An autonomous prediction lane information model training method and device

Abstract

An embodiment of the present invention discloses a method and device for training an autonomous lane information prediction model. The method includes: obtaining historical vehicle state information and historical lane information, where the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is the lane position information passed by the vehicle; training a first model according to the historical vehicle state information, the historical lane information, and a first loss function, where the first model is a long short-term memory recurrent neural network model, the first model is used to autonomously predict the lane position information that the vehicle will pass through, the first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function. The present disclosure solves the problem of lane mismatch caused by missing lane information.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and more specifically, to a method and device for training an autonomous lane information prediction model. Background Art

[0002] In an autonomous driving or vehicle-road cooperation system, many in-vehicle application models require an intelligent vehicle or an intelligent vehicle terminal to determine the lane in which the vehicle is located, so that the real-time lane position of the vehicle can be used as the starting point for subsequent models. Examples of in-vehicle application models include a vehicle acceleration and deceleration control model, a vehicle lane change model, etc.

[0003] The basic idea of vehicle lane matching is to match the high-precision positioning of the vehicle itself with the high-precision map information of the vehicle. Since the high-precision map data is much larger than the existing map data, and there is a great deal of uncertainty in the position where the vehicle travels, the high-precision map data of the vehicle is generally not directly stored locally in the vehicle. Instead, the MAP information is sent by a Road Side Unit (RSU) through vehicle-road cooperation, so that the relevant high-precision map data information is transmitted to the vehicle side of autonomous driving. Affected by the RSU deployment location, wireless signals affected by the environment, multi-path reflection, obstacle occlusion, etc., an On-board Unit (OBU) of a vehicle-mounted intelligent terminal cannot receive the MAP information from the RSU in real time and continuously. Therefore, it affects the judgment of the lane information in which the vehicle is located in the autonomous driving model, and further affects the accuracy of the output data of a series of subsequent in-vehicle application models of autonomous driving.

[0004] In addition, even if the OBU of the vehicle-mounted intelligent terminal successfully receives the complete lane data sent by the RSU and its own high-precision positioning data, there will still be a certain degree of lane matching error due to reasons such as its own high-precision positioning data error and high-precision map original data error.

[0005] A common method for solving the problem of missing lane information in the existing technology is to cache the MAP data received in the historical vehicle-road cooperation system for use when it cannot be received next time, but this method is not applicable to the sections where the vehicle passes for the first time.

[0006] A common method for solving the problem caused by data error in the existing technology is to introduce a redundant threshold range, that is, the distance between the vehicle and the center line of the lane calculated is used as a judgment threshold, and it is set according to experience to enlarge or reduce the judgment threshold to make the matching result consistent with the actual situation. However, due to the complexity of the actual road, the judgment error of the redundant threshold with a linear relationship is very large.

[0007] How to complete the lane information of the section where the vehicle is located and eliminate the error of the redundant threshold in lane matching is a problem that needs to be solved in the field of autonomous driving. Summary of the Invention

[0008] The present invention provides a method and device for training an autonomous prediction lane information model, which solves the problem of lane mismatch caused by missing lane information. The specific technical solution is as follows:

[0009] In a first aspect, an embodiment of the present invention provides a method for training an autonomous prediction lane information model, the method comprising:

[0010] Obtain historical vehicle state information and historical lane information, where the historical vehicle state information is the input data for training the model, and the historical lane information is the output data for training the model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle movement information, and the historical lane information is the lane position information passed by the vehicle;

[0011] Train a first model according to the historical vehicle state information, the historical lane information, and a first loss function. The first model is a long short-term memory recurrent neural network model, and the first model is used to autonomously predict the lane position information that the vehicle is about to pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0012] Optionally, the formula of the first model:

[0013] f t = σ g (W f x t + U f h t-1 + b f )

[0014] i t = σ g (W i x t + U i h t-1 + b i )

[0015] o t = σ g (W o x t + U o h t-1 + b o )

[0016]

[0017]

[0018]

[0019] Among them, x t is the input feature vector, h t is the output vector, f t is the excitation vector for controlling the forgetting gate, i t is the excitation vector for controlling the input gate, o t is the excitation vector for controlling the output gate, is the cell input excitation vector, c t is the cell state vector, W and U are weights, b is the offset, and σ is the activation function.

[0020] Optionally, the vehicle high-precision positioning information includes the longitude, latitude, and elevation of the vehicle, the vehicle motion state information includes at least one of the vehicle's x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear, and the historical lane information includes the longitude, latitude, and elevation of the lane passed by the vehicle.

[0021] In a second aspect, an embodiment of the present invention provides an adaptive redundancy coefficient model training method. The adaptive redundancy coefficient is applied to lane matching. The method includes:

[0022] Obtain vehicle high-precision positioning information, road curvature, and weather humidity. The vehicle high-precision positioning information, road curvature, and weather humidity are the input feature vectors for training the model;

[0023] Obtain redundancy coefficient data through manual or automatic annotation. The redundancy coefficient is the output feature vector of the training model;

[0024] Train a second model according to the vehicle high-precision positioning information, road curvature, weather humidity, redundancy coefficient data, second loss function, and first model. The second model is a shallow neural network model. The second model is used to obtain the adaptive redundancy coefficient of the vehicle matching lane. The second model is trained using the backpropagation algorithm. The second loss function is the cross-entropy loss function.

[0025] Optionally, the formula of the second model:

[0026] R = σ(W r N + b r )

[0027] Among them, N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, b r is the offset of the shallow neural network.

[0028] Optionally, the formula of the second loss function:

[0029] CrossEntropy(match(PointList,N0,N1|R),GroundTruth)

[0030] Among them, PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle dimension, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0031] In a third aspect, an embodiment of the present invention provides an algorithm for matching lanes based on a small amount of high-precision map data. The method includes;

[0032] Obtain the high-precision positioning information of the target vehicle and at least a part of the MAP information received by the target vehicle, where the MAP information corresponds to the high-precision map;

[0033] Match the target lane of the target vehicle according to the high-precision positioning information of the target vehicle, the MAP information, the first model, and the adaptive redundancy coefficient. The first model is trained by the autonomous prediction lane information model of any one of the above, and the adaptive redundancy coefficient is trained by the adaptive redundancy coefficient model of any one of the above.

[0034] In a fourth aspect, an embodiment of the present invention provides an apparatus for training an autonomous prediction lane information model. The apparatus includes;

[0035] The first acquisition module is used to acquire historical vehicle state information and historical lane information. The historical vehicle state information is the input data for training the model, and the historical lane information is the output data for training the model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical high-precision positioning information of the vehicle and historical vehicle movement information, and the historical lane information is the lane position information passed by the vehicle;

[0036] The first training module is used to train the first model according to the historical vehicle state information, the historical lane information, and the first loss function. The first model is a long short-term memory recurrent neural network model. The first model is used to autonomously predict the lane position information that the vehicle will pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0037] Optionally, the formula of the first model:

[0038] f t =σ g (W f x t +U f h t-1 +b f )

[0039] it = σ g (W i x t + U i h t-1 + b i )

[0040] o t = σ g (W o x t + U o h t-1 + b o )

[0041]

[0042]

[0043]

[0044] Among them, x t is the input feature vector, h t is the output vector, f t is the excitation vector for controlling the forgetting gate, i t is the excitation vector for controlling the input gate, o t is the excitation vector for controlling the output gate, is the cell input excitation vector, c t is the cell state vector, W and U are weights, b is the offset, and σ is the activation function.

[0045] Optionally, the vehicle high-precision positioning information includes the longitude, latitude, and altitude of the vehicle, the vehicle motion state information includes at least one of the vehicle's x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear, and the historical lane information includes the longitude, latitude, and altitude of the lanes passed by the vehicle.

[0046] In a fifth aspect, an embodiment of the present invention provides an adaptive redundancy coefficient model training device, the device includes;

[0047] A second acquisition module, configured to acquire vehicle high-precision positioning information, road curvature, and weather humidity, and the vehicle high-precision positioning information, road curvature, and weather humidity are input feature vectors for training the model;

[0048] A third acquisition module, configured to obtain redundancy coefficient data through manual or automatic annotation, and the redundancy coefficient is an output feature vector of the training model;

[0049] The second training module is used to train a second model according to the vehicle's high-precision positioning information, road curvature, weather humidity, redundancy coefficient data, the second loss function, and the first model. The second model is a shallow neural network model and is used to obtain the adaptive redundancy coefficient of the vehicle's matching lane. The second model is trained using the backpropagation algorithm, and the second loss function is the cross-entropy loss function.

[0050] Optionally, the formula of the second model:

[0051] R = σ(W r N + b r )

[0052] where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, and b r is the offset of the shallow neural network.

[0053] Optionally, the formula of the second loss function:

[0054] CrossEntropy(match(PointList, N0, N1|R), GroundTruth)

[0055] where PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle latitude, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0056] In a sixth aspect, an embodiment of the present invention provides a device for matching lanes based on a small amount of high-precision map data. The device includes:

[0057] A matching information acquisition module, configured to acquire the high-precision positioning information of the target vehicle and at least a part of the MAP information received by the target vehicle, where the MAP information corresponds to the high-precision map;

[0058] A matching module, configured to match the target lane of the target vehicle according to the high-precision positioning information of the target vehicle, the MAP information, the first model, and the adaptive redundancy coefficient. The first model is trained by the autonomous prediction lane information model of any of the above, and the adaptive redundancy coefficient is trained by the adaptive redundancy coefficient model of any of the above.

[0059] As can be seen from the above, the embodiments of the present invention provide a method and device for training an autonomous prediction lane information model, which obtain historical vehicle state information and historical lane information. The historical vehicle state information is the input data for training the model, and the historical lane information is the output data for training the model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is the lane position information passed by the vehicle. According to the historical vehicle state information, the historical lane information, and the first loss function, the first model is trained. The first model is a long short-term memory recurrent neural network model, which is used to autonomously predict the lane position information that the vehicle is about to pass. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0060] Applying the embodiments of the present invention solves the problem of lane mismatch or mis-match caused by missing lane information and data errors. Of course, implementing any product or method of the present invention does not necessarily require achieving all the above-mentioned advantages at the same time.

[0061] The innovative points of the embodiments of the present invention include:

[0062] 1. When the MAP information of a certain section is missing because the vehicle cannot receive it in real time, the existing solution is to use the historical MAP cache data of this section to complete the MAP information of this section. This method cannot solve the situation where the vehicle passes through this section for the first time, that is, the vehicle does not store the historical MAP cache information of this section. The embodiments of the present invention use historical vehicle state information and historical lane information to autonomously predict the lane position information that is about to pass, and convert the problem of completing the missing lane point sequence into a sequence prediction problem, which solves the problem that the vehicle cannot obtain the position information of this section when passing through a certain section for the first time, and also solves the problem of lane mismatch caused by missing lane information.

[0063] 2. When a vehicle matches a lane, it calculates which lane its own vehicle is in according to the vehicle's high-precision positioning information and high-precision map information. Due to various reasons, there will be errors between the data such as the vehicle's high-precision positioning information and high-precision map information obtained and their true values. The existing common method for solving problems caused by data errors is to introduce a redundant threshold range, but a fixed redundant threshold can only meet the correction of some actual situations. The embodiments of the present invention realize an adaptive redundancy coefficient through a modeling method, and adjust the redundancy coefficient according to the output result of the autonomous prediction lane information model, so that the redundancy coefficient can dynamically adaptively change according to the actual road conditions, and solves the problem of lane mis-match caused by incorrect redundant coefficients. Brief Description of the Drawings

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0065] Figure 1 A flowchart of a method for training an autonomous prediction lane information model provided by an embodiment of the present invention;

[0066] Figure 2 A flowchart of a method for training an adaptive redundancy coefficient model provided by an embodiment of the present invention;

[0067] Figure 3 A flowchart of an algorithm for matching lanes based on a small amount of high-precision map data provided by an embodiment of the present invention;

[0068] Figure 4 A structural diagram of an apparatus for training an autonomous prediction lane information model provided by an embodiment of the present invention;

[0069] Figure 5 A structural diagram of an apparatus for training an adaptive redundancy coefficient model provided by an embodiment of the present invention. Detailed implementation manners

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0071] It should be noted that the terms "include" and "have" and any variations thereof in the embodiments of the present invention and the accompanying drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally further include steps or units not listed, or may optionally further include other steps or units inherent to these processes, methods, products, or devices.

[0072] The present invention provides a method and an apparatus for training an autonomous prediction lane information model. The following will detail the embodiments of the present invention.

[0073] Figure 1 A flowchart of a method for training an autonomous prediction lane information model provided by an embodiment of the present invention. The method may include the following steps:

[0074] S101: Obtain historical vehicle status information and historical lane information. The historical vehicle status information is the input data for training the model, and the historical lane information is the output data for training the model. The historical lane information corresponds to the historical vehicle status information. The historical vehicle status information includes historical vehicle high-precision positioning information and historical vehicle movement information, and the historical lane information is the lane position information passed by the vehicle.

[0075] The vehicle high-precision positioning information is obtained based on the differential positioning algorithm model of ground-based enhanced RTK (Real-time kinematic). Vehicle high-precision positioning includes, but is not limited to, GPS differential positioning and Beidou differential positioning. Obtain the historical vehicle status information and historical lane information, and use the historical vehicle status information and historical lane information as the training set to train the first model mentioned in step S102. In the training set, the historical vehicle status information is the input data, and the historical lane information is the output data.

[0076] Vehicles receive lane data, usually through roadside RSU. That is, the RSU will send MAP information to the on-vehicle intelligent terminal OBU in real time. Affected by many factors, vehicles will lose some data during the data reception process. The common method to solve the problem of missing lane information is to cache historical MAP data for use when it cannot be received next time. However, this method cannot solve the situation where there is no historical MAP cache information when the vehicle first passes a certain section. The embodiment of the present invention can calculate and output the optimal predicted lane point information of the current section in real time through the historical data of the previous section and the historical operation status data of the vehicle itself in the previous section, so as to solve the problem of missing lane information. The historical operation status data of the vehicle itself in the previous section includes, but is not limited to, the high-precision positioning information, vehicle speed, yaw angle, etc. of the vehicle itself.

[0077] In an alternative embodiment, the vehicle high-precision positioning information includes the longitude, latitude, and elevation of the vehicle, and the vehicle movement state information includes at least one of the x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear of the vehicle. The historical lane information includes the longitude, latitude, and elevation of the lane passed by the vehicle.

[0078] The Long Short-Term Memory Network (LSTM) is a deep learning method and a type of recurrent neural network developed by adding support for long-term dependencies to optimize the general Recurrent Neural Network (RNN) model. LSRM mainly processes complete sequential data rather than the sequence at a single moment, such as predicting a complete sequence of device states. In an embodiment of the present invention, LSTM is used to fuse the historical data of the previous road section and the historical data of the running state of the previous vehicle itself to predict the lane position information that the vehicle is about to pass through.

[0079] In an implementable manner, the input feature vector and the output feature vector of the LSTM are constructed. Among them, the longitude of the vehicle's GPS positioning can be selected as x t0 , the latitude of the GPS positioning is x t1 , the altitude of the GPS positioning is x t2 , the velocity along the x-axis is x t3 , the velocity along the y-axis is x t4 , the velocity along the z-axis is x t5 , the acceleration along the x-axis is x t6 , the acceleration along the y-axis is x t7 , the acceleration along the z-axis is x t8 , the yaw angular velocity is x t9 , the opening degree of the throttle pedal is x t10 , the opening degree of the brake pedal is x t11 , the gear position is x t12 , and the input feature vector x t corresponding to the data of the t-th frame is constructed as = [x t0 , x t1 , x t2 , x t3 , … x t11 , x t12 . The longitude of the lane point track information of the current frame can be taken as h t0 , the latitude is h t1 , and the altitude is h t2 , and the output vector h t corresponding to the data of the t-th frame is constructed.

[0080] S102: According to the historical vehicle state information, the historical lane information, and the first loss function, train the first model. The first model is a long short-term memory recurrent neural network model, which is used to autonomously predict the lane position information that the vehicle is about to pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0081] Taking historical vehicle state information as input data and historical lane information as output data, train the first model, where the first model is an STM model. Since the output data is lane point track information, that is, continuous geographic coordinate data, the embodiments of the present invention use the mean squared error as the loss function for training the LSTM model.

[0082] In an alternative embodiment, the formula of the first model is:

[0083] f t =σ g (W f x t +U f h t-1 +b f )

[0084] i t =σ g (W i x t +U i h t-1 +b i )

[0085] o t =σ g (W o x t +U o h t-1 +b o )

[0086]

[0087]

[0088]

[0089] Among them, x t is the input feature vector, h t is the output vector, f t is the activation vector for controlling the forget gate, i t is the activation vector for controlling the input gate, o t is the activation vector for controlling the output gate, is the cell input activation vector, c t is the cell state vector, W and U are weights, b is the offset, and σ is the activation function.

[0090] In the above formula, σ g is the sigmoid function, σ c and σ h are both hyperbolic tangent functions, h t-1Lane position data of the vehicle passing through in the previous time period.

[0091] In an implementable manner, using a training set and a loss function, applying the BPTT (back-propagation through time) algorithm, training the first model on a high-performance supercomputer server to obtain the optimal weights and offset parameters of the W, U, and b parameters in the first model, and a validation set can also be used during the training process for overfitting verification to verify the first model.

[0092] In addition, a part of the historical vehicle state data and historical lane data can be divided as a test set. After overfitting verification, the test set is used to test the first model with the trained W, U, and b parameters.

[0093] Figure 2 It is a schematic flowchart of a process for training an adaptive redundancy coefficient model provided by an embodiment of the present invention. The method may include the following steps:

[0094] S201: Obtain vehicle high-precision positioning information, road curvature, and weather humidity. The vehicle high-precision positioning information, the road curvature, and the weather humidity are input feature vectors for training the model.

[0095] When the vehicle matches the lane, due to reasons such as vehicle high-precision positioning data errors and high-precision map original data errors, it is possible that the vehicle and the lane are mismatched. The common method to solve data errors in the prior art is to introduce a redundancy threshold range, which is set according to experience to enlarge or reduce the judgment threshold. The method of setting according to experience, for example, directly performs manual scaling based on conditions such as road curvature and urban road grade. However, the redundancy threshold range does not exist in isolation, and its effect will change due to factors such as actual road curvature, GPS positioning accuracy, and weather. The judgment error of the redundancy threshold for a linear relationship is very large, and it can only meet the correction of some actual situations. The embodiment of the present invention realizes an adaptive dynamic redundancy coefficient based on a shallow neural network, generates a dynamic redundancy coefficient according to the actual road conditions, and can eliminate the lane mis-matching problem to a certain extent.

[0096] In an implementable manner, to construct the input feature vector of the second model mentioned in step S203, the vehicle GPS positioning longitude can be selected as N 0 , the GPS positioning latitude as N 1 , the GPS positioning altitude as N 2 , the road curvature as N 3 , and the weather humidity as N 4 , and construct the second model input feature vector N = [N 0 , N 1 , N 2 , N4 .

[0097] S202: Obtain redundancy coefficient data through manual or automatic annotation. The redundancy coefficient is the output feature vector of the training model.

[0098] The manual method can be to manually annotate the redundancy coefficient required for the correct matching of the actual road. The automatic annotation method can be to automatically annotate the data of the redundancy coefficient according to the correspondence between the MAP of the matched lane information and the position information of the vehicle itself.

[0099] S203: Train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model. The second model is a shallow neural network model, which is used to obtain the adaptive redundancy coefficient of the vehicle matching lane. The second model is trained using the backpropagation algorithm, and the second loss function is the cross-entropy loss function.

[0100] Use the obtained vehicle high-precision positioning information, road curvature, weather humidity, redundancy coefficient and other data as the training set to train the second model, and the second model is a shallow neural network model. In an implementable way, use the training set and the loss function, apply the BP (back-propagation) algorithm, and perform model training on the shallow neural network on a high-performance supercomputer server to obtain the optimal weights and offset parameters of the parameters in the second model.

[0101] In addition, the obtained vehicle high-precision positioning information, road curvature, weather humidity, redundancy coefficient and other data can also be divided into three parts. In addition to using a part as the training set to train the second model, the remaining two parts are used as the overfitting verification model and the test model respectively to obtain a better second model.

[0102] In an alternative embodiment, the formula of the second model:

[0103] R = σ(W r N + b r )

[0104] where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, and b r is the offset of the shallow neural network.

[0105] In an alternative embodiment, the formula of the second loss function:

[0106] CrossEntropy(match(PointList,N0,N1|R),GroundTruth)

[0107] Among them, PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle dimension, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0108] PointList is the lane position information predicted by the vehicle, which may include information such as the longitude, dimension, and elevation of the upcoming lane. The redundancy coefficient is adjusted according to the output data result of the first model. The lane matching algorithm is a prior art. In one case, the lane matching algorithm calculates which lane the vehicle will be in by using the vehicle's GPS position information and the high-precision map data structure information obtained by the vehicle. The meaning of match(PointList,N0,N1|R) is that when the parameter value in the match function is equal to R, the current vehicle longitude N0, latitude N1, and the lane point sequence PointList are matched to calculate whether the match is successful.

[0109] Figure 3 It is a schematic flowchart of an algorithm for matching lanes based on a small amount of high-precision map data provided by an embodiment of the present invention. The algorithm may include the following steps:

[0110] S301: Obtain the high-precision positioning information of the target vehicle and at least a part of the MAP information received by the target vehicle, where the MAP information corresponds to the high-precision map.

[0111] The MAP information is sent to the vehicle by the RSU. During the transmission of the MAP information, it is affected by factors such as the RSU deployment location, the environment of the wireless signal, the influence of multipath reflection, and obstacle occlusion. Sometimes the vehicle cannot receive the complete MAP information. When the MAP information is partially missing, the embodiment of the present invention can be applied to complete the road position information. Therefore, the embodiment of the present invention only needs to obtain at least a part of the MAP information to match the lane. That is to say, when the MAP information received by the vehicle is complete, the high-precision positioning information of the vehicle and the complete real-time MAP information are used to calculate and match the lane. When the MAP information is missing, the historical high-precision positioning information of the vehicle and the historical lane position information are used to predict the upcoming lane position information to be matched, which not only completes the lane information but also avoids the problem in the prior art that if the MAP information is missing, the lane cannot be matched for the first time passing through a section.

[0112] S302: Match the target lane of the target vehicle according to the high-precision positioning information of the target vehicle, the MAP information, the first model, and the adaptive redundancy coefficient.

[0113] Currently, the redundancy threshold range for lane matching is fixed. The adaptive redundancy coefficient in the present invention can adjust the redundancy coefficient in real time according to the real-time road conditions, solving the problem of incorrect lane matching caused by incorrect redundancy coefficients.

[0114] The training method steps of the first model are as follows:

[0115] Obtain historical vehicle state information and historical lane information. The historical vehicle state information is the input data for training the model, and the historical lane information is the output data for training the model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle movement information, and the historical lane information is the lane position information passed by the vehicle.

[0116] Train the first model according to the historical vehicle state information, the historical lane information, and the first loss function. The first model is a long short-term memory recurrent neural network model. The first model is used to autonomously predict the lane position information that the vehicle will pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0117] In an alternative manner, the formula of the first model:

[0118] f t =σ g (W f x t +U f h t-1 +b f )

[0119] i t =σ g (W i x t +U i h t-1 +b i )

[0120] o t =σ g (W o x t +U o h t-1 +b o )

[0121]

[0122]

[0123]

[0124] Among them, x t is the input feature vector, h t is the output vector, f t is the excitation vector for controlling the forget gate, i t is the excitation vector for controlling the input gate, o t is the excitation vector for controlling the output gate, is the cell input excitation vector, c t is the cell state vector, W and U are weights, b is the offset, and σ is the activation function.

[0125] In an alternative manner, the vehicle high-precision positioning information includes the longitude, latitude, and elevation of the vehicle, the vehicle motion state information includes at least one of the vehicle's x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear, and the historical lane information includes the longitude, latitude, and elevation of the lanes passed by the vehicle.

[0126] The training method steps of the adaptive redundancy coefficient model are as follows:

[0127] Obtain the vehicle high-precision positioning information, road curvature, and weather humidity. The vehicle high-precision positioning information, the road curvature, and the weather humidity are the input feature vectors for training the model;

[0128] Obtain the redundancy coefficient data through manual or automatic annotation. The redundancy coefficient is the output feature vector of the training model;

[0129] Train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model. The second model is a shallow neural network model. The second model is used to obtain the adaptive redundancy coefficient of the vehicle matching lane. The second model is trained using the backpropagation algorithm. The second loss function is the cross-entropy loss function.

[0130] In an alternative manner, the formula of the second model:

[0131] R = σ(W r N + b r )

[0132] Among them, N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, b r is the offset of the shallow neural network.

[0133] In an alternative approach, the formula for the second loss function is:

[0134] CrossEntropy(match(PointList,N0,N1|R),GroundTruth)

[0135] where PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle dimension, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0136] Figure 4 FIG. 11 is a schematic structural diagram of a device for training an autonomous prediction lane information model provided by an embodiment of the present invention. An embodiment of the present invention provides a device for training an autonomous prediction lane information model, the device comprising:

[0137] A first acquisition module S401, configured to acquire historical vehicle state information and historical lane information, where the historical vehicle state information is input data for training the model, the historical lane information is output data for training the model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle movement information, and the historical lane information is lane position information passed by the vehicle;

[0138] A first training module S401, configured to train a first model according to the historical vehicle state information, the historical lane information, and a first loss function, where the first model is a long short-term memory recurrent neural network model, the first model is used to autonomously predict lane position information that the vehicle is about to pass through, the first model is trained using the backpropagation through time algorithm, and the first loss function is a mean squared error loss function.

[0139] In an alternative embodiment, the formula for the first model is:

[0140] f t =σ g (W f x t +U f h t-1 +b f )

[0141] i t =σ g (W i x t +U i h t-1 +b i )

[0142] ot = σ g (W o x t + U o h t-1 + b o )

[0143]

[0144]

[0145]

[0146] where x t is the input feature vector, h t is the output vector, f t is the activation vector for controlling the forget gate, i t is the activation vector for controlling the input gate, o t is the activation vector for controlling the output gate, is the cell input activation vector, c t is the cell state vector, W and U are weights, b is the bias, and σ is the activation function.

[0147] In an optional embodiment, the vehicle high-precision positioning information includes the longitude, latitude, and altitude of the vehicle, the vehicle motion state information includes at least one of the vehicle's x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear, and the historical lane information includes the longitude, latitude, and altitude of the lane passed by the vehicle.

[0148] Figure 5 This is a schematic structural diagram of an adaptive redundancy coefficient model training device provided by an embodiment of the present invention. An embodiment of the present invention also provides an adaptive redundancy coefficient model training device, where the adaptive redundancy coefficient is applied to lane matching, and the device includes:

[0149] A second acquisition module 501, configured to acquire vehicle high-precision positioning information, road curvature, and weather humidity, where the vehicle high-precision positioning information, the road curvature, and the weather humidity are input feature vectors for training the model;

[0150] A third acquisition module 502, configured to obtain redundancy coefficient data through manual or automatic annotation, where the redundancy coefficient is an output feature vector of the training model;

[0151] The second training module 503 is used to train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model. The second model is a shallow neural network model, and the second model is used to obtain the adaptive redundancy coefficient of the vehicle matching lane. The second model is trained using the backpropagation algorithm, and the second loss function is the cross-entropy loss function.

[0152] In an optional embodiment, the formula of the second model is:

[0153] R = σ(W r N + b r )

[0154] where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, and b r is the offset of the shallow neural network.

[0155] In an optional embodiment, the formula of the second loss function is:

[0156] CrossEntropy(match(PointList, N0, N1|R), GroundTruth)

[0157] where PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle dimension, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0158] Corresponding to the above lane matching algorithm, an embodiment of the present invention provides a device for matching lanes based on a small amount of high-precision map data. The device includes:

[0159] A matching information acquisition module, configured to acquire the high-precision positioning information of the target vehicle and at least a part of the MAP information received by the target vehicle, where the MAP information corresponds to the high-precision map;

[0160] A matching module, configured to match the target lane of the target vehicle according to the high-precision positioning information of the target vehicle, the MAP information, the first model, and the adaptive redundancy coefficient. The first model is trained through the autonomous prediction lane information model described above, and the adaptive redundancy coefficient is trained through the adaptive redundancy coefficient model described above.

[0161] The training device of the first model includes:

[0162] The first acquisition module S401 is configured to acquire historical vehicle state information and historical lane information. The historical vehicle state information is the input data for training the model, and the historical lane information is the output data of the training model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle movement information, and the historical lane information is the lane position information passed by the vehicle.

[0163] The first training module S402 is configured to train the first model according to the historical vehicle state information, the historical lane information, and the first loss function. The first model is a long short-term memory recurrent neural network model, and the first model is used to autonomously predict the lane position information that the vehicle is about to pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

[0164] In an optional manner, the formula of the first model is:

[0165] f t = σ g (W f x t + U f h t-1 + b f )

[0166] i t = σ g (W i x t + U i h t-1 + b i )

[0167] o t = σ g (W o x t + U o h t-1 + b o )

[0168]

[0169]

[0170]

[0171] Among them, x t is the input feature vector, h t is the output vector, f t is the activation vector for controlling the forget gate, i t is the activation vector for controlling the input gate, ot is the excitation vector for controlling the output gate, is the cell input excitation vector, c t is the cell state vector, W and U are weights, b is the offset, and σ is the activation function.

[0172] In an alternative embodiment, the vehicle high-precision positioning information includes the longitude, latitude, and altitude of the vehicle, the vehicle motion state information includes at least one of the x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, throttle pedal opening, brake pedal opening, and gear of the vehicle, and the historical lane information includes the longitude, latitude, and altitude of the lanes passed by the vehicle.

[0173] The training device for the adaptive redundancy coefficient model includes:

[0174] A second acquisition module S501, configured to acquire vehicle high-precision positioning information, road curvature, and weather humidity, where the vehicle high-precision positioning information, the road curvature, and the weather humidity are input feature vectors for training the model;

[0175] A third acquisition module S502, configured to obtain redundancy coefficient data through manual or automatic annotation, where the redundancy coefficient is an output feature vector of the training model;

[0176] A second training module S503, configured to train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model, where the second model is a shallow neural network model, the second model is used to obtain the adaptive redundancy coefficient of the vehicle matching lane, the second model is trained using the backpropagation algorithm, and the second loss function is a cross-entropy loss function.

[0177] In an alternative embodiment, the formula of the second model:

[0178] R = σ(W r N + b r )

[0179] where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, W r is the weight of the shallow neural network, and b r is the offset of the shallow neural network.

[0180] In an alternative embodiment, the formula of the second loss function:

[0181] CrossEntropy(match(PointList,N0,N1|R),GroundTruth)

[0182] Among them, PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle dimension, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information.

[0183] The above system and device embodiments correspond to the system embodiment and have the same technical effects as the method embodiment. For specific descriptions, please refer to the method embodiment. The device embodiment is obtained based on the method embodiment. For specific descriptions, please refer to the method embodiment section and will not be elaborated here. Those of ordinary skill in the art can understand that the drawings are only schematic diagrams of one embodiment, and the modules or processes in the drawings are not necessarily essential for implementing the present invention.

[0184] Those of ordinary skill in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be correspondingly changed and located in one or more devices different from this embodiment. The modules in the above embodiment can be combined into one module, or further split into multiple sub-modules.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training an adaptive redundancy coefficient model, where the adaptive redundancy coefficient is applied to lane matching, Characterized in that, The method includes: Obtain vehicle high-precision positioning information, road curvature, and weather humidity. The vehicle high-precision positioning information, the road curvature, and the weather humidity are input feature vectors for training the second model; Obtain redundancy coefficient data through manual or automatic annotation. The redundancy coefficient is the output feature vector for training the second model; According to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model, train the second model. The first model is used to autonomously predict the lane position information that the vehicle is about to pass through. The second model is a shallow neural network model. The second model is used to obtain the adaptive redundancy coefficient for the vehicle to match the lane. The second model is trained using the backpropagation algorithm. The second loss function is the cross-entropy loss function; The formula of the second loss function: CrossEntropy(match(PointList, N0, N1|R), GroundTruth) Where PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle latitude, R is the adaptive redundancy coefficient, and GroundTruth is the actual road information. The lane matching algorithm is a lane matching algorithm that calculates which lane the vehicle is about to be in through the vehicle GPS position information and the high-precision map data structure information obtained by the vehicle. And match(PointList, N0, N1|R) means that when the parameter value in the match function is equal to R, the vehicle longitude N0, vehicle latitude N1, and the first model output data PointList are matched.

2. The method according to claim 1, Characterized in that, The formula of the second model: ； Among them, N is the input feature vector of the second model, R is the adaptive redundancy coefficient, is the ReLU activation function, W r is the weight of the shallow neural network, b r is the offset of the shallow neural network.

3. A method for matching lanes based on a small amount of high-precision map data, Characterized in that, The method includes: Obtain the high-precision positioning information of the target vehicle and at least a part of the MAP information received by the target vehicle. The MAP information corresponds to the high-precision map; According to the high-precision positioning information of the target vehicle, the MAP information, the first model, and the adaptive redundancy coefficient, match the target lane of the target vehicle. The adaptive redundancy coefficient is obtained by training through the adaptive redundancy coefficient model described in any one of claims 1 to 2; The training method of the first model includes: Obtain historical vehicle state information and historical lane information. The historical vehicle state information is the input data for training the first model, and the historical lane information is the output data for training the first model. The historical lane information corresponds to the historical vehicle state information. The historical vehicle state information includes historical vehicle high-precision positioning information and historical vehicle movement information. The historical lane information is the lane position information passed by the vehicle; Train the first model according to the historical vehicle state information, the historical lane information, and the first loss function. The first model is a long short-term memory recurrent neural network model, which is used to autonomously predict the lane position information that the vehicle will pass through. The first model is trained using the backpropagation through time algorithm, and the first loss function is the mean squared error loss function.

4. The method according to claim 3, wherein, the formula of the first model: ； ； ； ； ； ； where x t is the input feature vector, h t is the output vector, f t is the activation vector for controlling the forget gate, i t is the activation vector for controlling the input gate, o t is the activation vector for controlling the output gate, is the cell input activation vector, is the cell state vector, W and U are weights, b is the bias, is the activation function.

5. The method according to claim 3, wherein, the vehicle high-precision positioning information includes the longitude, latitude, and elevation of the vehicle, and the vehicle motion state information includes at least one of the vehicle's x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw angular velocity, throttle pedal opening, brake pedal opening, and gear. The historical lane information includes the longitude, latitude, and elevation of the lane passed by the vehicle.

6. An adaptive redundancy coefficient model training device, where the adaptive redundancy coefficient is applied to lane matching, wherein, the device includes: A second acquisition module, configured to acquire vehicle high-precision positioning information, road curvature, and weather humidity. The vehicle high-precision positioning information, the road curvature, and the weather humidity are input feature vectors for training the second model; A third acquisition module, configured to obtain redundancy coefficient data through manual or automatic annotation. The redundancy coefficient is an output feature vector for training the second model; A second training module, configured to train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, the second loss function, and the first model. The first model is used to autonomously predict the lane position information that the vehicle will pass through. The second model is a shallow neural network model, and the second model is used to obtain the adaptive redundancy coefficient of the vehicle matching lane. The second model is trained using the backpropagation algorithm, and the second loss function is the cross-entropy loss function; The formula of the second loss function: CrossEntropy(match(PointList, N0, N1|R), GroundTruth) where PointList is the output data of the first model, match() is the lane matching algorithm, N0 is the vehicle longitude, N1 is the vehicle latitude, R is the adaptive redundancy coefficient, GroundTruth is the actual road information, and the lane matching algorithm is a lane matching algorithm that calculates which lane the vehicle will be in through the vehicle GPS position information and the high-precision map data structure information obtained by the vehicle. And match(PointList, N0, N1|R) means that when the parameter value in the match function is equal to R, the vehicle longitude N0, vehicle latitude N1, and the first model output data PointList are matched.

Images