A vehicle lane-changing behavior warning method based on hierarchical reasoning model

Abstract

The present invention provides a vehicle lane change behavior warning method based on a hierarchical reasoning model, which can be applied to the technical fields of advanced driver assistance and autonomous driving. The method includes: calculating the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle through a hierarchical reasoning model, wherein the hierarchical reasoning model includes a first module and a second module; according to the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle, calculating the lane change probability of the target vehicle through the attention weight generation network and lane change probability prediction network of the first module; according to the lane change probability of the target vehicle, determining the current state of the target vehicle through the Gaussian hidden Markov model of the second module; according to the current state of the target vehicle, deciding whether to issue an alarm signal or implement deceleration through the hierarchical reasoning model. The above method provided by the present invention can help the driver or the autonomous driving vehicle to react in advance, prevent collisions with lane-changing vehicles, and improve the safety of the traffic system.

Description

Technical Field

[0001] The present invention relates to the fields of advanced driver assistance and autonomous driving, and in particular to a vehicle lane change behavior warning method based on a hierarchical reasoning model. Background Art

[0002] Whether for advanced driver assistance or autonomous driving vehicles, identifying the target vehicle's possible lane change behavior in advance and implementing deceleration operations are of great significance to ensuring safety and improving comfort.

[0003] Different stages of lane change behavior need to be characterized by different features. Specifically: in the early stages before the lane change, the target vehicle will frequently adjust its speed to ensure a safe distance from other vehicles in its lane and the target lane; in the stage close to the lane change, the target vehicle will produce obvious lateral position offset and speed changes to smoothly cut into the target lane. To achieve reliable lane change behavior warning, it is necessary to fully consider the motion characteristics of the target vehicle at different stages. At the same time, considering that in actual driving, the lane keeping state accounts for the vast majority of all driving states, the lane change warning method should ensure a low false alarm rate to meet the requirements of actual scenarios. In summary, designing a reliable vehicle lane change behavior warning method that can identify the lane change intention as early as possible while ensuring a low false alarm rate is a hot issue in this field. Summary of the Invention

[0004] In view of the above problems, the present invention provides a vehicle lane change behavior warning method based on a hierarchical reasoning model, in order to solve at least one of the above technical problems.

[0005] According to a first aspect of the present invention, a vehicle lane change behavior warning method based on a hierarchical reasoning model is provided, characterized by comprising:

[0006] Calculating the autocorrelation motion features and the cross-correlation motion features of the target vehicle through a hierarchical reasoning model, wherein the hierarchical reasoning model includes a first module and a second module;

[0007] According to the autocorrelation motion features and cross-correlation motion features of the target vehicle, the lane change probability of the target vehicle is calculated through the attention weight generation network and lane change probability prediction network of the first module;

[0008] According to the lane change probability of the target vehicle, the current state of the target vehicle is determined through the Gaussian Hidden Markov Model of the second module;

[0009] According to the current state of the target vehicle, a hierarchical reasoning model is used to decide whether to issue an alarm signal or implement deceleration.

[0010] According to an embodiment of the present invention, the above-mentioned calculation of the autocorrelation motion features and cross-correlation motion features of the target vehicle by using the hierarchical reasoning model includes:

[0011] The autocorrelation motion characteristics of the target vehicle are calculated using a hierarchical reasoning model based on the longitudinal velocity, lateral velocity, longitudinal acceleration of the target vehicle and the distance between the target vehicle and the lane boundary.

[0012] The hierarchical reasoning module is used to vector-splice the characteristic information of the target vehicle's surrounding vehicles, the relative longitudinal speed of the target vehicle and the surrounding vehicles, and the longitudinal distance between the target vehicle and the surrounding vehicles to obtain the cross-correlation motion characteristics of the target vehicle.

[0013] According to an embodiment of the present invention, the above-mentioned calculation of the lane change probability of the target vehicle by the attention weight generation network and the lane change probability prediction network of the first module based on the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle includes:

[0014] The autocorrelation motion features and cross-correlation motion features are processed separately through the attention weight generation network to obtain the attention weights of the autocorrelation motion features and the attention weights of the cross-correlation motion features respectively;

[0015] Processing the autocorrelation motion feature through a first lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the autocorrelation feature;

[0016] Processing the cross-correlation motion features through a second lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the cross-correlation features;

[0017] The first module calculates the attention weight of the autocorrelation motion feature and the lane change probability based on the autocorrelation feature to obtain the autocorrelation feature calculation result;

[0018] The first module calculates the attention weight of the cross-correlation motion feature and the lane change probability based on the cross-correlation feature to obtain a cross-correlation feature calculation result;

[0019] The first module performs vector operations on the autocorrelation feature calculation results and the cross-correlation feature calculation results to obtain the lane change probability of the target vehicle.

[0020] According to an embodiment of the present invention, the above-mentioned attention weight generation network processes the autocorrelation motion features and the cross-correlation motion features respectively, and obtains the attention weights of the autocorrelation motion features and the attention weights of the cross-correlation motion features respectively, including:

[0021] The autocorrelation motion features and cross-correlation motion features are vector-concatenated through the attention weight generation network to obtain the feature vector concatenation result;

[0022] The feature vector concatenation result is classified by the learnable fully connected layer of the attention weight generation network to obtain the feature vector classification result;

[0023] The feature vector classification results are processed by normalized exponential function to obtain the attention weights of autocorrelation motion features and cross-correlation motion features.

[0024] According to an embodiment of the present invention, the first lane change probability prediction network processes the autocorrelation motion feature to obtain the lane change probability based on the autocorrelation feature, including:

[0025] Performing vector concatenation of the autocorrelation motion feature at the current moment and the hidden state of the first long short-term memory neural network of the first lane change probability prediction network at the previous moment to obtain the autocorrelation vector concatenation result at the current moment;

[0026] Processing the concatenated autocorrelation vectors at the current moment through the input gate of the first long short-term memory neural network and the learnable linear transformation layer of the first lane change probability prediction network, and performing a Hadamard product operation on the processed results to obtain a first Hadamard product;

[0027] The concatenation result of the autocorrelation vectors at the current moment is processed by the forget gate of the first long short-term memory neural network, and the processing result is subjected to a Hadamard product operation with the cell state of the first long short-term memory neural network at the previous moment to obtain a second Hadamard product;

[0028] Performing a vector element-wise addition operation on the first Hadamard product and the second Hadamard product to obtain the cell state of the first long short-term memory neural network at the current moment;

[0029] Processing the concatenated result of the autocorrelation vectors at the current moment through the output gate of the first long short-term memory neural network to obtain a first processing result;

[0030] Processing the cell state of the first long short-term memory neural network at the current moment through the learnable linear transformation layer of the first lane change probability prediction network to obtain a second processing result;

[0031] Performing a Hadamard product operation on the first processing result and the second processing result to obtain a hidden state of the first long short-term memory neural network at the current moment;

[0032] The hidden state of the first long short-term memory neural network at the current moment is processed by multiple nonlinear activation functions, multiple learnable fully connected layers of the first lane change probability prediction network, and a normalization function to obtain the lane change probability based on the autocorrelation feature.

[0033] According to an embodiment of the present invention, the second lane change probability prediction network processes the cross-correlation motion features to obtain the lane change probability based on the cross-correlation features, including:

[0034] Performing vector concatenation of the cross-correlation motion features at the current moment and the hidden state of the second long short-term memory neural network of the second lane change probability prediction network at the previous moment to obtain a cross-correlation vector concatenation result at the current moment;

[0035] Processing the concatenated cross-correlation vectors at the current moment through the input gate of the second long short-term memory neural network and the learnable linear transformation layer of the second lane change probability prediction network, and performing a Hadamard product operation on the processed results to obtain a third Hadamard product;

[0036] The concatenation result of the cross-correlation vectors at the current moment is processed by the forget gate of the second long short-term memory neural network, and the processing result is subjected to a Hadamard product operation with the cell state of the second long short-term memory neural network at the previous moment to obtain a fourth Hadamard product;

[0037] Performing a vector element-wise addition operation on the third Hadamard product and the fourth Hadamard product to obtain the cell state of the second long short-term memory neural network at the current moment;

[0038] Processing the concatenated result of the cross-correlation vectors at the current moment through the output gate of the second long short-term memory neural network to obtain a third processing result;

[0039] Processing the cell state of the second long short-term memory neural network at the current moment through the learnable linear transformation layer of the second lane change probability prediction network to obtain a fourth processing result;

[0040] Performing a Hadamard product operation on the third processing result and the fourth processing result to obtain a hidden state of the second long short-term memory neural network at the current moment;

[0041] The hidden state of the second long short-term memory neural network at the current moment is processed through multiple nonlinear activation functions, multiple learnable fully connected layers of the second lane change probability prediction network, and a normalization function to obtain the lane change probability based on the cross-correlation feature.

[0042] According to an embodiment of the present invention, the determining of the current state of the target vehicle by the Gaussian Hidden Markov Model of the second module based on the lane change probability of the target vehicle includes:

[0043] Construct a Gaussian hidden Markov model to describe the state changes of the target vehicle and initialize the parameters of the Gaussian hidden Markov model;

[0044] According to the lane-changing probability of the target vehicle, an observation sequence from the initial moment to the current moment is constructed;

[0045] The Gaussian Hidden Markov Model is used to process the observation sequence and obtain the current state of the target vehicle.

[0046] According to an embodiment of the present invention, the hidden state values ​​of the Gaussian Hidden Markov Model include lane keeping state, warning state and lane changing state;

[0047] Based on the Baum-Welch algorithm, the parameters of the Gaussian Hidden Markov Model are solved by maximizing the score of the observation sequence.

[0048] According to an embodiment of the present invention, the above-mentioned determination of whether to issue an alarm signal or implement deceleration based on the current state of the target vehicle through the hierarchical reasoning model includes:

[0049] When the current state of the target vehicle is a lane keeping state, the driver of the current vehicle or the autonomous driving vehicle performs the original operation;

[0050] If the current state of the target vehicle is an alarm state, the driver of the current vehicle or the autonomous driving vehicle performs an appropriate deceleration operation;

[0051] In the case where the current state of the target vehicle is a lane-changing state, the driver of the current vehicle or the autonomous driving vehicle performs an immediate deceleration operation.

[0052] According to an embodiment of the present invention, the attention weight generation network and lane change probability prediction network of the above-mentioned first module are obtained by training with a focused cross entropy loss function.

[0053] The present invention designs a vehicle lane-changing behavior warning method based on a hierarchical reasoning model. Based on the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle, the lane-changing probability is calculated using a lane-changing probability prediction network and an attention weight generation network. According to the lane-changing probability at historical moments, a Gaussian hidden Markov model of the target vehicle's state change is constructed to obtain the most likely target vehicle state at the current moment. Based on the target vehicle state at the current moment, it is determined whether an alarm signal needs to be sent or a deceleration action needs to be performed, thereby ensuring vehicle driving safety and improving the efficiency of the transportation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0055] Figure 1 is a flowchart of a vehicle lane change behavior warning method based on a hierarchical reasoning model according to an embodiment of the present invention;

[0056] Figure 2is a flow chart of calculating the autocorrelation motion features and cross-correlation motion features of a target vehicle according to an embodiment of the present invention;

[0057] Figure 3 is a schematic diagram of an application scenario of a vehicle lane change behavior warning method based on a hierarchical reasoning model according to an embodiment of the present invention;

[0058] Figure 4 is a flow chart of calculating the lane change probability of a target vehicle according to an embodiment of the present invention;

[0059] Figure 5 This is a data flow diagram of the first module calculating the lane change probability according to an embodiment of the present invention.

[0060] Figure 6 is a schematic diagram of the structure of a single lane change probability prediction network according to an embodiment of the present invention;

[0061] Figure 7 is a flow chart for determining the current state of a target vehicle according to an embodiment of the present invention;

[0062] FIG8( a ) is a motion trajectory of a target vehicle, a vehicle behind it in the target lane, and other vehicles when the target vehicle turns left according to an embodiment of the present invention;

[0063] FIG8( b ) is a lateral position change curve and a lane change curve of a target vehicle when the target vehicle turns left according to an embodiment of the present invention;

[0064] FIG8( c ) is a curve showing the target vehicle state output result at different times when the target vehicle turns left according to an embodiment of the present invention;

[0065] FIG8( d ) is a curve showing a change in the longitudinal position of a target vehicle when the target vehicle turns left according to an embodiment of the present invention;

[0066] FIG8(e) is a curve showing the speed change of the target vehicle when the target vehicle turns left according to an embodiment of the present invention.

[0067] FIG9( a ) is a motion trajectory of a target vehicle, a vehicle behind it in the target lane, and other vehicles when the target vehicle turns right according to an embodiment of the present invention;

[0068] FIG9( b ) is a lateral position change curve and a lane change curve of a target vehicle when the target vehicle turns right according to an embodiment of the present invention;

[0069] FIG9( c ) is a curve showing the target vehicle state output result at different times when the target vehicle turns right according to an embodiment of the present invention;

[0070] FIG9( d ) is a curve showing a change in the longitudinal position of a target vehicle when the target vehicle turns right according to an embodiment of the present invention;

[0071] FIG9( e ) is a speed change curve of the target vehicle when the target vehicle turns right according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0073] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concept of the present invention.

[0074] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise," "include," etc. used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0075] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0076] When expressions such as "at least one of A, B and C, etc." are used, they should generally be interpreted in accordance with the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0077] In the technical solution of the present invention, the acquisition, storage and application of vehicle information involved have obtained the authorization and permission of the information owner, complies with the provisions of relevant laws and regulations, takes necessary confidentiality measures, and does not violate public order and good morals.

[0078] Figure 1 4 is a flowchart of a vehicle lane change behavior warning method based on a hierarchical reasoning model according to an embodiment of the present invention.

[0079] like Figure 1 As shown, the above-mentioned vehicle lane change behavior warning method based on the hierarchical reasoning model is characterized by including operations S110 to S140.

[0080] In operation S110 , an autocorrelation motion feature and a cross-correlation motion feature of the target vehicle are calculated using a hierarchical reasoning model, wherein the hierarchical reasoning model includes a first module and a second module.

[0081] Figure 2 4 is a flow chart of calculating the autocorrelation motion features and cross-correlation motion features of a target vehicle according to an embodiment of the present invention.

[0082] like Figure 2 As shown, the above-mentioned calculation of the autocorrelation motion features and the cross-correlation motion features of the target vehicle through the hierarchical reasoning model includes operations S210 to S220.

[0083] In operation S210 , an autocorrelation motion feature of the target vehicle is calculated using a hierarchical reasoning model based on the longitudinal velocity, lateral velocity, longitudinal acceleration of the target vehicle and the distance between the target vehicle and the boundary lane line.

[0084] In operation S210, the hierarchical reasoning module performs vector concatenation on the feature information of the target vehicle's surrounding vehicles, the relative longitudinal speed of the target vehicle and the surrounding vehicles, and the longitudinal distance between the target vehicle and the surrounding vehicles to obtain the cross-correlation motion feature of the target vehicle.

[0085] Surrounding vehicles refer to vehicles that are adjacent to the target vehicle and can affect the driving of the target vehicle.

[0086] In order to better illustrate the advantages of the method provided by the present invention in calculating the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle, the following is combined with the attached Figure 3 The above operations S210 to S220 are further described in detail with reference to the specific embodiments.

[0087] Figure 3 3 is a schematic diagram of an application scenario of a vehicle lane change behavior warning method based on a hierarchical reasoning model according to an embodiment of the present invention.

[0088] like Figure 3 As shown, the target vehicle is denoted as TV. The surrounding vehicles include other vehicles traveling in the original lane and the target lane, which are denoted as the preceding vehicle PV in the original lane, the following vehicle FV in the original lane, the preceding vehicle SPV in the target lane, the adjacent vehicle SAV in the target lane, and the following vehicle SFV in the target lane. The target vehicle's cut-in time is defined as the time when the center point of the target vehicle crosses the boundary line between the original lane and the target lane, denoted as T. c The coordinate axis is defined as the horizontal axis along the lane forward, recorded as the x-axis, and the vertical axis perpendicular to the lane to the left, recorded as the y-axis. The vertical coordinate of the intersection lane line between the original lane and the target lane is y L .

[0089] Specifically, the autocorrelation motion feature is calculated according to the following formula (1):

[0090]

[0091] in, Represents the autocorrelation motion characteristics of the target vehicle at time t. Δy m Indicates the distance between the target vehicle and the boundary lane line, that is, Δy m =y TV -y L ,y TV is the vertical coordinate of the target vehicle. TV Indicates the longitudinal velocity of the target vehicle. u TV Indicates the lateral velocity of the target vehicle. a TV Indicates the longitudinal acceleration of the target vehicle. Motion characteristics It is only related to the state of the target vehicle itself and has nothing to do with the surrounding vehicles, so it is called autocorrelation motion feature.

[0092] Specifically, the cross-correlation motion feature is calculated according to the following formula (2):

[0093]

[0094] in, ) represents the cross-correlation motion characteristics of the target vehicle at time t. Represents the set of surrounding vehicles, corresponding to the front vehicle in the original lane, the rear vehicle in the original lane, the front vehicle in the target lane, the adjacent vehicle in the target lane, and the rear vehicle in the target lane. The subscript * in the variable represents a surrounding vehicle*. * Indicates whether the surrounding vehicle* exists, 1 if it exists, and 0 if it does not. Δx * Indicates the relative longitudinal distance between the target vehicle and surrounding vehicles*, i.e. Δx * =x TV -x * , x TV and x * are the horizontal coordinates of the target vehicle and the surrounding vehicles* respectively. Δv * Indicates the relative longitudinal velocity of the target vehicle and surrounding vehicles*, i.e. Δv * =v TV -v * , v TV and v * are the longitudinal speeds of the target vehicle and surrounding vehicles*, respectively. Symbol Express The feature vectors of all surrounding vehicles in the splicing operation are combined. It is related to both the target vehicle’s own state and the states of surrounding vehicles, so it is called cross-correlation motion feature.

[0095] In operation S120 , the lane change probability of the target vehicle is calculated by the attention weight generation network and the lane change probability prediction network of the first module based on the autocorrelation motion features and the cross-correlation motion features of the target vehicle.

[0096] The first module includes an attention weight generation network and two lane change probability prediction networks, wherein the first lane change probability prediction network is used to process autocorrelation motion features, and the second lane change probability prediction network is used to process cross-correlation motion features. The first lane change probability prediction network and the second lane change probability prediction network have similar network structures but different parameters.

[0097] The first module includes an attention weight generation network and two lane change probability prediction networks, which are trained through a preset training method, so that the parameters of the first module are optimized.

[0098] Figure 4 4 is a flowchart of calculating the lane change probability of a target vehicle according to an embodiment of the present invention.

[0099] like Figure 4 As shown, the above-mentioned calculation of the lane change probability of the target vehicle based on the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle through the attention weight generation network and the lane change probability prediction network of the first module includes operations S410 to S460.

[0100] In operation S410, the autocorrelation motion feature and the cross-correlation motion feature are processed separately through an attention weight generation network to obtain attention weights of the autocorrelation motion feature and attention weights of the cross-correlation motion feature, respectively.

[0101] In operation S420 , the autocorrelation motion feature is processed by a first lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the autocorrelation feature.

[0102] In operation S430 , the cross-correlation motion feature is processed by a second lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the cross-correlation feature.

[0103] In operation S440 , the attention weight of the autocorrelation motion feature and the lane change probability based on the autocorrelation feature are calculated by the first module to obtain an autocorrelation feature calculation result.

[0104] In operation S450 , the attention weight of the cross-correlation motion feature and the lane change probability based on the cross-correlation feature are calculated by the first module to obtain a cross-correlation feature calculation result.

[0105] In operation S460, the first module performs vector operation on the autocorrelation feature operation result and the cross-correlation feature operation result to obtain the lane change probability of the target vehicle.

[0106] The following is combined with Figure 5 and Figure 6 As well as specific embodiments, the above operations S410 to S460 are further described in detail.

[0107] Figure 5 is a data flow chart of calculating lane change probability according to the first module of an embodiment of the present invention; Figure 5 As shown in the figure, first, the lane change probability prediction network output based on the autocorrelation motion features and cross-correlation motion features is calculated respectively; secondly, the attention weight generation network output is calculated; finally, the lane change probability of the target vehicle is calculated.

[0108] Figure 6 This is a schematic diagram of the structure of a single lane change probability prediction network according to an embodiment of the present invention, based on the autocorrelation motion feature and cross-correlation motion features The lane change prediction network outputs can be calculated using the following formulas (3) to (6). The network inputs differ for different motion features, but the calculation process remains the same. For convenience, the superscript (i) of the variables is used to distinguish between autocorrelation motion features (i=1) and cross-correlation motion features (i=2), and no further explanation is given.

[0109]

[0110]

[0111]

[0112]

[0113] The subscript t of the variable represents the current time, and t-1 represents the previous time. Represents motion features, i=1 represents autocorrelation motion features, and i=2 represents cross-correlation motion features. Represents the hidden state of the long short-term memory neural network. Represents the cell state of a long short-term memory neural network. express and The concatenated feature vector, symbol Represents a vector concatenation operation. Represents the final output probability vector. The symbol ⊙ represents element-wise multiplication of vectors, and the symbol + represents element-wise addition of vectors. Represents the forget gate of the long short-term memory neural network, which consists of a learnable linear transformation layer and a sigmoid activation function. Represents the input gate of the long short-term memory neural network, which is composed of a learnable linear transformation layer and a sigmoid activation function. Represents the output gate of the long short-term memory neural network, which consists of a learnable linear transformation layer and a sigmoid activation function. Φ represents a learnable linear transformation layer with a tanh activation function as the output. and are two learnable fully connected layers. Tanh and ReLU are two nonlinear activation functions. SoftMax is a normalized exponential function.

[0114] See also Figure 6 , the output dimension of the long short-term memory network is 256, and the fully connected layer The output dimension of the fully connected layer is 32. The output dimension of is 2, and the dimension of the probability vector output after SoftMax is 2.

[0115] See also Figure 6 , to examine a certain motion feature Under the condition of , the output of the lane change probability prediction network at the current moment is calculated according to the following operations:

[0116] From time tT o +1 to time t, total T o The motion states of each historical moment are sequentially input into the long short-term memory network in time sequence, and the hidden state of the long short-term memory network at the current moment t is output.

[0117] Hidden state After the tanh activation function, the fully connected layer ReLU activation function, fully connected layer SoftMax normalized exponential function, based on motion features The output of the lane change probability prediction network at the current moment.

[0118] Specifically, in operation S410, the attention weight is calculated according to the following formula (7) to generate the network output:

[0119]

[0120] in, and denote the attention weights of the autocorrelation motion features and the cross-correlation motion features, respectively. and Represent the autocorrelation motion features and cross-correlation motion features respectively, represents the vector concatenation operation, It is a learnable fully connected layer in the attention weight generation network, and SoftMax is a normalized exponential function.

[0121] Specifically, the target vehicle lane change probability is calculated according to the following formula (8):

[0122]

[0123] in, and They represent the lane changing probability and lane keeping probability of the target vehicle respectively.

[0124] The following further describes in detail the process of obtaining the lane change probability based on the autocorrelation motion feature by the first lane change probability prediction network processing the autocorrelation motion feature in conjunction with a specific embodiment.

[0125] According to an embodiment of the present invention, processing the autocorrelation motion feature through the first lane change probability prediction network of the lane change probability prediction network to obtain the lane change probability based on the autocorrelation feature includes: performing vector splicing on the autocorrelation motion feature at the current moment and the hidden state of the first long short-term memory neural network of the first lane change probability prediction network at the previous moment to obtain the autocorrelation vector splicing result at the current moment; processing the autocorrelation vector splicing result at the current moment through the input gate of the first long short-term memory neural network and the learnable linear transformation layer of the first lane change probability prediction network, and performing a Hadamard product operation on the processing result to obtain a first Hadamard product; processing the autocorrelation vector splicing result at the current moment through the forget gate of the first long short-term memory neural network, and performing a Hadamard product operation on the processing result and the cell state of the first long short-term memory neural network at the previous moment to obtain second Hadamard product; performing element-wise addition operation on the first Hadamard product and the second Hadamard product to obtain the cell state of the first long short-term memory neural network at the current moment; processing the autocorrelation vector splicing result at the current moment through the output gate of the first long short-term memory neural network to obtain a first processing result; processing the cell state of the first long short-term memory neural network at the current moment through the learnable linear transformation layer of the first lane change probability prediction network to obtain a second processing result; performing Hadamard product operation on the first processing result and the second processing result to obtain the hidden state of the first long short-term memory neural network at the current moment; processing the hidden state of the first long short-term memory neural network at the current moment through multiple nonlinear activation functions, multiple learnable fully connected layers of the first lane change probability prediction network and a normalization function to obtain a lane change probability based on the autocorrelation feature.

[0126] The following further describes in detail the process of the second lane change probability prediction network processing the cross-correlation motion features to obtain the lane change probability based on the cross-correlation motion features in conjunction with specific embodiments.

[0127] According to an embodiment of the present invention, the second lane change probability prediction network of the lane change probability prediction network processes the cross-correlation motion features to obtain the lane change probability based on the cross-correlation features, including: performing vector splicing on the cross-correlation motion features at the current moment and the hidden state of the second long short-term memory neural network of the second lane change probability prediction network at the previous moment to obtain the cross-correlation vector splicing result at the current moment; processing the cross-correlation vector splicing result at the current moment through the input gate of the second long short-term memory neural network and the learnable linear transformation layer of the second lane change probability prediction network, and performing a Hadamard product operation on the processing result to obtain a third Hadamard product; processing the cross-correlation vector splicing result at the current moment through the forget gate of the second long short-term memory neural network, and performing a Hadamard product operation on the processing result and the cell state of the second long short-term memory neural network at the previous moment to obtain fourth Hadamard product; performing element-wise addition operation on the third Hadamard product and the fourth Hadamard product to obtain the cell state of the second long short-term memory neural network at the current moment; processing the cross-correlation vector splicing result at the current moment through the output gate of the second long short-term memory neural network to obtain a third processing result; processing the cell state of the second long short-term memory neural network at the current moment through the learnable linear transformation layer of the second lane change probability prediction network to obtain a fourth processing result; performing Hadamard product operation on the third processing result and the fourth processing result to obtain the hidden state of the second long short-term memory neural network at the current moment; processing the hidden state of the second long short-term memory neural network at the current moment through multiple nonlinear activation functions, multiple learnable fully connected layers of the second lane change probability prediction network and a normalization function to obtain the lane change probability based on the cross-correlation feature.

[0128] In operation S130 , the current state of the target vehicle is determined by the Gaussian Hidden Markov Model of the second module according to the lane change probability of the target vehicle.

[0129] Figure 7 is a flow chart of determining the current state of a target vehicle according to an embodiment of the present invention.

[0130] like Figure 7 As shown, the above-mentioned determination of the current state of the target vehicle by the Gaussian Hidden Markov Model of the second module according to the lane change probability of the target vehicle includes operations S610 to S630.

[0131] In operation S710 , a Gaussian Hidden Markov Model for describing a state change of a target vehicle is constructed, and parameters of the Gaussian Hidden Markov Model are initialized.

[0132] In operation S720 , an observation sequence from an initial time to a current time is constructed according to the lane change probability of the target vehicle.

[0133] In operation S730 , the observation sequence is processed using a Gaussian Hidden Markov Model to obtain the current state of the target vehicle.

[0134] According to an embodiment of the present invention, the hidden state values ​​of the above-mentioned Gaussian hidden Markov model include lane keeping state, warning state and lane changing state; based on the Baum-Welch algorithm and by maximizing the score of the observation sequence, the parameters of the Gaussian hidden Markov model are solved.

[0135] The following is a collection of specific implementation methods to further illustrate the above operations S610 to S630 in detail.

[0136] In operation S610, a Gaussian hidden Markov model of the target vehicle state change is constructed by the following formulas (9)-(12):

[0137]

[0138]

[0139]

[0140]

[0141] in, represents a Gaussian hidden Markov model. Represents the hidden state value set, H k Indicates lane keeping status, H w Indicates alarm status, H c Indicates lane change status. t Represents the hidden state at time t. π0 ​​represents the distribution law of each hidden state at the initial moment, which belongs to subspace; where each element π 0i Represents the hidden state H i The probability of the initial moment. Α represents the transition probability matrix between the hidden states between two consecutive moments, which belongs to subspace; where each element a j,i Represents the hidden state H at the previous moment j Transfer to the current hidden state H i The probability of B is obtained according to the hidden state. The emission probability matrix belongs to subspace; where each element b i Represents the current hidden state H i Get observations The probability of b.i Subject to the mean μ i , the variance is Gaussian distribution

[0142] The probability of each hidden state of the target vehicle at the current moment is calculated according to the following formulas (13) and (14):

[0143]

[0144]

[0145] in, represents an observation sequence consisting of the lane change probability of the target vehicle from the initial time to the current time. Represents the current moment, given the observation sequence Under the condition that the hidden state is H i The conditional probability of . Indicates the current moment, obtains the observation sequence And the hidden state is H i The probability of is obtained by the recursive form described by formula (14).

[0146] Specifically, the most likely target vehicle state at the current moment is determined according to the following formula (15):

[0147]

[0148] in, That is the most likely target vehicle state at the current moment.

[0149] In operation S140 , a decision is made through a hierarchical reasoning model based on the current state of the target vehicle whether to issue an alarm signal or implement deceleration.

[0150] According to an embodiment of the present invention, the above-mentioned decision of whether to issue an alarm signal or implement deceleration based on the current state of the target vehicle through the hierarchical reasoning model includes the following situations.

[0151] In the case where the current state of the target vehicle is the lane keeping state, the driver of the current vehicle or the autonomous driving vehicle performs the original operation.

[0152] In the case where the current state of the target vehicle is an alarm state, the driver of the current vehicle or the autonomous driving vehicle performs an appropriate deceleration operation.

[0153] In the case where the current state of the target vehicle is a lane-changing state, the driver of the current vehicle or the autonomous driving vehicle performs an immediate deceleration operation.

[0154] According to the most likely target vehicle state at the current moment Determine the specific actions the driver or self-driving car should take:

[0155] when When the target vehicle is in lane keeping mode, the driver or the autonomous vehicle does not need to take any additional measures and continues to perform the original operation such as following the vehicle in front or cruising at a constant speed.

[0156] when When the target vehicle is in an alarm state, the driver or the autonomous vehicle should always pay attention to the possible lane change of the target vehicle and slow down in advance to ensure a safe following distance between them.

[0157] when When the target vehicle is in a lane-changing state, the driver or the autonomous vehicle should immediately decelerate until the distance between them and the target vehicle reaches a safe distance to prevent a collision with the target vehicle.

[0158] The present invention designs a vehicle lane-changing behavior warning method based on a hierarchical reasoning model. Based on the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle, the lane-changing probability is calculated using a lane-changing probability prediction network and an attention weight generation network. According to the lane-changing probability at historical moments, a Gaussian hidden Markov model of the target vehicle's state change is constructed to obtain the most likely target vehicle state at the current moment. Based on the target vehicle state at the current moment, it is determined whether an alarm signal needs to be sent or a deceleration action needs to be performed, thereby ensuring vehicle driving safety and improving the efficiency of the transportation system.

[0159] In the above method provided by the present invention, the first module having an attention weight generation network and two lane change probability prediction networks is trained by using a focused cross entropy loss function, wherein the focused cross entropy loss function is shown in formula (16):

[0160]

[0161] Among them, Loss FCE Represents the cross entropy loss function value with focus; α F It represents the balance factor, which is the proportion of lane keeping data in all data; γ represents the focal length, which is 2; lg() represents the logarithm with base 10; represents the probability of the target vehicle changing lanes at time t; represents the probability that the target vehicle maintains its lane at time t

[0162] The second module with Gaussian Hidden Markov Model uses Baum-Welch Algorithm to solve the optimal Gaussian Hidden Markov Model parameters by maximizing the score of the observation sequence. The score of the observation sequence is shown in formula (17)

[0163]

[0164] Among them, Score represents the score of the observation sequence; represents an observation sequence consisting of the probability of the target vehicle changing lanes from the initial moment to the current moment; Represents the set of hidden state values ​​of the Gaussian Hidden Markov Model; Indicates the current moment, obtains the observation sequence And the hidden state is H i probability.

[0165] The following is a specific implementation method and combined with the attached Figure 7 and 8, further describing in detail the advantages of the above method provided by the present invention.

[0166] Figure 8 shows a situation where the target vehicle turns left according to an embodiment of the present invention, wherein Figure 8(a) shows the motion trajectories of the target vehicle, the vehicle behind the target lane, and other vehicles according to an embodiment of the present invention; Figure 8(b) shows the lateral position change curve and lane change curve of the target vehicle according to an embodiment of the present invention; Figure 8(c) shows the target vehicle state output result curve at different times according to an embodiment of the present invention; Figure 8(d) shows the target vehicle longitudinal position change curve according to an embodiment of the present invention; and Figure 8(e) shows the target vehicle speed change curve according to an embodiment of the present invention.

[0167] Figure 9 illustrates a target vehicle turning right according to an embodiment of the present invention. Figure 9(a) shows the motion trajectories of the target vehicle, the vehicle behind it in the target lane, and other vehicles according to an embodiment of the present invention; Figure 9(b) shows the lateral position change curve and lane change curve of the target vehicle according to an embodiment of the present invention; Figure 9(c) shows the target vehicle state output curves at different times according to an embodiment of the present invention; Figure 9(d) shows the longitudinal position change curve of the target vehicle according to an embodiment of the present invention; and Figure 9(e) shows the speed change curve of the target vehicle according to an embodiment of the present invention.

[0168] Specifically, in the driver-in-the-loop experiment, the target vehicle is allowed to be controlled by a human driver through a driving simulator, and the operating objects include the steering wheel, accelerator pedal, brake pedal, and lane change indicator button; the surrounding vehicles are controlled by programs to achieve autonomous driving lane keeping and adaptive following functions, and the reference speed is set to 40km / h; the specific controller function is implemented using the intelligent driver model introduced in the reference document Congested traffic states in empirical observations and microscopic simulations (M.Treiber, 2020). The moment when the human driver presses the lane change indicator button is recorded as the start time of the target vehicle's lane change behavior, and the moment when the center point of the target vehicle's body crosses the intersection lane line is recorded as the cut-in time. The effectiveness of the technical solution of the present invention is determined by observing the driving behavior of the vehicle behind the target lane.

[0169] The “5% false alarm” and “1% false alarm” in FIG8(c) and FIG9(c) indicate that the lane change probability of the target vehicle is 0. t LC Select the binary classification threshold and divide the output state into lane keeping state H k and lane change status H c When the false alarm rate is 5% or 1%, the corresponding output state curve is obtained, and "GHMM" represents the output state curve of the lower layer Gaussian Hidden Markov Module designed by the present invention.

[0170] In Figures 8(d)(e) and 9(d)(e), “No intent recognition” indicates the result without using the technical solution of the present invention, “No GHMM” indicates the result using only the upper-layer module designed by the present invention, and “Complete method” indicates the result using the complete technical solution of the present invention.

[0171] As shown in Figures 8(a), (b), and (c), the target vehicle begins a left lane change at 125.9 seconds and crosses the lane marking to the left at 129.7 seconds. As shown in Figure 8(c), the "5% false alarm" solution exhibits a long period of false alarms before the start time, misinterpreting lane keeping as lane change. However, the "1% false alarm" solution and the "GHMM" solution exhibit no false alarm issues. Furthermore, compared to the "1% false alarm" solution, the "GHMM" solution provides an earlier warning state, prompting the driver or the autonomous vehicle to decelerate. As shown in Figures 8(d) and (e), compared to the "no intention recognition" solution, the "no GHMM" solution provides earlier deceleration and a greater distance from the target vehicle at the time of cut-in. However, compared to the "no GHMM" solution, the "full method" solution of the present invention relies on the warning state to achieve earlier deceleration, and the deceleration during the warning phase is more moderate than that of the "no GHMM" solution.

[0172] As shown in Figures 9(a), (b), and (c), the target vehicle begins to change lanes to the right at 142 seconds and crosses the lane line to the right at 146.2 seconds. As shown in Figure 9(c), both the "5% false alarm" and "GHMM" schemes experience false alarms before the start time, but the "GHMM" scheme has a shorter false alarm duration, lasting only 0.5 seconds. The "1% false alarm" scheme experiences a period of missed alarms after the start time, meaning it fails to successfully determine the lane change status. As shown in Figures 9(d) and (e), compared to the "no intention recognition" scheme, the "no GHMM" scheme provides earlier deceleration and a greater distance from the target vehicle at the time of cut-in. However, compared to the "no GHMM" scheme, the "full method" scheme of the present invention can achieve earlier deceleration based on the warning state, and the deceleration during the warning phase is more moderate than that of the "no GHMM" scheme.

[0173] Those skilled in the art will appreciate that various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, even if such combinations and / or combinations are not explicitly described in the present invention. In particular, various combinations and / or combinations of features described in the various embodiments and / or claims of the present invention may be made, without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0174] The above describes embodiments of the present invention. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be advantageously used in combination. The scope of the present invention is defined by the appended claims and their equivalents. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which are intended to fall within the scope of the present invention.

Claims

1. A vehicle lane change behavior warning method based on a hierarchical reasoning model, characterized in that: include: Calculating the autocorrelation motion features and the cross-correlation motion features of the target vehicle through the hierarchical reasoning model, wherein the hierarchical reasoning model includes a first module and a second module; Calculating the lane change probability of the target vehicle through the attention weight generation network and the lane change probability prediction network of the first module based on the autocorrelation motion characteristics and cross-correlation motion characteristics of the target vehicle; Determining the current state of the target vehicle using a Gaussian Hidden Markov Model of the second module according to the lane change probability of the target vehicle; According to the current state of the target vehicle, determining whether to issue an alarm signal or implement deceleration through the hierarchical reasoning model; Calculating the lane change probability of the target vehicle by using the attention weight generation network and the lane change probability prediction network of the first module according to the autocorrelation motion characteristics and the cross-correlation motion characteristics of the target vehicle includes: Processing the autocorrelation motion feature and the cross-correlation motion feature respectively through the attention weight generation network to obtain the attention weight of the autocorrelation motion feature and the attention weight of the cross-correlation motion feature respectively; processing the autocorrelation motion feature by a first lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the autocorrelation feature; processing the cross-correlation motion feature by a second lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the cross-correlation feature; calculating, by the first module, the attention weight of the autocorrelation motion feature and the lane change probability based on the autocorrelation feature to obtain an autocorrelation feature calculation result; calculating, by the first module, the attention weight of the cross-correlation motion feature and the lane change probability based on the cross-correlation feature to obtain a cross-correlation feature calculation result; The first module performs vector operation on the autocorrelation feature operation result and the cross-correlation feature operation result to obtain the lane change probability of the target vehicle.

2. The method according to claim 1, characterized in that Calculating the autocorrelation motion features and cross-correlation motion features of the target vehicle through the hierarchical reasoning model includes: Calculating the autocorrelation motion characteristics of the target vehicle using the hierarchical reasoning model according to the longitudinal speed, lateral speed, longitudinal acceleration of the target vehicle and the distance between the target vehicle and the boundary lane line; The hierarchical reasoning model is used to perform vector splicing on the characteristic information of the surrounding vehicles of the target vehicle, the relative longitudinal speed of the target vehicle and the surrounding vehicles, and the longitudinal distance between the target vehicle and the surrounding vehicles to obtain the cross-correlated motion characteristics of the target vehicle.

3. The method according to claim 1, characterized in that Processing the autocorrelation motion feature and the cross-correlation motion feature respectively through the attention weight generation network to obtain the attention weight of the autocorrelation motion feature and the attention weight of the cross-correlation motion feature respectively includes: Performing vector splicing on the autocorrelation motion feature and the cross-correlation motion feature through the attention weight generation network to obtain a feature vector splicing result; Classify the feature vector concatenation result through the learnable fully connected layer of the attention weight generation network to obtain a feature vector classification result; The feature vector classification result is processed by a normalized exponential function to obtain the attention weight of the autocorrelation motion feature and the attention weight of the cross-correlation motion feature.

4. The method according to claim 1, wherein Processing the autocorrelation motion feature by a first lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the autocorrelation feature includes: Performing vector concatenation of the autocorrelation motion feature at the current moment and the hidden state of the first long short-term memory neural network of the first lane change probability prediction network at the previous moment to obtain an autocorrelation vector concatenation result at the current moment; Processing the concatenated autocorrelation vectors at the current moment through the input gate of the first long short-term memory neural network and the learnable linear transformation layer of the first lane change probability prediction network, respectively, and performing a Hadamard product operation on the processed results to obtain a first Hadamard product; Processing the concatenated autocorrelation vectors at the current moment through a forget gate of the first long short-term memory neural network, and performing a Hadamard product operation on the processed result and the cell state of the first long short-term memory neural network at the previous moment to obtain a second Hadamard product; Performing a vector element-wise addition operation on the first Hadamard product and the second Hadamard product to obtain a cell state of the first long short-term memory neural network at a current moment; Processing the autocorrelation vector splicing result at the current moment through the output gate of the first long short-term memory neural network to obtain a first processing result; Processing the cell state of the first long short-term memory neural network at the current moment through the learnable linear transformation layer of the first lane change probability prediction network to obtain a second processing result; Performing a Hadamard product operation on the first processing result and the second processing result to obtain a hidden state of the first long short-term memory neural network at a current moment; The hidden state of the first long short-term memory neural network at the current moment is processed by multiple nonlinear activation functions, multiple learnable fully connected layers of the first lane change probability prediction network, and a normalization function to obtain the lane change probability based on the autocorrelation feature.

5. The method according to claim 1, wherein Processing the cross-correlation motion feature by a second lane change probability prediction network of the lane change probability prediction network to obtain a lane change probability based on the cross-correlation feature includes: Performing vector concatenation of the cross-correlation motion feature at the current moment and the hidden state of the second long short-term memory neural network of the second lane change probability prediction network at the previous moment to obtain a cross-correlation vector concatenation result at the current moment; Processing the concatenated result of the cross-correlation vectors at the current moment through the input gate of the second long short-term memory neural network and the learnable linear transformation layer of the second lane change probability prediction network, respectively, and performing a Hadamard product operation on the processed result to obtain a third Hadamard product; Processing the concatenated result of the cross-correlation vectors at the current moment through the forget gate of the second long short-term memory neural network, and performing a Hadamard product operation on the processed result and the cell state of the second long short-term memory neural network at the previous moment to obtain a fourth Hadamard product; Performing a vector element-wise addition operation on the third Hadamard product and the fourth Hadamard product to obtain a cell state of the second long short-term memory neural network at the current moment; Processing the cross-correlation vector splicing result at the current moment through the output gate of the second long short-term memory neural network to obtain a third processing result; Processing the cell state of the second long short-term memory neural network at the current moment through the learnable linear transformation layer of the second lane change probability prediction network to obtain a fourth processing result; Performing a Hadamard product operation on the third processing result and the fourth processing result to obtain a hidden state of the second long short-term memory neural network at the current moment; The hidden state of the second long short-term memory neural network at the current moment is processed by multiple nonlinear activation functions, multiple learnable fully connected layers of the second lane change probability prediction network, and a normalization function to obtain the lane change probability based on the cross-correlation feature.

6. The method according to claim 1, characterized in that Determining the current state of the target vehicle by using the Gaussian Hidden Markov Model of the second module according to the lane change probability of the target vehicle includes: Constructing a Gaussian hidden Markov model for describing the state change of the target vehicle, and initializing parameters of the Gaussian hidden Markov model; Constructing an observation sequence from an initial moment to a current moment according to the lane change probability of the target vehicle; The Gaussian Hidden Markov Model is used to process the observation sequence to obtain the current state of the target vehicle.

7. The method according to claim 6, characterized in that The hidden state values ​​of the Gaussian Hidden Markov Model include lane keeping state, warning state and lane changing state; The parameters of the Gaussian Hidden Markov Model are solved based on the Baum-Welch algorithm by maximizing the score of the observation sequence.

8. The method according to claim 7, characterized in that Determining whether to issue an alarm signal or implement deceleration based on the current state of the target vehicle through the hierarchical reasoning model includes: When the current state of the target vehicle is a lane keeping state, the driver of the current vehicle or the autonomous driving vehicle performs the original operation; When the current state of the target vehicle is an alarm state, the driver of the current vehicle or the autonomous driving vehicle performs an appropriate deceleration operation; In a case where the current state of the target vehicle is a lane-changing state, the driver of the current vehicle or the automatic driving vehicle performs an immediate deceleration operation.

9. The method according to any one of claims 1 to 8, characterized in that The attention weight generation network and lane change probability prediction network of the first module are trained through a focused cross-entropy loss function.

Images