CN114169463A - Autonomous prediction lane information model training method and device - Google Patents
Autonomous prediction lane information model training method and device Download PDFInfo
- Publication number
- CN114169463A CN114169463A CN202111555631.6A CN202111555631A CN114169463A CN 114169463 A CN114169463 A CN 114169463A CN 202111555631 A CN202111555631 A CN 202111555631A CN 114169463 A CN114169463 A CN 114169463A
- Authority
- CN
- China
- Prior art keywords
- model
- information
- vehicle
- lane
- historical
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/29—Geographical information databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The embodiment of the invention discloses an autonomous prediction lane information model training method and an autonomous prediction lane information model training device, wherein the method comprises the following steps: acquiring historical vehicle state information and historical lane information, wherein the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where vehicles pass; training a first model according to historical vehicle state information, historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model and is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a back propagation algorithm along with time, and the first loss function is a minimum mean square error loss function. The method and the device solve the problem that the lanes cannot be matched due to the loss of lane information.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to an autonomous prediction lane information model training method and device.
Background
In an automatic driving or lane-changing cooperative system, many application models of vehicle ends, such as a vehicle acceleration and deceleration control model, a vehicle lane-changing model and the like, require an intelligent vehicle or an intelligent vehicle terminal to determine a lane where the vehicle is located so that the real-time lane position of the vehicle is used as a starting point of a subsequent model.
The basic idea of vehicle lane matching is to match the high-precision positioning of the vehicle with the high-precision MAP information of the vehicle, because the high-precision MAP data is much larger than the data volume of the existing MAP data and the traveling position of the vehicle has great uncertainty, the high-precision MAP data of the vehicle generally does not directly exist in the local area of the vehicle, but the MAP information is issued through a Road Side RSU (Road Side Unit) in cooperation with the Road, so that the related high-precision MAP data information is transmitted to the end of the vehicle which is driven automatically. Due to the deployment position of the RSU, the influence of the environment, multipath reflection On wireless signals, barrier shielding and the like, the On Board Unit (OBU) cannot continuously receive the MAP information from the RSU in real time, so that the judgment of the lane information of the vehicle in the automatic driving model is influenced, and the accuracy of the output data of a series of subsequent automatic driving vehicle-end application models is further influenced.
In addition, even if the on-board intelligent terminal OBU successfully receives complete lane data and self high-precision positioning data sent by the RSU, the lane matching error problem still exists to a certain extent due to self high-precision positioning data errors, high-precision map original data errors and other reasons.
The conventional method for solving the problem of lane information loss is to cache MAP data received in a historical vehicle-road cooperative system so as to be used when the MAP data cannot be received next time, but the method is not suitable for a road section which a vehicle passes through for the first time.
The conventional common method for solving the problem caused by data error is to introduce a redundant threshold range, namely, the calculated distance between a vehicle and a lane central line is used as a judgment threshold, and the judgment threshold is set to be enlarged or reduced according to experience so as to realize that a matching result is consistent with a real situation. However, due to the complexity of the actual road, the judgment error of the redundant threshold of the linear relation is large.
How to complement the lane information of the road section where the vehicle is located and eliminate the error of the redundant threshold value in lane matching is a problem to be solved in the field of automatic driving.
Disclosure of Invention
The invention provides an autonomous prediction lane information model training method and device, and solves the problem that lanes cannot be matched due to lane information loss. The specific technical scheme is as follows:
in a first aspect, an embodiment of the present invention provides an autonomous predictive lane information model training method, where the method includes:
acquiring historical vehicle state information and historical lane information, wherein the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where vehicles pass;
training a first model according to historical vehicle state information, historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model and is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a back propagation algorithm along with time, and the first loss function is a minimum mean square error loss function.
Optionally, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
Optionally, the vehicle high-precision positioning information includes longitude, latitude and elevation of the vehicle, the vehicle motion state information includes at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear of the vehicle, and the historical lane information includes longitude, latitude and elevation of a lane through which the vehicle passes.
In a second aspect, an embodiment of the present invention provides a method for training a self-adaptive redundancy coefficient model, where a self-adaptive redundancy coefficient is applied to lane matching, and the method includes:
acquiring vehicle high-precision positioning information, road curvature and weather humidity, wherein the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of a training model;
obtaining redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output characteristic vector of the training model;
training a second model according to the high-precision positioning information of the vehicle, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, wherein the second model is a shallow neural network model and is used for obtaining the self-adaptive redundancy coefficient of the matched lane of the vehicle, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
Optionally, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
Optionally, the second loss function formula:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
In a third aspect, an embodiment of the present invention provides an algorithm for lane matching based on a small amount of high-precision map data, where the method includes;
acquiring high-precision positioning information of a target vehicle and at least a part of MAP information received by the target vehicle, wherein the MAP information corresponds to a high-precision MAP;
and matching the target lane of the target vehicle according to the high-precision positioning information, the MAP information, the first model and the adaptive redundancy coefficient of the target vehicle, wherein the first model is obtained by training the autonomous prediction lane information model of any item, and the adaptive redundancy coefficient is obtained by training the adaptive redundancy coefficient model of any item.
In a fourth aspect, an embodiment of the present invention provides an autonomous predictive lane information model training apparatus, including;
the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring historical vehicle state information and historical lane information, the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where a vehicle passes;
the first training module is used for training a first model according to historical vehicle state information, historical lane information and a first loss function, the first model is a long-short term memory recurrent neural network model, the first model is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a back propagation algorithm along with time, and the first loss function is a minimum mean square error loss function.
Optionally, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor controlling the output gateThe excitation vector of (a) is calculated,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
Optionally, the vehicle high-precision positioning information includes longitude, latitude and elevation of the vehicle, the vehicle motion state information includes at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear of the vehicle, and the historical lane information includes longitude, latitude and elevation of a lane through which the vehicle passes.
In a fifth aspect, an embodiment of the present invention provides an adaptive redundancy coefficient model training apparatus, including;
the second acquisition module is used for acquiring vehicle high-precision positioning information, road curvature and weather humidity, and the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of the training model;
the third acquisition module is used for acquiring redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output feature vector of the training model;
the second training module is used for training a second model according to the high-precision positioning information of the vehicle, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, the second model is a shallow neural network model and is used for obtaining the self-adaptive redundancy coefficient of the vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
Optionally, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
Optionally, the second loss function formula:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
In a sixth aspect, an embodiment of the present invention provides an apparatus for lane matching based on a small amount of high-precision map data, the apparatus including;
the matching information acquisition module is used for acquiring high-precision positioning information of the target vehicle and at least part of MAP information received by the target vehicle, wherein the MAP information corresponds to the high-precision MAP;
and the matching module is used for matching 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, wherein the first model is obtained by training any one of the self-prediction lane information models, and the adaptive redundancy coefficient is obtained by training any one of the self-adaptive redundancy coefficient models.
As can be seen from the above, the embodiment of the present invention provides an autonomous prediction lane information model training method and apparatus, acquiring historical vehicle state information and historical lane information, where the historical vehicle state information is input data of a training model, the historical lane information is 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 motion information, and the historical lane information is lane position information where a vehicle passes through; training a first model according to historical vehicle state information, historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model and is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a back propagation algorithm along with time, and the first loss function is a minimum mean square error loss function.
By applying the embodiment of the invention, the problem that the lanes cannot be matched or are matched wrongly caused by lane information loss and data errors is solved. Of course, not all of the advantages described above need to be achieved at the same time in the practice of any one product or method of the invention.
The innovation points of the embodiment of the invention comprise:
1. when MAP information of a road section is lost due to the fact that a vehicle cannot receive MAP information of the road section in real time, the MAP information of the road section is supplemented by historical MAP cache data of the road section, and the method cannot solve the problem that the vehicle firstly passes through the road section, namely the vehicle does not store the historical MAP cache information of the road section. The embodiment of the invention autonomously predicts the position information of the lane to be passed by using the historical vehicle state information and the historical lane information, converts the completion problem of the missing lane point sequence into the sequence prediction problem, solves the problem that the position information of a road section cannot be obtained when the vehicle passes through the road section for the first time, and also solves the problem that the lanes cannot be matched due to the missing lane information.
2. The vehicle matching lane is that a lane on which a vehicle is located is calculated according to vehicle high-precision positioning information and high-precision map information, errors exist between data such as the obtained vehicle high-precision positioning information and the high-precision map information and true values of the data due to various reasons, a common method for solving the problem caused by data errors is to introduce a redundancy threshold range, but a fixed redundancy threshold can only meet the correction of partial actual conditions. According to the embodiment of the invention, the self-adaptive redundancy coefficient is realized in a modeling mode, and the redundancy coefficient is adjusted according to the output result of the autonomous prediction lane information model, so that the redundancy coefficient can be dynamically and self-adaptively changed according to the actual road condition, and the problem of lane mismatching caused by the error of the redundancy coefficient is solved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It is to be understood that the drawings in the following description are merely exemplary of some embodiments of the invention. For a person skilled in the art, without inventive effort, further figures can be obtained from these figures.
Fig. 1 is a schematic flow chart of an autonomous predictive lane information model training method according to an embodiment of the present invention;
FIG. 2 is a schematic flow chart of a method for training an adaptive redundancy coefficient model according to an embodiment of the present invention;
FIG. 3 is a schematic flow chart of an algorithm for lane matching based on a small amount of high-precision map data according to an embodiment of the present invention;
fig. 4 is a schematic structural diagram of an autonomous predictive lane information model training apparatus according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an adaptive redundancy coefficient model training apparatus according to an embodiment of the present invention.
Detailed Description
The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It is to be understood that the described embodiments are merely a few embodiments of the invention, and not all embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive effort based on the embodiments of the present invention, are within the scope of the present invention.
It is to be noted that the terms "comprises" and "comprising" and any variations thereof in the embodiments and drawings of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those steps or elements listed, but may alternatively include other steps or elements not listed, or inherent to such process, method, article, or apparatus.
The invention provides an autonomous prediction lane information model training method and device. The following provides a detailed description of embodiments of the invention.
Fig. 1 is a schematic flow chart of an autonomous predictive lane information model training method according to an embodiment of the present invention. The method may comprise the steps of:
s101: the method comprises the steps of obtaining historical vehicle state information and historical lane information, wherein the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where vehicles pass.
The high-precision positioning information of the vehicle is obtained based on a differential positioning algorithm model of ground-based enhanced RTK (Real-time kinematic) and includes but is not limited to GPS differential positioning and Beidou differential positioning. Historical vehicle state information and historical lane information are acquired, the historical vehicle state information and the historical lane information are used as a training set, the first model mentioned in the step S102 is trained, the historical vehicle state information is input data in the training set, and the historical lane information is output data.
The vehicle receives the lane data, which is generally received through the road-side RSU, that is, the RSU sends MAP information to the on-board intelligent terminal OBU in real time, and due to the influence of many factors, a part of data is lost during the data receiving process of the vehicle. The conventional method for solving the problem of lane information loss is historical MAP data caching so as to be used when the vehicle cannot receive the data next time, but the method cannot solve the problem that the vehicle firstly passes through a certain road section without historical MAP caching information. According to the embodiment of the invention, the optimal predicted lane point information of the current road section can be calculated and output in real time through the historical data of the preorder road section and the historical data of the running state of the preorder vehicle, so that the problem of lane information loss is solved. The running state historical data of the preceding vehicle comprises but is not limited to high-precision positioning information, vehicle speed, heading angle and the like of the vehicle.
In an optional embodiment, the vehicle high-precision positioning information comprises longitude, latitude and elevation of the vehicle, the vehicle motion state information comprises at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear position of the vehicle, and the historical lane information comprises longitude, latitude and elevation of a lane passed by the vehicle.
A Long short-term memory Network (LSTM) is a deep learning method, and a Recurrent Neural Network developed by supporting and optimizing Long-term dependence is added on the basis of a general Neural Network (RNN) model. The LSRM mainly processes complete sequence-type data, not a sequence at a single moment, such as predicting a complete device state sequence. According to the embodiment of the invention, LSTM is utilized to fuse the historical data of the preorder road section and the historical data of the running state of the preorder self-vehicle, and the lane position information about the vehicle to pass is predicted.
In one implementation, the input and output eigenvectors for the LSTM are constructed, where the vehicle GPS location longitude can be chosen to be xt0And the GPS positioning latitude is xt1And the GPS positioning elevation is xt2X-axis velocity of xt3Y-axis velocity of xt4Z-axis velocity xt5X-axis acceleration of xt6Acceleration of y-axis of xt7Z-axis acceleration of xt8Yaw rate xt9The opening degree of an accelerator pedal is xt10The opening degree of the brake pedal is xt11The gear is xt12Constructing the input feature vector x corresponding to the t-th frame datat＝[xt0，xt1，xt2，xt3，…xt11，xt12]The longitude of the current frame lane trace point information can be taken as ht0Latitude of ht1Elevation of ht2Constructing a corresponding output vector h of the t-th frame datat。
S102: training a first model according to the historical vehicle state information, the historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model, the first model is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a time-dependent back propagation algorithm, and the first loss function is a minimum mean square error loss function.
And training a first model by taking the historical vehicle state information as input data and the historical lane information as output data, wherein the first model is an STM model. Because the output data is lane trace point information, i.e. continuous geographical coordinate data, the embodiment of the present invention uses the minimum mean square error as a loss function for the training of the LSTM model.
In an alternative embodiment, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
In the above formula, σgIs sigmoid function, σc、σhAre hyperbolic tangent function，ht-1Lane position data that the vehicle has passed in the last time period.
In an implementation manner, the first model is model-trained on a high-performance supercomputing server by using a training set and a loss function and applying a BPTT (back-propagation time back propagation) algorithm to obtain the optimal weight and offset parameter of the W, U, b parameter in the first model, and the first model can be verified by using a verification set to perform overfitting verification in the training process.
In addition, a part of historical vehicle state data and historical lane data can be divided to serve as a test set, and after overfitting verification is conducted, the first model with the trained W, U, b parameters is tested by the aid of the test set.
Fig. 2 is a schematic flowchart of adaptive redundancy coefficient model training according to an embodiment of the present invention. The method may comprise the steps of:
s201: the method comprises the steps of obtaining vehicle high-precision positioning information, road curvature and weather humidity, wherein the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of a training model.
When the vehicle matches the lane, the vehicle may be mistakenly matched with the lane due to errors of high-precision positioning data of the vehicle, errors of original data of a high-precision map and the like. In the prior art, a common method for solving data errors is to introduce a redundant threshold range, set according to experience to enlarge or reduce the judgment threshold, and directly perform manual scaling according to the experience setting method based on conditions such as road curvature, urban road grade and the like, but the redundant threshold range does not exist independently, the action effect can be changed due to the change of factors such as actual road curvature, GPS positioning accuracy and weather, and the judgment error of the redundant threshold for linear relation is very large, and only can meet the correction of partial actual conditions. The embodiment of the invention realizes the self-adaptive dynamic redundancy coefficient based on the shallow neural network, generates the dynamic redundancy coefficient according to the actual road condition and can eliminate the problem of lane mismatching to a certain extent.
In an implementable manner, the constructing stepThe second model mentioned in step S203 inputs feature vectors, and can select the vehicle GPS positioning longitude to be N0GPS positioning latitude of N1GPS positioning elevation of N2Road curvature of N3Weather humidity of N4Constructing a second model input feature vector N ═ N0,N1,N2,N4]。
S202: and obtaining redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output characteristic vector of the training model.
The manual mode can be that the redundancy coefficient that needs to set up is annotated to the correct matching of actual road by hand, and the automatic standard mode can be that according to MAP with the matched lane information and the corresponding relation of the position information of the self-parking, the automatic marking obtains the data of redundancy coefficient.
S203: training the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, wherein the second model is a shallow neural network model and is used for obtaining a self-adaptive redundancy coefficient of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
And training a second model by taking the acquired data such as high-precision positioning information of the vehicle, road curvature, weather humidity, redundancy coefficients and the like as a training set, wherein the second model is a shallow neural network model. In one implementation, a model training of the shallow neural network is performed on the high-performance supercomputing server using a training set and a loss function, applying a BP (back-propagation) algorithm, to obtain optimal weights and offset parameters for the parameters in the second model.
In addition, the acquired data such as the high-precision positioning information of the vehicle, the road curvature, the weather humidity, the redundancy coefficient and the like can be divided into three parts, one part is used as a training set for training the second model, and the other two parts are used as an overfitting verification model and a testing model respectively, so that a better second model is obtained.
In an alternative embodiment, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
In an alternative embodiment, the second loss function equation:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
The poitlist is lane position information predicted by the vehicle, and may include longitude, latitude, elevation and other information of the lane to be passed, and the redundancy coefficient is adjusted according to the output data result of the first model. The lane matching algorithm is the prior art, and in one case, the lane matching algorithm is a lane matching algorithm for calculating which lane the vehicle is to be in through the GPS position information of the vehicle and the high-precision map data structure information obtained by the vehicle. match (PointList, N0, N1| R) means that when the parameter value in the match function is equal to R, the matching between the current vehicle longitude N0 and latitude N1 and the lane point sequence PointList is calculated to be successful.
Fig. 3 is a schematic flowchart of an algorithm for matching lanes based on a small amount of high-precision map data according to an embodiment of the present invention. The algorithm may include the steps of:
s301: and acquiring high-precision positioning information of the target vehicle and at least a part of MAP information received by the target vehicle, wherein the MAP information corresponds to the high-precision MAP.
The MAP information is sent to the vehicle through the RSU, the vehicle cannot receive complete MAP information sometimes under the influence of the RSU deployment position, the influence of environment and multipath reflection on wireless signals, barrier shielding and other factors in the MAP information transmission process, and when the MAP information is partially lost, the embodiment of the invention can be applied to complete road position information. Therefore, the embodiment of the invention can match the lane only by acquiring at least a part of MAP information. When the MAP information received by the vehicle is complete, the matched lane is calculated by using the high-precision positioning information of the vehicle and the complete real-time MAP information, and when the MAP information is missing, the lane position information to be matched is predicted by using the high-precision positioning information of the historical vehicle and the historical lane position information, so that the lane information is completed, and the problem that the lane cannot be matched when the MAP information is missing and passes through a road section for the first time in the prior art is solved.
S302: and matching 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 self-adaptive redundancy coefficient.
The range of the redundancy threshold value of the current lane matching is fixed, the self-adaptive redundancy coefficient can adjust the redundancy coefficient in real time according to the real-time road condition, and the problem of lane mismatching caused by the error of the redundancy coefficient is solved.
The training method of the first model comprises the following steps:
the method comprises the steps of obtaining historical vehicle state information and historical lane information, wherein the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where vehicles pass.
Training a first model according to the historical vehicle state information, the historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model, the first model is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a time-dependent back propagation algorithm, and the first loss function is a minimum mean square error loss function.
In an alternative, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
In an optional mode, the vehicle high-precision positioning information comprises longitude, latitude and elevation of the vehicle, the vehicle motion state information comprises at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear position of the vehicle, and the historical lane information comprises longitude, latitude and elevation of a lane passed by the vehicle.
The training method of the self-adaptive redundancy coefficient model comprises the following steps:
acquiring vehicle high-precision positioning information, road curvature and weather humidity, wherein the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of a training model;
obtaining redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output characteristic vector of a training model;
training the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, wherein the second model is a shallow neural network model and is used for obtaining a self-adaptive redundancy coefficient of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
In an alternative, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
In an alternative, the second loss function equation:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
Fig. 4 is a schematic structural diagram of an autonomous predictive lane information model training apparatus according to an embodiment of the present invention. The embodiment of the invention provides an autonomous prediction lane information model training device, which comprises:
the first obtaining module S401 is configured to obtain historical vehicle state information and historical lane information, where the historical vehicle state information is input data of a training model, the historical lane information is 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 motion information, and the historical lane information is lane position information where a vehicle passes through;
the first training module S401 is 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 a vehicle will pass through, the first model is trained by using a time-based back propagation algorithm, and the first loss function is a minimum mean square error loss function.
In an alternative embodiment, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor controlling the input gateExcitation vector of otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
In an optional embodiment, the vehicle high-precision positioning information comprises longitude, latitude and elevation of the vehicle, the vehicle motion state information comprises at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear position of the vehicle, and the historical lane information comprises longitude, latitude and elevation of a lane passed by the vehicle.
Fig. 5 is a schematic structural diagram of an adaptive redundancy coefficient model training apparatus according to an embodiment of the present invention. The embodiment of the invention also provides a training device for the adaptive redundancy coefficient model, wherein the adaptive redundancy coefficient is applied to lane matching, and the training device comprises:
a second obtaining module 501, configured to obtain 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 of a training model;
a third obtaining module 502, configured to obtain redundancy coefficient data in a manual or automatic labeling manner, where the redundancy coefficient is an output feature vector of a training model;
the second training module 503 is configured to train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function, and the first model, where the second model is a shallow neural network model, the second model is used to obtain an adaptive redundancy coefficient of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
In an alternative embodiment, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
In an alternative embodiment, the second loss function equation:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
Corresponding to the lane matching algorithm, an embodiment of the present invention provides an apparatus for matching lanes based on a small amount of high-precision map data, the apparatus including:
the matching information acquisition module is used for acquiring high-precision positioning information of the target vehicle and at least part of MAP information received by the target vehicle, wherein the MAP information corresponds to a high-precision MAP;
and the matching module is used for matching a 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, wherein the first model is obtained by training the autonomous prediction lane information model, and the adaptive redundancy coefficient is obtained by training the adaptive redundancy coefficient model.
The training device of the first model comprises:
the first obtaining module S401 is configured to obtain historical vehicle state information and historical lane information, where the historical vehicle state information is input data of a training model, the historical lane information is 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 motion information, and the historical lane information is lane position information where a vehicle passes through.
The first training module S402 is 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 a vehicle will pass through, the first model is trained by using a time-based back propagation algorithm, and the first loss function is a minimum mean square error loss function.
In an alternative, the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctIs the cell state vector, W, U is the weight, b is the offset, σ is the excitation function。
In an optional mode, the vehicle high-precision positioning information comprises longitude, latitude and elevation of the vehicle, the vehicle motion state information comprises at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear position of the vehicle, and the historical lane information comprises longitude, latitude and elevation of a lane passed by the vehicle.
The training device of the adaptive redundancy coefficient model comprises:
a second obtaining module S501, configured to obtain 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 of a training model;
a third obtaining module S502, configured to obtain redundancy coefficient data in a manual or automatic labeling manner, where the redundancy coefficient is an output feature vector of the training model;
the second training module S503 is configured to train the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function, and the first model, where the second model is a shallow neural network model, the second model is used to obtain an adaptive redundancy coefficient of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
In an alternative, the second model formula:
R＝σ(WrN+br)
where N is the input feature vector of the second model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, and WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
In an alternative, the second loss function equation:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
The system and apparatus embodiments correspond to the system embodiments, and have the same technical effects as the method embodiments, and for the specific description, refer to the method embodiments. The device embodiment is obtained based on the method embodiment, and for specific description, reference may be made to the method embodiment section, which is not described herein again. Those of ordinary skill in the art will understand that: the figures are merely schematic representations of one embodiment, and the blocks or flow diagrams in the figures are not necessarily required to practice the present invention.
Those of ordinary skill in the art will understand that: modules in the devices in the embodiments may be distributed in the devices in the embodiments according to the description of the embodiments, or may be located in one or more devices different from the embodiments with corresponding changes. The modules of the above embodiments may be combined into one module, or further split into multiple sub-modules.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.
Claims (9)
1. An autonomous predictive lane information model training method, the method comprising:
acquiring historical vehicle state information and historical lane information, wherein the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where a vehicle passes;
training a first model according to the historical vehicle state information, the historical lane information and a first loss function, wherein the first model is a long-short term memory recurrent neural network model, the first model is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a time-dependent back propagation algorithm, and the first loss function is a minimum mean square error loss function.
2. The method of claim 1, wherein the first model formula:
ft＝σg(Wfxt+Ufht-1+bf)
it＝σg(Wixt+Uiht-1+bi)
ot＝σg(Woxt+Uoht-1+bo)
wherein x istFor inputting the feature vector, htAs an output vector, ftFor the excitation vector for controlling the forgetting gate, itFor the excitation vector used to control the input gate, otFor the excitation vector used to control the output gates,
input an excitation vector for the cell, ctFor the cell state vector, W, U is the weight, b is the offset, and σ is the excitation function.
3. The method of claim 1, wherein the vehicle high-precision positioning information comprises longitude, latitude and elevation of the vehicle, the vehicle motion state information comprises at least one of x-axis speed, y-axis speed, z-axis speed, x-axis acceleration, y-axis acceleration, z-axis acceleration, yaw rate, accelerator pedal opening, brake pedal opening and gear position of the vehicle, and the historical lane information comprises longitude, latitude and elevation of a lane through which the vehicle passes.
4. A method for training an adaptive redundancy coefficient model, wherein the adaptive redundancy coefficient is applied to lane matching, the method comprising:
acquiring vehicle high-precision positioning information, road curvature and weather humidity, wherein the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of a training model;
obtaining redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output characteristic vector of a training model;
training the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, wherein the second model is a shallow neural network model and is used for obtaining a self-adaptive redundancy coefficient of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
5. The method of claim 4, wherein the second model formula:
R＝σ(WrN+br)
wherein N is the secondThe input feature vector of the model, R is the adaptive redundancy coefficient, σ is the ReLU activation function, WrWeight of the shallow neural network, brIs the offset of the shallow neural network.
6. The method of claim 4, wherein the second loss function formula:
CrossEntropy(match(PointList,N0,N1|R),GroundTruth)
the PointList is output data of the first model, match () is a lane matching algorithm, N0 is vehicle longitude, N1 is vehicle latitude, R is a self-adaptive redundancy coefficient, and GroundTruth is actual road information.
7. An algorithm for lane matching based on a small amount of high precision map data, characterized in that the method comprises:
acquiring high-precision positioning information of a target vehicle and at least a part of MAP information received by the target vehicle, wherein the MAP information corresponds to a high-precision MAP;
matching a target lane of a target vehicle according to the target vehicle high-precision positioning information, the MAP information, the first model and the adaptive redundancy coefficient, wherein the first model is obtained by training the autonomous prediction lane information model according to any one of claims 1 to 3, and the adaptive redundancy coefficient is obtained by training the adaptive redundancy coefficient model according to any one of claims 4 to 6.
8. An autonomous predictive lane information model training device, characterized by comprising:
the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring historical vehicle state information and historical lane information, the historical vehicle state information is input data of a training model, the historical lane information is output data of the training model, the historical lane information corresponds to the historical vehicle state information, the historical vehicle state information comprises historical vehicle high-precision positioning information and historical vehicle motion information, and the historical lane information is lane position information where a vehicle passes;
the first training module is used for training 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-term and short-term memory recurrent neural network model, the first model is used for autonomously predicting lane position information about a vehicle to pass through, the first model is trained by using a time-based back propagation algorithm, and the first loss function is a minimum mean square error loss function.
9. An adaptive redundancy coefficient model training apparatus, the adaptive redundancy coefficient being applied to lane matching, the apparatus comprising:
the second acquisition module is used for acquiring vehicle high-precision positioning information, road curvature and weather humidity, wherein the vehicle high-precision positioning information, the road curvature and the weather humidity are input feature vectors of the training model;
the third acquisition module is used for acquiring redundancy coefficient data in a manual or automatic marking mode, wherein the redundancy coefficient is an output feature vector of the training model;
the second training module is used for training the second model according to the vehicle high-precision positioning information, the road curvature, the weather humidity, the redundancy coefficient data, a second loss function and the first model, the second model is a shallow neural network model and is used for obtaining self-adaptive redundancy coefficients of a vehicle matching lane, the second model is trained by using a back propagation algorithm, and the second loss function is a cross entropy loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111555631.6A CN114169463A (en) | 2021-12-17 | 2021-12-17 | Autonomous prediction lane information model training method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111555631.6A CN114169463A (en) | 2021-12-17 | 2021-12-17 | Autonomous prediction lane information model training method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114169463A true CN114169463A (en) | 2022-03-11 |
Family
ID=80487314
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111555631.6A Pending CN114169463A (en) | 2021-12-17 | 2021-12-17 | Autonomous prediction lane information model training method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114169463A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114973616A (en) * | 2022-05-19 | 2022-08-30 | 北京京天威科技发展有限公司 | Construction safety monitoring method and system for mechanical equipment |
-
2021
- 2021-12-17 CN CN202111555631.6A patent/CN114169463A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114973616A (en) * | 2022-05-19 | 2022-08-30 | 北京京天威科技发展有限公司 | Construction safety monitoring method and system for mechanical equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11713006B2 (en) | Systems and methods for streaming processing for autonomous vehicles | |
| CN110850861B (en) | Attention-based hierarchical lane-changing depth reinforcement learning | |
| US11269329B2 (en) | Dynamic model with learning based localization correction system | |
| US11755014B2 (en) | Jointly learnable behavior and trajectory planning for autonomous vehicles | |
| CN112099496B (en) | Automatic driving training method, device, equipment and medium | |
| US11537127B2 (en) | Systems and methods for vehicle motion planning based on uncertainty | |
| EP3959576A1 (en) | Methods and systems for trajectory forecasting with recurrent neural networks using inertial behavioral rollout | |
| CN110850854A (en) | Autonomous driver agent and policy server for providing policies to autonomous driver agents | |
| CN110673602B (en) | Reinforced learning model, vehicle automatic driving decision method and vehicle-mounted equipment | |
| JP2018037064A (en) | Online learning method and vehicle control method based on reinforcement learning without active search | |
| WO2021178909A1 (en) | Learning point cloud augmentation policies | |
| EP4020113A1 (en) | Dynamic model evaluation package for autonomous driving vehicles | |
| US11673584B2 (en) | Bayesian Global optimization-based parameter tuning for vehicle motion controllers | |
| JP2019031268A (en) | Control policy learning and vehicle control method based on reinforcement learning without active exploration | |
| EP4113393A2 (en) | Learning-based critic for tuning a motion planner of autonomous driving vehicle | |
| CN112406904A (en) | Method and device for training automatic driving strategy, automatic driving method, equipment, vehicle and computer readable storage medium | |
| CN111257853B (en) | Automatic driving system laser radar online calibration method based on IMU pre-integration | |
| CN114169463A (en) | Autonomous prediction lane information model training method and device | |
| US20220204016A1 (en) | Enhanced vehicle operation | |
| WO2021052383A1 (en) | Methods and systems for observation prediction in autonomous vehicles | |
| CN113424209A (en) | Trajectory prediction using deep learning multi-predictor fusion and bayesian optimization | |
| EP4171080A1 (en) | Intelligent messaging framework for vehicle ecosystem communication | |
| CN109492835B (en) | Method for determining vehicle control information, method for training model and related device | |
| EP4270352A1 (en) | Controlling a future traffic state on a road segment | |
| US20230227061A1 (en) | Systems and Methods for Pareto Domination-Based Learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |