Vehicle longitudinal speed prediction method based on vehicle-road cloud cooperation

Abstract

The invention relates to a vehicle longitudinal speed prediction method based on vehicle-road cloud cooperation. According to the method, a plurality of single prediction models capable of being used for longitudinal speed estimation are determined according to the actual composition of a system. Then determining available single prediction models, selecting a plurality of single prediction models from the plurality of available single prediction models obtained in the step of determining the available single prediction models, endowing the selected single prediction models with corresponding weight values, and combining the selected single prediction models into a combined prediction model; and the prediction validity of the obtained combined prediction model is not less than the prediction validity of any single prediction model. And finally, estimating the longitudinal speed of the autonomous vehicle at the moment through the constructed combined prediction model. According to the method, multiple different longitudinal vehicle speed prediction models are combined, and more accurate longitudinal vehicle speed estimation is provided when the vehicle speed of the automatic driving vehicle is determined in real time. According to the longitudinal vehicle speed prediction method, results from different longitudinal vehicle speed prediction models can be fused, and a better longitudinal vehicle speed prediction result is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of longitudinal vehicle speed prediction for autonomous driving, and particularly to a method for predicting the longitudinal vehicle speed of a vehicle with vehicle-road-cloud cooperation. Background Art

[0002] According to statistics, all major and extremely serious traffic accidents in open-pit mines occur on curved road sections with poor visibility and dangerous roadside conditions. When a mining vehicle is driving on a heavy-load curve, it often experiences dangerous situations such as rollover and sideslip due to excessive speed. Therefore, how to accurately predict the longitudinal vehicle speed of a mining vehicle when driving on a curve and maintain the vehicle speed is of great significance for reducing the occurrence of such accidents. Summary of the Invention

[0003] In order to more accurately determine the longitudinal vehicle speed of an autonomous vehicle and reduce the safety risks during the operation of the autonomous driving system, the present invention provides a method for predicting the longitudinal vehicle speed of a vehicle with vehicle-road-cloud cooperation.

[0004] The present application provides a method for predicting the longitudinal vehicle speed of a vehicle with vehicle-road-cloud cooperation, including the following steps: S1. Determine available single prediction models: Determine multiple single prediction models that can be used for longitudinal vehicle speed estimation according to the actual composition of the system.

[0005] S2. Construct a combined prediction model: Select several single prediction models from the multiple available single prediction models obtained in step S1 of determining available single prediction models, assign corresponding weight values to the selected single prediction models to form a combined prediction model, and ensure that the prediction validity of the obtained combined prediction model is not less than that of any single prediction model.

[0006] S3. Estimate the safe speed: Estimate the longitudinal vehicle speed of the autonomous vehicle at this moment through the combined prediction model constructed in step S2 of constructing the combined prediction model.

[0007] Preferably, in the step of S1 determining available single prediction models, the multiple different longitudinal vehicle speed prediction models include a cloud longitudinal vehicle speed prediction model dominated by a cloud system, a vehicle-end longitudinal vehicle speed prediction model dominated by an autonomous mining vehicle, and a roadside longitudinal vehicle speed prediction model dominated by a roadside system.

[0008] Preferably, in the step of S3 estimating the safe speed, based on the combined prediction model determined in step S2 of constructing the combined prediction model, the longitudinal vehicle speeds of different models are weighted and averaged based on the assigned weights to obtain the final estimation of the longitudinal vehicle speed.

[0009] Preferably, the step of S2 constructing the combined prediction model includes the following sub-steps: Complete without order S21. Calculate the prediction effectiveness of the single prediction model; S22. Calculate the confidence level of the single prediction model; Then execute S23. Determine the combined prediction model: Try to construct combinations of different single prediction models, construct the weights of each single prediction model based on the confidence level of the single prediction model in this combination, and then determine the prediction effectiveness of the combined prediction model based on the weights of each single prediction model and the prediction effectiveness; Keep the one with the highest prediction effectiveness of the combined prediction model as the finally determined combined prediction model.

[0010] Preferably, in the step of S23 for determining the combined prediction model, the prediction effectiveness of different single prediction models is sorted to obtain .

[0011] According to the sorting result, the result of combining adjacent single prediction models is recursively combined with the subsequent single prediction model for rolling update to obtain the final combined prediction model.

[0012] Preferably, the result of combining adjacent single prediction models is determined as follows. Determine as the result of the combination; If , then determine as the result of the combination.

[0013] The vehicle longitudinal speed prediction method of vehicle-road-cloud cooperation in this application combines multiple different longitudinal speed prediction models to provide a more accurate longitudinal speed estimate when determining the speed of an autonomous vehicle in real time. This longitudinal speed prediction method can fuse the results from different longitudinal speed prediction models to achieve a better longitudinal speed prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is the overall flow schematic diagram of the vehicle longitudinal speed prediction method of vehicle-road-cloud cooperation of the present invention; Figure 2 is the schematic diagram of a specific embodiment of the vehicle longitudinal speed prediction method of vehicle-road-cloud cooperation of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. In this specification, the scale ratio of the drawings does not represent the actual scale ratio. It is only used to reflect the relative positional relationship and connection relationship between components. Components with the same name or the same reference numeral represent similar or the same structures, and are for illustrative purposes only.

[0016] During the process of autonomous driving, it is of great significance to determine the longitudinal vehicle speed of an autonomous vehicle in real time. By restricting the longitudinal vehicle speed of the autonomous vehicle within a given safe speed range, the safety risks brought about by the excessive speed of the autonomous vehicle can be prevented. Taking the autonomous driving planning process of mining vehicles as an example, due to their large load and many curves, if the mining vehicle travels too fast, accidents such as rollover and capsizing may occur.

[0017] In the current mining autonomous driving system, the control of the longitudinal vehicle speed is mainly completed based on unilateral information. Specifically in practical applications, it is mostly generated by the vehicle end based on navigation information, or based on the wheel speed sensors of the vehicle end. The accuracy of the navigation information generation method is sufficient in most cases. However, in the mining scenario, due to the relatively complex environmental conditions and signal occlusion, the navigation information is often lost, so there are certain stability problems. Moreover, the speed information generated by the vehicle's own wheel speed sensors, etc., has a certain difference from the actual running vehicle speed of the vehicle because it cannot record the slip between the wheels and the ground. Therefore, in practical applications, this means that the obtained longitudinal vehicle speed is still rough and there is a certain probability of safety risks.

[0018] The roadside V2X vehicle speed measurement hardware is relatively accurate under the working conditions where the vehicle speed is extremely low and there is no intense acceleration or deceleration. However, in the case of extremely low speeds, its update period becomes longer, so there will be a problem of slow update in the longitudinal vehicle speed estimation. As a platform with comprehensive capabilities in multiple aspects such as comprehensive application, data service, and data interface, the cloud control platform can obtain the historical driving speeds of each vehicle and can effectively avoid noise interference. To give full play to the advantages of multiple estimation models, this application provides a vehicle-road-cloud collaborative vehicle longitudinal vehicle speed prediction method, which combines multiple different longitudinal vehicle speed prediction models to provide a more accurate longitudinal vehicle speed estimation when determining the vehicle speed of an autonomous vehicle in real time. This longitudinal vehicle speed prediction method can fuse the results from different longitudinal vehicle speed prediction models to achieve a better longitudinal vehicle speed prediction result.

[0019] The steps of the vehicle-road-cloud collaborative vehicle longitudinal vehicle speed prediction method of this application are generally as follows: S1. Determine the available single prediction models. Determine multiple single prediction models that can be used for longitudinal vehicle speed estimation according to the actual composition of the system.

[0020] S2. Construct a combined prediction model. Select several single prediction models from the multiple available single prediction models obtained in S1, assign corresponding weight values to the selected single prediction models to form a combined prediction model, and ensure that the prediction effectiveness of the obtained combined prediction model is not less than that of any single prediction model.

[0021] S3. Safe speed estimation. Estimate the longitudinal vehicle speed of the autonomous vehicle at this moment through the combined prediction model constructed in S2.

[0022] In the step of S1 for determining the available single prediction models, since there may be a combination of different single prediction models used to comprehensively estimate the predicted value of the final longitudinal vehicle speed subsequently. Therefore, in this process, different single prediction models are preferably independent of each other, that is, the processes of realizing the longitudinal vehicle speed estimation are not related to each other.

[0023] Taking the mine scenario as an example, in the step of S1 for determining the available single prediction models, multiple different longitudinal vehicle speed prediction models can be respectively the cloud longitudinal vehicle speed prediction model dominated by the cloud system, the vehicle-end longitudinal vehicle speed prediction model dominated by the autonomous mining vehicle, and the roadside longitudinal vehicle speed prediction model dominated by a package of roadside systems, etc. The cloud longitudinal vehicle speed prediction model is generally realized relying on the cloud system of the intelligent mine. As a platform with comprehensive capabilities in multiple aspects such as comprehensive application, data service, and data interface, the cloud system can obtain information such as the historical driving speed of the autonomous vehicle, the global map of the driving scenario, and the planned driving path of the vehicle, and thus determine a longitudinal vehicle speed at this moment based on the information obtained. The vehicle-end longitudinal vehicle speed prediction model mainly determines the appropriate longitudinal vehicle speed based on the state of the autonomous mining vehicle itself, such as the current vehicle speed, the working state of the vehicle braking system, the working state of sensors such as lidar and millimeter-wave radar, etc. The roadside longitudinal vehicle speed prediction model mainly determines the appropriate longitudinal vehicle speed based on the information obtained by the roadside sensors. Depending on the types of sensors installed, the roadside longitudinal vehicle speed prediction model may be able to obtain meteorological conditions in the scenario, such as visibility conditions, which can be estimated from the images obtained by the imaging equipment in the roadside system, or from the effective point ratio of the point cloud data fed back by sensors such as radar installed in the roadside system. In addition, the roadside system may also include real-time updates of the local road conditions and road surface quality, such as the number of local vehicles, the number of gravel and foreign objects on the road surface, etc. The roadside system determines the longitudinal vehicle speed at this time based on the above information, etc.

[0024] S2. Construct a combined prediction model. As Figure 2 shown, two tasks need to be completed when constructing the combined prediction model. One is to determine which single prediction models to select for combination, and the other is to determine how to match any single prediction model when combined into the combined prediction model. Or in the embodiments of the present application, this means how to allocate the weights of the single prediction models during the combination process.

[0025] The following is an embodiment for realizing S2 to construct the combined prediction model.

[0026] S21. Calculate the prediction effectiveness of the single prediction model.

[0027] S22. Calculate the confidence level of the single prediction model.

[0028] S23. Determine the combined prediction model. Try to construct combinations of different single prediction models. In this combination, based on the confidence level of the single prediction model, construct the weight of each single prediction model, and then based on the weight of each single prediction model and the prediction effectiveness, determine the prediction effectiveness of the combined prediction model. Retain the one with the highest prediction effectiveness of the combined prediction model as the finally determined combined prediction model.

[0029] The following is an optional implementation. For the prediction effectiveness of different single prediction models (where i is the counting flag of the single prediction model) are sorted, and assume it is .

[0030] The determination of the combined prediction model lies in the determination of the single prediction model and the weight configuration therein. Combine the single prediction models with the top two prediction effectiveness and , configure the weights to initially determine the linear combined prediction model . If indicates that the introduction of the prediction model does not improve the prediction effectiveness, then is considered an effectiveness redundant model, and is determined as the result of the combination; if , it indicates that the introduction of the single prediction model has improved the prediction effectiveness, then is determined as the preliminary combined prediction model for time t. According to this scheme, combined with the sorting result of the prediction performance of the single model at time t, the obtained preliminary combined prediction model is recursively combined with the subsequent single prediction model for rolling update to obtain the final combined prediction model.

[0031] The prediction effectiveness evaluation solves the problems of sorting and selecting the optimal single prediction model. In the step of calculating the prediction effectiveness of the single prediction model in S21, let be the statistical data of the speed of vehicle y up to time t (determine t as the acquisition point of the discontinuous moment), be the estimated vehicle speed of the i-th prediction method of vehicle y up to time t, then the prediction error can be determined as , and the prediction error vector formed by the prediction errors is ), and then the prediction error vector of all models is determined as , and the sum of the squares of the prediction errors . The covariance of different prediction errors ; A linear combination prediction model based on the above single prediction models.

[0032]

[0033] Weight coefficient vector , and satisfy = 1, , denote as the predicted value of vehicle y at time t based on the combined prediction model.

[0034] It is stipulated that the relative error of vehicle y at time t based on the i-th single prediction model is , and

[0035] Then it is called as the vehicle y speed prediction accuracy based on the i-th single prediction model at the cut-off time t.

[0036] Let be the prediction effectiveness of vehicle y method . Where:

[0037]

[0038] Among them, is the mathematical expectation of the prediction accuracy sequence of vehicle y method at the cut-off time t. The larger its value, the better the prediction effect. It is an increasing function of the prediction accuracy sequence; is the standard deviation of the prediction accuracy sequence of vehicle y single prediction model at the cut-off time t, used to describe the instability of the prediction accuracy sequence. The smaller the value, the better the prediction effect. It is a decreasing function of the prediction accuracy sequence.

[0039] The above method is also suitable for determining the prediction effectiveness of the combined model, and can effectively screen out redundant single prediction models. The so-called redundant single prediction model is the added prediction method that does not reduce the sum of squared prediction errors of the combined model and does not improve its prediction effectiveness.

[0040] S22. Calculate the confidence of the single prediction model. The core idea of the information fusion in this application is that each prediction model makes a local decision on the vehicle to be supervised according to the results of its respective historical prediction database, and then transmits the evaluation results to the fusion center via the communication channel. The fusion center then comprehensively processes each local decision using certain criteria to make a final decision. However, due to various interferences, the uncertainty of the local decision results will occur, making various criteria in the decision-level fusion not necessarily "optimal" or globally "optimal" for the system. To solve this problem, a confidence function is used to improve the confidence of the system.

[0041] Multiple single prediction models participate in the evaluation of the same parameter of the vehicle to be predicted. The data measured by the i-th prediction model and the j-th prediction model are respectively and . If is more real, is more trusted by the remaining data. The so-called is trusted by degree of trust, that is, from viewpoint the possibility of being real data. This degree of trust between the data measured by multiple prediction models is called confidence.

[0042] To further uniformly quantify the confidence between the information fed back by each prediction model, a confidence function is defined. Represents is trusted by degree. According to the definition of confidence, let:

[0043] In the formula, is a continuous decreasing function, and .

[0044] The general fusion method is to give a fusion upper limit ( >0). For , let

[0045] If =0, it is considered that the i-th prediction model and the j-th prediction model do not trust each other. If If \(T_{ij} = 1\), it is considered that the \(i\)-th prediction model fully trusts the \(j\)-th prediction model. If a prediction model is not trusted by other prediction models or is only trusted by a few prediction models, the information provided by this prediction model will be deleted during information fusion. Such processing is not conducive to making an objective judgment on the actual situation, and further makes the fusion result overly affected by subjective factors.

[0046] Based on the above analysis, the trust degree function can be defined in the form of an exponential function. Let:

[0047] It can be seen that the smaller the value of , the larger is, that is, the mutual trust degree between data and is greater; when the value of is 0, at this time is very small. Conversely, when the value of

[0048] is very large, Since the exponential function takes values from 1 to 0 monotonically decreasing on , it satisfies the properties that the trust degree function in Equation (5.2) should have. In practical applications, when the value of exceeds the set upper limit value , it can be considered that these two data no longer trust each other, and at this time

[0049] Define in the form of a fuzzy exponential function. In this way, it can not only make full use of the advantage of the range determination of the membership function in fuzzy theory, but also avoid the absolutization of the mutual trust degree between data, which is more in line with the authenticity of practical problems. At the same time, it is convenient for specific implementation and can make the fused result more accurate and stable.

[0050] Suppose there are prediction models providing the same index of the same vehicle to be predicted. According to the trust degree function between data, a trust degree matrix is established as follows:

[0051] For the \(i\)-th row element in , if A larger value indicates that the information provided by the prediction model i is trusted by most prediction models; conversely, it indicates that the information provided by the prediction model i is less likely to be real data.

[0052] S3. Safe speed estimation. Estimate the longitudinal vehicle speed of the autonomous vehicle at this moment through the combined prediction model constructed in S2. Briefly, in the combined prediction model determined in step S2, the longitudinal vehicle speeds of different models are weighted and averaged based on the assigned weights to obtain the final estimated longitudinal vehicle speed.

[0053] The principle is as follows. Let represent the weight of the information provided by the prediction model i in the fusion process. Since the value reflects the comprehensive trust degree of the information of other prediction models in , we can use to perform a weighted sum on to obtain the expression for data fusion: ,

[0054] In the formula, the weight coefficient of should satisfy:

[0055] In the trust matrix B, the trust function only represents the trust degree of the measured data in , and cannot reflect the trust degree of the measured data of all prediction models in the system in , while the true degree should actually be reflected by comprehensively.

[0056] should comprehensively consider all the information of each in a trust degree system about , so a set of non - negative numbers needs to be obtained such that:

[0057] Rewritten in matrix form as:

[0058] In the formula, , .

[0059] Because , so the trust matrix B is a non - negative matrix, and this symmetric matrix has a maximum - modulus eigenvalue , such that:

[0060] Find out and the corresponding eigenvector A, and the components in A satisfy (i = 1, 2, …, n). Substituting gives:

[0061] It can be used as a measure of the comprehensive trust degree among the information provided by each prediction model, that is:

[0062] Considering the weight coefficient should satisfy the aforementioned conditions, it is necessary to normalize to obtain:

[0063] The final result of the information fusion measured for all prediction models is: .

[0064] The above content is only a description of the preferred embodiments of the present invention and does not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention should fall within the protection scope determined by the claims of the present invention.

Claims

1. A vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration, characterized in that: The steps include: S1. Determine available single prediction models: determine multiple single prediction models that can be used for longitudinal vehicle speed estimation according to the actual configuration of the system; S2. Constructing a combined prediction model: selecting a number of single prediction models from the multiple available single prediction models obtained in the step of determining the available single prediction models in S1, assigning corresponding weight values ​​to the selected single prediction models and combining them into a combined prediction model, and ensuring that the prediction validity of the obtained combined prediction model is not less than the prediction validity of any single prediction model; S3. Safe speed estimation: Estimate the longitudinal speed of the autonomous driving vehicle at that moment through the combined prediction model constructed in step S2.

2. The vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration as claimed in claim 1, characterized in that: In the step of S1 determining the available single prediction models, the multiple different longitudinal vehicle speed prediction models include a cloud-based longitudinal vehicle speed prediction model dominated by a cloud system, a vehicle-side longitudinal vehicle speed prediction model dominated by an autonomous driving mining car, and a roadside longitudinal vehicle speed prediction model dominated by a roadside system.

3. The vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration as claimed in claim 1, characterized in that: In the safety speed estimation step S3, based on the combined prediction model determined in the combined prediction model construction step S2, the longitudinal vehicle speeds of different models are weighted averaged based on the assigned weights to obtain the final longitudinal vehicle speed estimate.

4. The vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration as claimed in claim 1, characterized in that: S2 step of building a combined prediction model includes the following sub-steps: Complete in no particular order S21, calculating the prediction effectiveness of the single prediction model; S22, calculating the confidence of a single prediction model; Then execute S23. Determine the combined prediction model: try to construct a combination of different single prediction models, in which the weight of each single prediction model is constructed based on the trust of the single prediction model, and then determine the prediction effectiveness of the combined prediction model based on the weight and prediction effectiveness of each single prediction model; retain the combined prediction model with the highest prediction effectiveness as the final combined prediction model.

5. The vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration as claimed in claim 4, characterized in that: S23 determines the prediction effectiveness of different single prediction models in the combined prediction model step Sort, where i takes the count of 1-n, and get ; According to the sorting result, the combination results of adjacent single prediction models are recursively combined with a subsequent single prediction model to perform rolling update to obtain the final combined prediction model.

6. The vehicle longitudinal speed prediction method based on vehicle-road-cloud collaboration as claimed in claim 4, characterized in that: The results of the combination of adjacent single prediction models are determined as follows, Sure is the result of the combination; if , then determine is the result of the combination.

Images