Lane changing intention identification method and system under two scenes of expressway section and urban section

Abstract

The invention discloses a vehicle lane changing intention recognition method and system. The method comprises the following steps that the speed of a recognized vehicle and the speed of vehicles around the recognized vehicle are obtained; if the speeds of the two vehicles are both greater than a preset value, inputting characteristic parameters of the identified vehicle and surrounding vehicles thereof obtained in a real-time scene into a first lane change identification model in a highway section scene, otherwise, inputting the characteristic parameters into a second lane change identification model in an urban road section scene; and identifying the lane changing intention of the vehicle through the first lane changing identification model or the second lane changing identification model and outputting the lane changing intention. According to the invention, identification models under different scenes of the expressway section and the urban section are constructed according to the vehicle speed, and the characteristic parameters of the identified vehicle and the surrounding vehicles in the real-time scene are respectively input into the corresponding models for vehicle lane changing intention identification. The problem that in an existing method, a lane changing intention recognition application scene is single or the scene is not clear can be solved.

Description

Technical Field

[0001] The present invention relates to the field of traffic safety technology, and in particular to a method and system for recognizing lane change intentions in two scenarios: expressway sections and urban sections. Background Art

[0002] Since traffic accidents caused by improper lane-changing operations by drivers account for the majority of traffic accidents, it is necessary to accurately identify the driver's lane-changing intention to reduce the losses caused by such accidents. At present, the methods for identifying vehicle lane-changing intentions are mostly targeted at a single scenario or the application scenario is unclear. In recent years, with the development of deep learning, the identification method of lane-changing intentions has changed from ordinary algorithms to deep learning methods to improve the recognition accuracy. Summary of the invention

[0003] The main purpose of the present invention is to provide a method and system for identifying lane changing intention in two scenarios, namely, expressway sections and urban sections, which can improve the accuracy of identifying lane changing intention.

[0004] The technical solution adopted by the present invention is:

[0005] A method for identifying a vehicle lane change intention is provided, comprising the following steps:

[0006] Obtain the speed of the identified vehicle and the speed of the vehicles around it;

[0007] If both speeds are greater than the preset value, the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario are input into the first lane change recognition model in the highway section scenario; otherwise, the characteristic parameters are input into the second lane change recognition model in the urban section scenario; wherein the characteristic parameters of the identified vehicle include speed, acceleration, horizontal coordinate and vertical coordinate, and the characteristic parameters of the surrounding vehicles include speed and acceleration, as well as the horizontal relative position and vertical relative position with the identified vehicle;

[0008] The lane change intention of the vehicle is recognized and outputted through the first lane change recognition model or the second lane change recognition model.

[0009] Following the above technical solution, the construction of the first lane change recognition model includes the following steps:

[0010] Construct the first lane-changing recognition model, including the GRU recurrent layer, the attention mechanism layer, the fully connected layer, and the output layer; its input is the characteristic parameters of the target vehicle and its surrounding vehicles, and its output is the probability of the target vehicle's three different lane-changing categories;

[0011] A historical vehicle trajectory dataset of a highway section is selected from a public vehicle trajectory dataset and processed to obtain vehicles whose lane numbers have changed once, and vehicles that have stably traveled for a preset time under the lane number after changing left or right are screened out, and straight-moving vehicles whose lane numbers remain unchanged are screened out, and the screened-out vehicles are used as sample vehicles; the IDs of left- and right-changing vehicles and the data of each frame are obtained from the sample vehicles, and the frame in which the lane number of each vehicle changes is taken as the starting point to obtain a certain number of frames of data before the lane change, and the speed, acceleration, horizontal coordinate and vertical coordinate of the surrounding vehicles during this period are obtained accordingly, and the lateral relative position and longitudinal relative position of the surrounding vehicles relative to the sample vehicle in each frame are calculated; at the same time, the corresponding data of the straight-moving vehicle and its surrounding vehicles are obtained; the data of each sample vehicle and its surrounding vehicles are taken as a sample, and are labeled accordingly according to three different lane change categories to obtain a sample set;

[0012] The sample set is divided into a training set and a test set, and the first lane change recognition model is trained and tested to obtain a final first lane change recognition model that meets the target requirements.

[0013] Following the above technical solution, the construction of the second lane change recognition model is the same as the first lane change recognition model, except that the data of the training set and the test set are replaced by the corresponding urban road section data.

[0014] According to the above technical solution, the surrounding vehicles include vehicles in six positions: left front, left rear, directly in front, directly behind, right front, and right rear of the identified vehicle.

[0015] Following the above technical solution, during training, the labels of the samples are as follows: left lane change, straight driving, and right lane change are (1, 0, 0), (0, 1, 0), and (0, 0, 1), respectively. Each sample contains a feature sequence of n time steps, and the feature sequence of the i-th time step is [V i , a i , x i ,y i , V ij , a ij , Δx ij , Δy ij ] T , where V i is the speed of the sample vehicle at the i-th time step, a i is the acceleration of the sample vehicle at the i-th time step, x i is the longitudinal position of the sample vehicle at the i-th time step, y i is the lateral position of the sample vehicle at the i-th time step, V ij is the speed of the jth vehicle around the sample vehicle at the i-th time step, a ij is the acceleration of the jth vehicle around the sample vehicle at the i-th time step, Δx ijis the longitudinal relative position of the jth vehicle at the i-th time step around the sample vehicle, Δy ij is the lateral relative position of the jth vehicle around the sample vehicle at the i-th time step, i = 1, 2…150, j = 1, 2, 3, 4, 5, 6, n = 150. The labeled samples are randomly divided into batches and enter the GRU recurrent layer for processing. After each time step is processed, a hidden state containing important vehicle features of the current time step will be output. After each sample is processed, a sequence of n hidden states will be output.

[0016] Following the above technical solution, the hidden state sequence of n time steps output by the GRU recurrent layer of each sample is used as the input of the attention mechanism layer. The attention mechanism layer assigns different weights to the hidden state of each sample at each time step, and the final hidden state step is obtained by weighting.

[0017] Following the above technical solution, the input of the fully connected layer is the final hidden state of the sample. By performing a linear transformation on the final hidden state, the probability estimation of left lane change, straight driving, and right lane change for each sample is output.

[0018] Following the above technical solution, the output layer calculates and normalizes the probability estimates of left lane change, straight driving, and right lane change through an activation function so that the sum of the three is equal to 1, and finally outputs the probabilities of three different lane change categories.

[0019] The present invention also provides a vehicle lane change intention recognition system, comprising the following steps:

[0020] An acquisition module, used to acquire the speed of the identified vehicle and the speeds of the vehicles around it;

[0021] The intention recognition module inputs the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario into the first lane change recognition model in the highway section scenario if both speeds are greater than the preset value; otherwise, the characteristic parameters are input into the second lane change recognition model in the urban section scenario; wherein the characteristic parameters of the identified vehicle include speed, acceleration, abscissa and ordinate, and the characteristic parameters of the surrounding vehicles include speed and acceleration, as well as the lateral relative position and longitudinal relative position to the identified vehicle;

[0022] The output module is used to identify the lane-changing intention of the vehicle through the first lane-changing recognition model or the second lane-changing recognition model and output it.

[0023] The present invention also provides a computer storage medium, which stores a computer program that can be executed by a processor, and the computer program executes the vehicle lane change intention recognition method described in the above technical solution.

[0024] The beneficial effects of the present invention are as follows: the present invention corresponds different scenarios to different lane change intention recognition models, divides different scenarios of expressway sections and urban sections according to vehicle speed, inputs characteristic parameters of the identified vehicle and its surrounding vehicles obtained in real-time scenarios into corresponding models respectively, and recognizes the vehicle lane change intentions in two accident-prone scenarios of expressway sections and urban sections, so as to solve the problem of single application scenario or unclear scenario of lane change intention recognition in existing methods.

[0025] Furthermore, compared with traditional methods, the introduction of the GRU recurrent layer can effectively extract the time dependency of the lane-changing vehicle feature parameters. The temporal attention mechanism is introduced after the GRU layer. This mechanism actively screens out key vehicle information by assigning different weight coefficients to the vehicle lane-changing features at different time steps, thereby improving the recognition performance of the model and helping to improve driving safety.

[0026] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0028] Figure 1 is a flow chart of a method for identifying a vehicle lane-changing intention according to an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of the structure of a vehicle lane change intention recognition system according to an embodiment of the present invention;

[0030] Figure 3 is a schematic structural diagram of a vehicle lane change intention recognition system according to another embodiment of the present invention; DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] It should be noted that the illustrations provided in the embodiments of the present invention are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0033] In the present invention, it is also necessary to explain that, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, if the terms "first" and "second" appear, they are only used for description and distinction purposes, and cannot be understood as indicating or implying relative importance.

[0034] like Figure 1 As shown, the vehicle lane change intention recognition method according to the embodiment of the present invention comprises the following steps:

[0035] S1. Obtain the speed of the identified vehicle and the speeds of its surrounding vehicles;

[0036] S2, determining whether the speeds of both are greater than a preset value;

[0037] S3, if both speeds are greater than a preset value, input the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario into the first lane change recognition model in the highway section scenario, otherwise, execute step S4;

[0038] S4, inputting characteristic parameters into a second lane change recognition model in an urban road scene; wherein the characteristic parameters of the identified vehicle include speed, acceleration, abscissa and ordinate, and the characteristic parameters of surrounding vehicles include speed and acceleration, as well as a lateral relative position and a longitudinal relative position to the identified vehicle;

[0039] S5. Identify the lane-changing intention of the vehicle through the first lane-changing recognition model or the second lane-changing recognition model and output it.

[0040] In the embodiment, the system can determine which scenario the lane change intention recognition model runs according to the average speed of the target vehicle and surrounding vehicles in the selected time window. If the average speed is ≥ 60km / h, the system runs the lane change intention recognition model in the highway scene, otherwise, the system runs the lane change intention recognition model in the urban scene. The judgment is based on the traffic regulations that the speed of urban roads is usually 30-60km / h, and the speed of highways is usually 60-120km / h.

[0041] It can be seen that this embodiment corresponds different scenarios to different lane change intention recognition models, divides different scenarios of highway sections and urban sections according to vehicle speed, inputs the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in real-time scenarios into corresponding models, and recognizes the vehicle's lane change intention in two accident-prone scenarios of highway sections and urban sections, so as to solve the problem of single application scenario or unclear scenario of lane change intention recognition in existing methods.

[0042] Furthermore, the construction of the first lane change recognition model includes the following steps:

[0043] Construct the first lane-changing recognition model, including the GRU recurrent layer, the attention mechanism layer, the fully connected layer, and the output layer; its input is the characteristic parameters of the target vehicle and its surrounding vehicles, and its output is the probability of the target vehicle's three different lane-changing categories;

[0044] A historical vehicle trajectory dataset of a highway section is selected from a public vehicle trajectory dataset and processed to obtain vehicles whose lane numbers have changed once, and vehicles that have stably traveled for a preset time under the lane number after changing left or right are screened out, and straight-moving vehicles whose lane numbers remain unchanged are screened out, and the screened-out vehicles are used as sample vehicles; the IDs of left- and right-changing vehicles and the data of each frame are obtained from the sample vehicles, and the frame in which the lane number of each vehicle changes is taken as the starting point to obtain a certain number of frames of data before the lane change, and the speed, acceleration, horizontal coordinate and vertical coordinate of the surrounding vehicles during this period are obtained accordingly, and the lateral relative position and longitudinal relative position of the surrounding vehicles relative to the sample vehicle in each frame are calculated; at the same time, the corresponding data of the straight-moving vehicle and its surrounding vehicles are obtained; the data of each sample vehicle and its surrounding vehicles are taken as a sample, and are labeled accordingly according to three different lane change categories to obtain a sample set;

[0045] The sample set is divided into a training set and a test set, and the first lane change recognition model is trained and tested to obtain a final first lane change recognition model that meets the target requirements.

[0046] The construction of the second lane change recognition model is the same as that of the first lane change recognition model, except that the data of the training set and the test set are replaced by the corresponding urban road section data.

[0047] Among them, the surrounding vehicles include vehicles in six positions: left front, left rear, directly in front, directly behind, right front, and right rear of the identified vehicle.

[0048] During training, the labels of the samples are: left lane change, straight driving, and right lane change are (1, 0, 0), (0, 1, 0), and (0, 0, 1), respectively. Each sample contains a feature sequence of n time steps, and the feature sequence of the i-th time step is [V i , a i , xi ,y i , V ij , a ij , Δx ij , Δy ij ] T , where V i is the speed of the sample vehicle at the i-th time step, a i is the acceleration of the sample vehicle at the i-th time step, x i is the longitudinal position of the sample vehicle at the i-th time step, y i is the lateral position of the sample vehicle at the i-th time step, V ij is the speed of the jth vehicle around the sample vehicle at the i-th time step, a ij is the acceleration of the jth vehicle around the sample vehicle at the i-th time step, Δx ij is the longitudinal relative position of the jth vehicle at the i-th time step around the sample vehicle, Δy ij is the lateral relative position of the jth vehicle around the sample vehicle at the i-th time step, i = 1, 2…150, j = 1, 2, 3, 4, 5, 6, n = 150. The labeled samples are randomly divided into batches and enter the GRU recurrent layer for processing. After each time step is processed, a hidden state containing important vehicle features of the current time step will be output. After each sample is processed, a sequence of n hidden states will be output.

[0049] Specifically, the hidden state sequence of n time steps output by the GRU recurrent layer of each sample is used as the input of the attention mechanism layer. The attention mechanism layer assigns different weights to the hidden state of each sample at each time step, and the final hidden state is obtained by weighting.

[0050] The input of the fully connected layer is the final hidden state of the sample. By performing a linear transformation on the final hidden state, the probability estimate of changing lanes left, going straight, and changing lanes right for each sample is output.

[0051] The output layer calculates and normalizes the probability estimates of left lane change, straight driving, and right lane change through an activation function so that the sum of the three is equal to 1, and finally outputs the probabilities of three different lane change categories.

[0052] Compared with the traditional method, the lane changing intention recognition model for the two scenarios established above can effectively extract the time dependence of the lane changing vehicle feature parameters by introducing the GRU recurrent layer. The temporal attention mechanism layer is introduced after the GRU recurrent layer. Through the temporal attention mechanism, different weight coefficients are assigned to the hidden states containing the vehicle lane changing features at different time steps output by the GRU recurrent layer, thereby actively screening out the hidden states containing the vehicle features at the time steps that have a greater impact on lane changing recognition, thereby improving the recognition performance of the model and helping to improve driving safety.

[0053] Furthermore, the present invention can achieve diversified recognition scenarios. Given that the overall framework and methods used in the lane change intention recognition methods for highway sections and urban sections are the same, they are only trained with different training samples. The recognition models for the two scenarios are trained with the same steps.

[0054] First, sequence processing is performed, specifically inputting the vehicle behavior sequence (including the feature data of the vehicle and surrounding vehicles) into the GRU network. When the GRU network processes the sequence, each time step will output a hidden state (including the important vehicle features of the current time step). This hidden state will be output by the GRU layer and passed to the next time step, while combining the input data of the current time step.

[0055] Secondly, the final hidden state is obtained. Specifically, after processing all input sequences of each sample, the attention mechanism calculates the weight of the hidden state output by the GRU layer of each time step of the sample, and weights the above hidden states according to the weights to obtain the final hidden state of the sample.

[0056] Third, a fully connected layer is constructed, specifically passing the final hidden state to the fully connected layer. The fully connected layer performs further nonlinear transformation on the data.

[0057] Fourth, output is performed through the output layer. There is usually an output layer that uses a softmax activation function to generate the final classification probability. The category with a high probability is the lane changing behavior category output by the overall recognition method.

[0058] Finally, during the training process, the loss function is used for optimization. Specifically, the loss function (cross entropy loss) is used to measure the difference between the predicted value and the true value, and the back-propagation algorithm is used to update the parameters of the GRU layer, attention mechanism, and fully connected layer.

[0059] Specifically, each of the above processes is described in detail as follows.

[0060] 1. Sequence processing

[0061] 1.1 Data acquisition and processing

[0062] The dataset used for training comes from the public dataset Citysim. The sampling frequency of the Citysim dataset is 30hz. The data of the FreewayB section in the dataset is selected as the data for high-speed section training. The data of the ExpresswayA section is used as the data for urban section training.

[0063] The redundant data is removed from the acquired data set, specifically, the vehicle data of non-car types is deleted, and the characteristic data such as vehicle ID, speed, acceleration, horizontal coordinate, vertical coordinate, sampling time and lane number are retained.

[0064] The lane-changing vehicles on expressways and urban roads and their corresponding feature data are screened separately. Specifically, the vehicle IDs whose lane numbers change once are obtained from the data preliminarily processed above. To ensure that the data obtained are vehicle data for successful lane changes, the frame where the lane number corresponding to the vehicle changes is taken as the starting point. When the vehicle drives stably for 2S or more, i.e., 60 frames, under the changed lane number after the lane change, the vehicle can be used as the identified vehicle for lane change intention recognition training. According to the changes in the lane numbers of the vehicles that have undergone a lane change behavior mentioned above and the diagrams in the data set, these lane-changing vehicles are divided into two categories: left lane change and right lane change. Thus, the lane-changing vehicles on expressways and urban roads and their corresponding feature data are obtained. At the same time, the vehicles whose lane numbers remain unchanged in the two data sets and their corresponding feature data are obtained as the vehicle data for the straight-ahead category in the two scenarios.

[0065] Obtain the six vehicles around the lane-changing vehicle on the expressway and urban road sections and their corresponding characteristic parameters. Specifically, the left and right lane-changing vehicle IDs and each frame of data in the two selected scenarios are used. The frame in which the lane number corresponding to each vehicle changes is used as the starting point to obtain the data of these vehicles for the time period before the vehicle changes lanes, that is, 150 frames. The data of each lane-changing vehicle are filtered from the two data sets according to the corresponding frame number and lane number to obtain the speed, acceleration, horizontal coordinate, and vertical coordinate of the six surrounding vehicles within 5S before the identified vehicle crosses the lane line. And the lateral relative position and longitudinal relative position of these vehicles in each frame relative to the identified vehicle are calculated by formulas (1-1) and (1-2). For the identified vehicles of the straight-moving category in the two scenarios, the speed, acceleration, horizontal coordinate, vertical coordinate, and the speed, acceleration, lateral relative position, and longitudinal relative position of the six vehicles around the identified vehicle are also obtained for 5S and 150 frames on their trajectories.

[0066] where y ij is the lateral position of the jth vehicle in the i-th frame around the identified vehicle, x ij is the longitudinal position of the jth vehicle in the i-th frame around the identified vehicle, y is is the lateral position of the identified vehicle in the i-th frame, x is is the longitudinal position of the identified vehicle in the i-th frame, Δy ij The relative lateral position between the jth vehicle and the identified vehicle around the i-th frame, Δx ij The relative longitudinal position of the jth vehicle around the i-th frame and the identified vehicle, i=1, 2…150, j=1, 2, 3, 4, 5, 6.

[0067] y ij -y is =Δy ij (1-1)

[0068] x ij -x is =Δx ij (1-2)

[0069] The time window length is selected as 5S before the three types of identified vehicles in the above two scenes cross the lane line, and the selected time window contains 150 time steps. The speed, acceleration, lateral position, longitudinal position of the three types of identified vehicles (left lane change, straight driving, right lane change) in the above two scene data sets within 5S, and the speed, acceleration, lateral relative position, and longitudinal relative position of the six vehicles around the identified vehicle are used as feature parameters.

[0070] The above features (such as speed, acceleration, position, etc.) have different dimensions and value ranges. Min-Max normalization is used for the above features to eliminate these differences for subsequent modeling. The specific calculation method takes speed V as an example, and the normalization method of other features is the same.

[0071]

[0072] Where min(V) and max(V) are the minimum and maximum values ​​of the speed characteristic sequence in the above time window of the vehicle, respectively.

[0073] The speed, acceleration, lateral position, longitudinal position of each frame of the identified vehicle, the speed, acceleration, lateral relative position and longitudinal relative position of each frame of the surrounding vehicles are normalized using the above method, and the processed values ​​are used to replace the original data to obtain a normalized feature sequence.

[0074] 1.2 GRU layer processing

[0075] The before-and-after correlation of vehicle lane-changing behavior is strong. Compared with traditional models, the network model with memory function in the deep learning model is more suitable for lane-changing intention recognition modeling. The GRU (Gated Recurrent Unit) model is a network model with memory function. It can effectively extract the spatial correlation and temporal dependency between vehicles, and has higher recognition accuracy. The GRU model is a variant of the LSTM (Long Short-Term Memory Network). In comparison, the GRU model has higher computational efficiency. The training speed of the GRU model is usually faster than that of the LSTM model, especially when processing large-scale data sets. This difference is more obvious. Compared with the standard RNN (recurrent neural network), GRU alleviates the gradient vanishing problem and effectively avoids the gradient vanishing in long-term sequence training. Therefore, the GRU model was selected in comparison.

[0076] The feature parameters of the identified vehicle and its corresponding surrounding vehicles extracted in the two scenes within the above 5S time window are respectively processed by the GRU circulation layers established in the two scenes.

[0077] In each scenario, the characteristic parameters of each identified vehicle and its corresponding surrounding vehicles within a 5S time window are recorded as an identified sample. Each sample is labeled according to the three categories of data processing before entering the GRU layer. Left lane change, straight driving, and right lane change are (1, 0, 0), (0, 1, 0), and (0, 0, 1), respectively. Each sample contains a feature sequence of 150 time steps. The feature sequence of the i-th time step is [V i , a i , x i ,y i , V ij , a ij , Δx ij , Δy ij ] T , where V i is the speed of the identified vehicle at the i-th time step, a i is the acceleration of the identified vehicle at the i-th time step, x i is the longitudinal position of the identified vehicle at the i-th time step, y i is the lateral position of the identified vehicle at the i-th time step, V ij is the speed of the jth vehicle around the identified vehicle at the i-th time step, a ij is the acceleration of the jth vehicle around the identified vehicle at the i-th time step, Δx ij is the longitudinal relative position of the jth vehicle around the identified vehicle at the i-th time step, Δy ij is the lateral relative position of the jth vehicle around the identified vehicle at the i-th time step, i = 1, 2…150, j = 1, 2, 3, 4, 5, 6. The labeled samples are randomly divided into batches and enter the GRU recurrent layer for processing.

[0078] The recurrent layer in each scenario contains two GRU units with the same parameters. The working process is consistent. The core of GRU (Gated Recurrent Unit) is the two built-in gates, namely the reset gate and the update gate. Formula 1-3 is the reset gate calculation formula, Formula 1-4 is the update gate calculation formula, Formula 1-5 is the candidate hidden state calculation formula, and Formula 1-6 is the hidden state calculation formula. The reset gate determines how to combine the new input information with the previous memory, and the update gate defines the amount of previous memory saved to the current time step. The new hidden state generated at the current moment is the result obtained after processing by the GRU unit, which is determined by Formula 1-6. The final output hidden state contains the feature sequence of the vehicle lane changing behavior input at the current time step and the important feature parameters of the vehicle lane changing behavior of the hidden state of the previous time step. It is the key to the GRU recurrent layer to realize the memory function. In Formulas 1-3 to 1-5, W i and U i (i takes r, z, h) is the weight matrix in the formula, b i (i takes r, z, h) is the bias vector in the formula. i and U i (i takes r, z, h) and initializes the matrix by randomly selecting values ​​from a uniform distribution. i (i takes r, z, h) as the bias vector in the formula, which is initialized to zero vector. i and U i (i takes r, z, h) is initialized to a 28×28 dimensional matrix, h t-1 is a 28-dimensional vector, b i (i takes r, z, h) is initialized to a 28-dimensional zero vector, and the hidden state of the GRU recurrent layer at the first time step of the sample input is a zero vector.

[0079] r t =σ(W r x t +U r h t-1 +b r ) (Formula 1-3)

[0080] Z t =σ(W z x t +U z h t-1 +b z ) (Formula 1-4)

[0081] H t =tanh(W h x t +(r t ⊙h t-1 )W h +bh )(Formula 1-5)

[0082] h t =Z t ⊙h t-1 +(1-Z t )⊙H t (Formula 1-6)

[0083] The feature sequences of 150 time steps of each sample in each batch are processed as above by the two GRU units of the recurrent layer. After processing, each time step will output a hidden state containing the important vehicle features of the current time step. After each sample is processed, 150 hidden states will be output, namely {h1, h2, ..., h150}.

[0084] 2. Working process of attention mechanism

[0085] Each sample in each batch outputs a hidden state sequence of 150 time steps through the GRU recurrent layer, which contains important feature parameters at each time step during the vehicle lane changing process. That is, {h1, h2, ..., h150} is used as the input of the attention mechanism. For each sample, the attention mechanism gives different weights to the hidden state sequence of each sample at each time step, so that the model pays more attention to the time feature sequence that has a greater impact on the lane changing behavior and actively queries important information. The weighting of the hidden state sequence at each time step is mainly calculated by calculating the attention score of the hidden state at each time step, and the attention score is normalized by the softmax function to obtain the attention weight of the hidden state at that time step. The calculation method of the attention score ei of the output hidden state of each sample at each time step is illustrated by taking the hidden state input attention mechanism of the i-th time step as an example.

[0086] The attention score of the hidden state of the i-th time step containing important feature parameters in the vehicle lane changing process is as follows:

[0087] ei=q T Whi

[0088] Among them, q is the query vector, W is the weight matrix, and hi is the hidden state containing the characteristics of the vehicle lane-changing behavior in the attention mechanism at the i-th time step of each sample. q is initialized to a 28-dimensional vector using a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. W is initialized by randomly selecting elements in the matrix from a uniform distribution. W is a 28×28 matrix.

[0089] Repeat the above steps to calculate the attention scores of the hidden states of other time steps in each sample time window.

[0090] The calculation method of the attention weight αi of the output hidden state of each time step is illustrated by taking the hidden state input attention mechanism of the i-th time step as an example.

[0091]

[0092] Among them, ei is the hidden state attention score of each sample at the i-th time step, and exp(ei) is e ei , Σexp(ei) is the sum of the exp(ei) of the hidden state attention score ei at each time step of 150 time steps for each sample.

[0093] Repeat the above steps to calculate the attention weights of the hidden states of other time steps in each sample time window.

[0094] Use the hidden state attention weight αi (i is 1, 2…150) of each time step to perform weighted summation on the corresponding hidden state sequence to obtain the weighted hidden state h * .

[0095]

[0096] Among them, α i is the hidden state attention weight of the i-th time step, h i is the hidden state at the i-th time step.

[0097] The attention mechanism is based on the weighted hidden state h * As output, h * Contains the integrated information of the important time steps of the time series during the lane changing process of the current vehicle of each sample.

[0098] Repeat the above step of weighting the hidden states of a sample at different time steps to obtain the final hidden state. Weight the hidden states of each batch of samples at different time steps input for training in the two scenarios to obtain the final hidden state.

[0099] 3. Fully connected layer

[0100] The final hidden state h of a batch of samples trained with the above input in the two scenarios is obtained * Then they are input into their respective fully connected layers for processing. * Contains the feature integration information of important time steps in the time series during the current vehicle lane changing process.

[0101] For each sample, the fully connected layer first uses the following method to weight the hidden state h * Perform a linear transformation.

[0102] z=W * h* +b *

[0103] Where z = (z1, z2, z3) T is the output vector of the fully connected layer. In the lane change intention recognition problem, there are three possible intentions (for example: left lane change, straight driving, right lane change). In this case, the output z of the final fully connected layer will be a vector of length 3. z1, z2, z3 are the combination of weights and biases learned by the fully connected layer, which can be understood as the probability estimation of the lane change model for different lane change categories. * Initialized to a 3D zero vector, W * The matrix is ​​initialized by randomly selecting values ​​from uniform distribution. W is a 3×28 matrix.

[0104] z=(z1,z2,z3) T The z1, z2, and z3 in it are the estimated probabilities of changing lanes left, going straight, and changing lanes right for each sample, respectively.

[0105] Repeat the above fully connected layer processing steps for one sample to perform fully connected layer processing on each batch of samples input for training in the two scenarios to obtain the probability estimates of left lane change, straight driving, and right lane change for each sample.

[0106] 4. Output Layer

[0107] The Softmax function is an activation function commonly used in multi-classification problems, which converts a real number vector into a probability distribution.

[0108] Since z1, z2, and z3 in the z vector obtained by the fully connected layer are the probability estimates of changing lanes to the left, going straight, and changing lanes to the right, respectively, the above calculation may result in an element value greater than 1, which does not meet the definition of probability less than 1. Therefore, the output layer uses the Softmax function to calculate the exponent of each element, and then normalizes these values ​​so that their sum is equal to 1, and each element is between 0 and 1, which can be interpreted as the probability of the corresponding category.

[0109] After obtaining the vector z of the probability estimate of each category of each sample of the above input training in the two scenarios, each of them is input into the softmax function for processing. The processing steps for each sample are as follows.

[0110] The output vector is processed by the sofmax function to output the probability distribution of each lane change category. The output of the softmax function is defined as a vector a = (a1, a2, a3), where a1, a2, and a3 are the probabilities of left lane change, straight driving, and right lane change, respectively. Taking the left lane change probability calculation as an example, it is shown in the following formula.

[0111]

[0112] Repeat the above steps to calculate the probability of each lane-changing behavior for each sample.

[0113] Repeat the above softmax function processing steps for one sample to perform softmax function processing on each batch of samples input for training in the two scenarios to obtain the probability estimates of left lane change, straight driving, and right lane change for each sample.

[0114] Note: Considering that this method is used to identify the lane-changing intention of the vehicle before actually starting to change lanes, it is more rigorous to calculate the probability of each type of intention to obtain the probability of each type of lane change. In model training, the category with the largest probability of the three categories is used as the category recognition result obtained during training. For example, the probability of changing lanes to the left is the largest, which is considered to be the intention of changing lanes to the left. In this way, this result is compared with the label result, and the parameters are optimized according to the deviation, so that the prediction of this method in subsequent use is closer to the accurate result.

[0115] 5. Loss function optimization

[0116] For each sample of a batch of samples input for training in the two scenarios, the cross entropy loss function is used to calculate the difference between the predicted probability and the true label. Among them, yi is the i-th element of the true label, and yi' is the i-th category probability predicted by the model.

[0117]

[0118] After each batch of samples is processed, the loss function values ​​of the two scenarios are calculated. The losses of all samples in the same batch in each scenario are summed up and then divided by the batch size (i.e. the number of samples) to get the average loss. When the average loss of a batch of samples is less than the set threshold of 0.0001, the entire method model stops iterative optimization.

[0119] In parameter optimization, after a batch of samples is processed, the chain rule is used to calculate the gradient of the loss function to the weight matrix and bias vector of the GRU layer using the back propagation algorithm, the gradient of the loss function to the weight matrix and query vector of the attention mechanism, and the gradient of the loss function to the weight matrix and bias vector of the fully connected layer, so as to update the parameters.

[0120]

[0121] Where Wg and bg are the weight matrix and bias vector of the GRU layer, Wa and q are the weight matrix and query vector of the attention mechanism, W and b are the weight matrix and bias vector of the fully connected layer, and η is the learning rate, which is 0.01.

[0122] The present invention also provides an embodiment of a vehicle lane change intention recognition system, which is mainly used to implement the above method embodiment, such as Figure 2 As shown, the system includes:

[0123] An acquisition module, used to acquire the speed of the identified vehicle and the speeds of the vehicles around it;

[0124] The intention recognition module inputs the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario into the first lane change recognition model in the highway section scenario if both speeds are greater than the preset value; otherwise, the characteristic parameters are input into the second lane change recognition model in the urban section scenario; wherein the characteristic parameters of the identified vehicle include speed, acceleration, abscissa and ordinate, and the characteristic parameters of the surrounding vehicles include speed and acceleration, as well as the lateral relative position and longitudinal relative position to the identified vehicle;

[0125] The output module is used to identify the lane-changing intention of the vehicle through the first lane-changing recognition model or the second lane-changing recognition model and output it.

[0126] like Figure 3 As shown, the lane change intention recognition system of another embodiment of the present invention includes a memory, a processor, an input port and an output port. The lane change intention recognition method program obtained by the above two scene training is stored in the memory of the system. The lane change intention recognition program is executed by the processor in the system. The input port inputs vehicle characteristic parameters, including characteristic parameters of the identified vehicle, i.e., the self-vehicle and surrounding vehicles, and is connected to the output port of the vehicle-mounted speed sensor, acceleration sensor, GPS, and laser radar. Among them, the vehicle-mounted speed sensor and acceleration sensor are used to obtain the speed and acceleration of the identified vehicle, i.e., the self-vehicle, respectively. The GPS is used to obtain the horizontal and vertical position information of the identified vehicle, i.e., the self-vehicle, and the laser radar is used to obtain the speed, acceleration and position information of the identified vehicle, i.e., the vehicles around the self-vehicle relative to the identified vehicle. The output port outputs the execution command of the processor and is connected to the actuator in the assisted driving system. The system can be used as a lane change intention recognition module in the assisted driving system to help the assisted driving system accurately identify whether the self-vehicle changes lanes according to the current driving state of the self-vehicle and the driving state of the surrounding vehicles, so as to ensure driving safety.

[0127] The present application also provides a computer-readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, a server, an App application store, etc., on which a computer program is stored, and when the program is executed by a processor, a corresponding function is realized. When the computer-readable storage medium of this embodiment is executed by a processor, the vehicle lane change intention recognition method of the method embodiment is realized.

[0128] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0129] The order of execution of each step in the above embodiment does not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0130] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for identifying a vehicle's lane-changing intention, characterized in that: The following steps are involved: Obtain the speed of the identified vehicle and the speed of the vehicles around it; If both speeds are greater than the preset value, the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario are input into the first lane change recognition model in the highway section scenario; otherwise, the characteristic parameters are input into the second lane change recognition model in the urban section scenario; wherein the characteristic parameters of the identified vehicle include speed, acceleration, horizontal coordinate and vertical coordinate, and the characteristic parameters of the surrounding vehicles include speed and acceleration, as well as the horizontal relative position and vertical relative position with the identified vehicle; The lane change intention of the vehicle is recognized and outputted through the first lane change recognition model or the second lane change recognition model.

2. The method for recognizing vehicle lane change intention according to claim 1, characterized in that: The construction of the first lane change recognition model includes the following steps: Construct the first lane-changing recognition model, including the GRU recurrent layer, the attention mechanism layer, the fully connected layer, and the output layer; its input is the characteristic parameters of the target vehicle and its surrounding vehicles, and its output is the probability of the target vehicle's three different lane-changing categories; A historical vehicle trajectory dataset of a highway section is selected from a public vehicle trajectory dataset and processed to obtain vehicles whose lane numbers have changed once, and vehicles that have stably traveled for a preset time under the lane number after changing left or right are screened out, and straight-moving vehicles whose lane numbers remain unchanged are screened out, and the screened-out vehicles are used as sample vehicles; the IDs of left- and right-changing vehicles and the data of each frame are obtained from the sample vehicles, and the frame in which the lane number of each vehicle changes is taken as the starting point to obtain a certain number of frames of data before the lane change, and the speed, acceleration, horizontal coordinate and vertical coordinate of the surrounding vehicles during this period are obtained accordingly, and the lateral relative position and longitudinal relative position of the surrounding vehicles relative to the sample vehicle in each frame are calculated; at the same time, the corresponding data of the straight-moving vehicle and its surrounding vehicles are obtained; the data of each sample vehicle and its surrounding vehicles are taken as a sample, and are labeled accordingly according to three different lane change categories to obtain a sample set; The sample set is divided into a training set and a test set, and the first lane change recognition model is trained and tested to obtain a final first lane change recognition model that meets the target requirements.

3. The method for recognizing vehicle lane change intention according to claim 2, characterized in that: The construction of the second lane change recognition model is the same as that of the first lane change recognition model, except that the data of the training set and the test set are replaced by the corresponding urban road section data.

4. The method for recognizing vehicle lane-changing intention according to claim 1, characterized in that: The surrounding vehicles include vehicles in six positions: left front, left rear, directly in front, directly behind, right front, and right rear of the identified vehicle.

5. The method for recognizing vehicle lane-changing intention according to claim 2, characterized in that: During training, the labels of the samples are: left lane change, straight driving, and right lane change are (1, 0, 0), (0, 1, 0), and (0, 0, 1), respectively. Each sample contains a feature sequence of n time steps, and the feature sequence of the i-th time step is [V i , a i , x i ,y i , V ij , a ij , Δx ij , Δy ij ] T , where V i is the speed of the sample vehicle at the i-th time step, a i is the acceleration of the sample vehicle at the i-th time step, x i is the longitudinal position of the sample vehicle at the i-th time step, y i is the lateral position of the sample vehicle at the i-th time step, V ij is the speed of the jth vehicle around the sample vehicle at the i-th time step, a ij is the acceleration of the jth vehicle around the sample vehicle at the i-th time step, Δx ij is the longitudinal relative position of the jth vehicle at the i-th time step around the sample vehicle, Δy ij is the lateral relative position of the jth vehicle around the sample vehicle at the i-th time step, i = 1, 2…150, j = 1, 2, 3, 4, 5, 6, n = 150. The labeled samples are randomly divided into batches and enter the GRU recurrent layer for processing. After each time step is processed, a hidden state containing important vehicle features of the current time step will be output. After each sample is processed, a sequence of n hidden states will be output.

6. The method for recognizing vehicle lane-changing intention according to claim 5, characterized in that: The hidden state sequence of n time steps output by the GRU recurrent layer of each sample is used as the input of the attention mechanism layer. The attention mechanism layer assigns different weights to the hidden state of each sample at each time step, and the final hidden state is obtained by weighting.

7. The method for recognizing vehicle lane-changing intention according to claim 6, characterized in that: The input of the fully connected layer is the final hidden state of the sample. By performing a linear transformation on the final hidden state, the probability estimate of changing lanes left, going straight, and changing lanes right for each sample is output.

8. The method for recognizing vehicle lane-changing intention according to claim 7, characterized in that: The output layer calculates and normalizes the probability estimates of left lane change, straight driving, and right lane change through an activation function so that the sum of the three is equal to 1, and finally outputs the probabilities of three different lane change categories.

9. A vehicle lane change intention recognition system, characterized in that: include: An acquisition module, used to acquire the speed of the identified vehicle and the speeds of the vehicles around it; The intention recognition module inputs the characteristic parameters of the identified vehicle and its surrounding vehicles obtained in the real-time scenario into the first lane change recognition model in the highway section scenario if both speeds are greater than the preset value; otherwise, the characteristic parameters are input into the second lane change recognition model in the urban section scenario; wherein the characteristic parameters of the identified vehicle include speed, acceleration, abscissa and ordinate, and the characteristic parameters of the surrounding vehicles include speed and acceleration, as well as the lateral relative position and longitudinal relative position to the identified vehicle; The output module is used to identify the lane-changing intention of the vehicle through the first lane-changing recognition model or the second lane-changing recognition model and output it.

10. A computer storage medium, characterized in that: A computer program executable by a processor is stored therein, and the computer program executes the vehicle lane change intention recognition method described in any one of claims 1-8.

Images