033899A (en) | Unmanned adjacent vehicle track prediction method | |
| CN113484875A (en) | Laser radar point cloud target hierarchical identification method based on mixed Gaussian ordering | |
| CN112927511A (en) | Vehicle early warning method based on driver age and gender identification | |
| CN115186594A (en) | Energy-saving speed optimization method under influence of man-vehicle-road coupling |
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 |