Vehicle speed prediction method and system considering dynamic following scenario in traffic environment

Abstract

A method and system for predicting vehicle speed in a dynamic following scenario that considers a traffic environment. The method includes extracting case data by setting standards and thresholds for the following process in a natural driving dataset; using a target detection algorithm to obtain traffic environment characteristics related to the following case data, extracting vehicle dynamics parameters from the following case data, fusing the vehicle dynamics parameters with the traffic environment characteristics, and normalizing them to obtain fused features; using a combined feature selection method to remove redundant variables from the fused features; dividing the fused features after removing redundant variables into a dataset, training a pre-established Informer-FDR vehicle speed prediction model, and using the trained speed prediction model to obtain vehicle speed prediction results. This invention effectively reflects the dynamic interaction characteristics between the driver and the traffic environment and potential traffic conflicts, enhances the model's understanding of complex driving environments, and improves the real-time and accuracy of vehicle speed prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle speed prediction, and in particular to a method and system for predicting vehicle speed in a dynamic following scenario taking into account a traffic environment. Background Art

[0002] Driven by the rapid development of driver assistance systems, intelligent vehicles have become a key solution for reducing traffic accidents and alleviating traffic congestion. Research has shown that if driver assistance systems can provide timely warnings to drivers, the accident rate can be reduced by 60%, highlighting the critical role of driver assistance systems in improving road safety. Vehicle speed prediction models are particularly important in driver assistance systems. Accurately predicting vehicle speed is one of the keys to achieving effective driver assistance systems and intelligent driving technologies. Furthermore, rear-end collisions account for 60%-70% of traffic accidents. By accurately predicting vehicle speed and behavior, driver assistance systems can better avoid potential collisions, thereby reducing the incidence of rear-end collisions. Therefore, accurately predicting vehicle speed in following vehicle scenarios has significant research significance and practical value for improving the effectiveness of driver assistance systems, enhancing human-like driving capabilities, and reducing rear-end collisions.

[0003] However, achieving accurate vehicle speed prediction is a complex task. In the following vehicle scenario, the driver is not only affected by the driving behavior of the vehicle in front, but also by the complex interactions of motor vehicles, non-motor vehicles and pedestrians in the surrounding lanes, causing the speed of the following vehicle to exhibit obvious fluctuation characteristics during the following cycle.

[0004] Existing methods rarely consider the impact of these dynamic environmental factors on the driver's speed choices during vehicle following. Furthermore, existing neural networks can only look back linearly to predict vehicle speed. This results in a limited receptive field, computational redundancy, and uncertainty about the impact of future information on time series predictions. Furthermore, increasing network depth can slow model training and reduce performance. Summary of the Invention

[0005] The purpose of the present invention is to address the problems in the above-mentioned prior art and provide a vehicle speed prediction method and system in a dynamic following scenario taking into account the traffic environment, so as to improve the real-time performance and accuracy of vehicle speed prediction in actual complex driving environments.

[0006] In order to achieve the above object, the present invention has the following technical solutions:

[0007] In a first aspect, a method for predicting vehicle speed in a dynamic following scenario taking into account a traffic environment is provided, comprising the following steps:

[0008] By setting standards and thresholds for the car-following process in the natural driving dataset, we extract car-following case data;

[0009] Use the target detection algorithm to obtain the traffic environment characteristics related to the car following case data, and extract the vehicle dynamics parameters in the car following case data. Then, fuse the vehicle dynamics parameters with the traffic environment characteristics and obtain the fused features after normalization.

[0010] Combined feature selection method is used to remove redundant variables in fusion features;

[0011] The fused features after removing redundant variables are divided into data sets, and the pre-established Informer-FDR vehicle speed prediction model is trained. The trained Informer-FDR vehicle speed prediction model is used to obtain the vehicle speed prediction results.

[0012] As a preferred solution, in the step of extracting the car-following case data by setting the standards and thresholds for the car-following process in the natural driving dataset, the standards and thresholds for the car-following process are as follows, and the car-following case data is selected when the following conditions are met:

[0013] The relative speed of the front and rear vehicles is less than 2.5m / s;

[0014] The lateral offset between the front and rear vehicles is less than 2.5m;

[0015] The following time is greater than 10 seconds;

[0016] The extracted car-following case data are divided into two road types: urban roads and highways.

[0017] As a preferred solution, the step of using the target detection algorithm to obtain traffic environment characteristics related to the vehicle following case data includes:

[0018] The yolov5 target detection algorithm is used to detect other vehicles and traffic participants in adjacent lanes. The traffic density is defined according to the number of vehicles traveling in the same direction as the vehicle as detected by the camera. The case where the number of vehicles traveling in the same direction is 1-2 is defined as low traffic density, the case where the number of vehicles traveling in the same direction is 3-4 is defined as medium traffic density, and the case where the number of vehicles traveling in the same direction is greater than 4 is defined as high traffic density. Pedestrians and cyclists identified by the target detection algorithm are combined into a unified category, defined as vulnerable traffic participants, which are used as input features of the Informer-FDR vehicle speed prediction model.

[0019] As a preferred solution, when extracting vehicle dynamic parameters from the vehicle following case data, the vehicle's own driving information is read through the CAN bus. The vehicle's own driving information includes speed, acceleration, relative distance, accelerator pedal opening, and headway.

[0020] As a preferred solution, the combined feature selection method for removing redundant variables in the fusion feature includes:

[0021] Use message integrity check (MIC), classification and regression tree (CART), random forest (RF), and gradient boosting ensemble learning (XGBoost) algorithms to calculate the correlation between candidate input features and vehicle speed;

[0022] The average value of each candidate variable is calculated by adding the relevant results of the four different algorithms and dividing it by the total number of algorithms. All variables are calculated and ranked in turn. Based on the ranking, the top few variables with the highest importance are selected as the input of the model.

[0023] As a preferred solution, the Informer-FDR vehicle speed prediction model establishes an Informer deep learning model in the following manner:

[0024] In the rolling forecast setting with a fixed window size, the input at time point t is:

[0025]

[0026] Where L is the length of the sample sequence, D is the feature dimension, and in X, a single data sample is:

[0027]

[0028] Where, Represents the data of the input of the i-th model at the j-th time step;

[0029] The vehicle speed is used as the output indicator:

[0030]

[0031] Where y i represents the output of the i-th model at each time step;

[0032] The encoder of the Informer deep learning model is composed of a multi-head sparse probabilistic self-attention mechanism and a stack of distillation operations:

[0033]

[0034] in, is a sparse matrix containing a sparsity measurement value; Q and K generated during encoder training represent the query vector and key vector respectively, V represents the value vector, and d represents the input dimension; the sparse probabilistic self-attention mechanism is defined as:

[0035]

[0036] Where, LK 1 / L K =q(k j |q i ); define the i queries of sparsity measurement as q i ;q i and k j from Q and K respectively;

[0037] The Informer deep learning model adds a convolutional layer and a max pooling layer between each pair of self-attention layers to reduce the input length. After the previous self-attention layer, a convolutional layer with a kernel size of 3 and a stride of 1 is used to enhance the perception of contextual information features. Subsequently, a max pooling layer with a kernel size of 3 and a stride of 2 is used to prioritize the main features, thereby generating smaller but more focused feature maps for the subsequent self-attention layer:

[0038]

[0039] Where, sparse probabilistic self-attention is contained in [·] AB In the figure, 1-D convolution filtering is implemented by Conv(·) with ELU(·) as the activation function. The encoder combines sparse probabilistic self-attention and distillation operations to generate feature maps.

[0040] Time series input to the decoder of the Informer deep learning model It is divided into two parts, namely the known sequence before the predicted point and forecast series that require masking future data

[0041]

[0042] Where, L token is the length of the starting tag sequence; L y is the length of the predicted sequence; d model is the model dimension;

[0043] The position encoding of each input information of the sequence is performed as follows:

[0044]

[0045] Where pos represents the position of the feature vector, d model Represents the dimension of the feature vector. Given the importance of time information for time series prediction, the time information in the time series is encoded, and the position information and time information are integrated into the input model.

[0046] As a preferred solution, the Informer-FDR vehicle speed prediction model replaces the maximum pooling layer of the distillation layer in the decoder of the Informer deep learning model with an average pooling layer, and changes the convolution layer to a dilated causal convolution. The dilated causal convolution consists of two parts: causal convolution and dilated convolution. The causal convolution is subject to time constraints and convolves the output at time t only with the time t and earlier elements of the previous layer to ensure causality in time series prediction, so that the network output information is only affected by past input information.

[0047] The dilated convolution operation is as follows:

[0048]

[0049] Where K is the convolution kernel size, s is the element of X, and F i is the weight of the convolution kernel, t is the time step, and td·i represents the past direction.

[0050] As a preferred solution, the Informer-FDR vehicle speed prediction model introduces PRELU as the activation function of the distillation layer. PRELU pushes the output mean to zero. The improved distillation operation is as follows:

[0051]

[0052] Where, sparse probabilistic self-attention is contained in [·] AB Middle; Dilated causal convolution is implemented by DCConv(·), the activation function is PReLU(·), and the final output is generated by AvgPool(·).

[0053] As a preferred solution, the Informer-FDR vehicle speed prediction model adds a residual connection module, which consists of three sparse probabilistic self-attention layers and two distillation layers. In order to ensure that the feature maps output by these two layers have the same shape when added together, a 1*1 convolution is used to merge the output of the first sparse probabilistic self-attention layer in the main encoder with the output of the third sparse probabilistic self-attention layer to ensure the effective transmission of feature information.

[0054] Secondly, a vehicle speed prediction system is provided for a dynamic vehicle following scenario taking into account the traffic environment, including:

[0055] A car-following case data extraction module is used to extract car-following case data by setting standards and thresholds for the car-following process in the natural driving dataset;

[0056] The feature fusion and normalization processing module is used to use the target detection algorithm to obtain the traffic environment features related to the car following case data, extract the vehicle dynamics parameters in the car following case data, fuse the vehicle dynamics parameters with the traffic environment features, and obtain the fused features after normalization.

[0057] Redundant variable removal module, used to remove redundant variables in fusion features using combined feature selection method;

[0058] The vehicle speed prediction model training and prediction output module is used to divide the fused features after removing redundant variables into data sets, train the pre-established Informer-FDR vehicle speed prediction model, and use the trained Informer-FDR vehicle speed prediction model to obtain vehicle speed prediction results.

[0059] Compared with the prior art, the present invention has at least the following beneficial effects:

[0060] By setting standards and thresholds for the vehicle-following process in a natural driving dataset, the model effectively extracts case data for vehicle-following, enabling targeted mining of this data and providing a more comprehensive understanding and analysis of the dynamic behavior of vehicle-following in the driving environment. A target detection algorithm is used to capture traffic environment features associated with the case data, extract vehicle dynamics parameters from the data, and fuse these parameters with traffic environment features. By integrating these features, the model effectively reflects the dynamic interaction between the driver and the traffic environment and potential traffic conflicts, enhancing the model's understanding of complex driving environments. The fused features, after removing redundant variables, are then divided into datasets, trained using a pre-established Informer-FDR speed prediction model, and speed prediction results are obtained using the trained Informer-FDR speed prediction model. Compared with other models, the Informer-FDR speed prediction model employed in this paper achieves the best prediction results, more accurately capturing the fluctuation trends of the data series and exhibiting a high degree of fit to the actual data. In particular, the Informer-FDR speed prediction model exhibits smaller prediction deviations at extreme points, demonstrating its ability to more accurately capture randomness in traffic conditions.

[0061] Furthermore, the present invention reduces the high computational complexity of the Informer-FDR vehicle speed prediction model by using a sparse probabilistic self-attention mechanism, which helps to solve the difficulties in applying Transformer-type deep learning models to vehicle-mounted platforms.

[0062] Furthermore, the Informer-FDR vehicle speed prediction model established in the present invention takes into account the impact of the dynamic traffic environment on vehicle speed fluctuations, realizes global feature extraction over a long vehicle following period, and solves the degradation problem of deep networks.

[0063] Furthermore, the present invention introduces dilated causal convolution to expand the receptive field of the Informer-FDR vehicle speed prediction model, ensure the causality of time prediction, and avoid the loss of extreme position information and information leakage.

[0064] Furthermore, the Informer-FDR vehicle speed prediction model of the present invention adds a residual connection module, which accelerates the convergence of the Informer-FDR vehicle speed prediction model and enhances the stability of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0066] Figure 1 This is a flow chart of a vehicle speed prediction method in a dynamic following vehicle scenario taking into account a traffic environment according to an embodiment of the present invention;

[0067] Figure 2 This is a network structure diagram of the Informer-FDR vehicle speed prediction model established in an embodiment of the present invention;

[0068] Figure 3 This is a diagram of the dilated causal convolution structure of an embodiment of the present invention;

[0069] Figure 4 This is a test result diagram of a vehicle speed prediction method in a dynamic following vehicle scenario in a traffic environment according to an embodiment of the present invention;

[0070] Figure 5 This is a comparison chart of the prediction performance of the Informer-FDR vehicle speed prediction model of an embodiment of the present invention and other algorithm models. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, ordinary technicians in this field can also derive other embodiments without making any creative work.

[0072] like Figure 1As shown, an embodiment of the present invention proposes a vehicle speed prediction method in a dynamic following scenario taking into account the traffic environment, which mainly includes: first, by setting the standards and thresholds of the following process in the natural driving data set, the following case data is effectively extracted; secondly, the target detection algorithm is used to obtain relevant environmental features, and the vehicle dynamics parameters under the following segment are extracted from the CAN bus, and integrated with the environmental features; then, a combined feature selection method is used to eliminate redundant variables, improve the model prediction accuracy and perform normalization processing; finally, the data set is divided into a training set, a test set, and a validation set, and a pre-established Informer-FDR vehicle speed prediction model is input for training and testing to finally obtain the vehicle speed prediction result. The implementation of the complete vehicle speed prediction method of the embodiment of the present invention is mainly divided into four steps: acquisition of following data, acquisition of traffic environment information based on the target detection algorithm, feature fusion and elimination of redundant variables, and establishment of the Informer-FDR vehicle speed prediction model for vehicle speed prediction.

[0073] Step 1: Acquisition of vehicle tracking data, including the following steps:

[0074] Step 1.1: Install a comprehensive data acquisition system consisting of four cameras on the ego vehicle: a facial recognition camera, a pedal camera, a front camera, and a wide-angle front camera.

[0075] Step 1.2: Video data captured by the front camera is used to identify traffic participants, while data related to the actual movement of the vehicle, such as speed, acceleration, steering wheel angle, and mileage, is extracted from the CAN-BUS system.

[0076] Step 1.3: Extract the following case data by setting the criteria and thresholds for the following process. All following case data must meet the specified thresholds. If any of these thresholds are not met, the corresponding event will not be included in the initial following case database. The thresholds set are:

[0077] 1) The relative speed of the front and rear vehicles is less than 2.5m / s;

[0078] 2) The lateral offset between the front and rear vehicles is less than 2.5m;

[0079] 3) The following time is greater than 10 seconds.

[0080] Furthermore, because urban roads and highways differ significantly in traffic conditions and driving behavior, which leads to differences in vehicle following behavior, the embodiment of the present invention divides vehicle following cases into two road types: urban roads and highways. These serve as input features for the model, enabling the model to better understand and learn vehicle following behavior under different road conditions.

[0081] Step 2: Obtaining traffic environment information based on the target detection algorithm, specifically including the following steps:

[0082] Step 2.1: Use the YOLOv5 object detection algorithm to detect other vehicles and traffic participants in adjacent lanes. Define traffic density based on the number of vehicles traveling in the same direction as the vehicle, as captured by the front camera. A medium traffic density is defined as a situation where there are 3 to 4 vehicles traveling in the same direction, and a high traffic density is defined as a situation where there are more than 4 vehicles traveling in the same direction.

[0083] Step 2.2: In traffic environments, drivers are often influenced by individuals who may be at higher risk on the road, which in turn affects their driving behavior. To this end, we merge pedestrians and cyclists identified by the object detection algorithm into a unified category, named vulnerable traffic participants, as input features for the model.

[0084] Step 3: Feature fusion and elimination of redundant variables, specifically including the following steps:

[0085] Step 3.1: The vehicle dynamics parameters extracted from the CAN-BUS system, including speed, acceleration, relative distance, accelerator pedal opening, headway, etc., are integrated with the traffic environment information obtained by the target detection algorithm and normalized.

[0086] Step 3.2: Use the Message Integrity Check (MIC), Classification and Regression Tree (CART), Random Forest (RF), and Gradient Boosting (XGBoost) ensemble learning algorithms to calculate the correlation between candidate input features and vehicle speed. Calculate the average value for each candidate variable by adding the correlation results from the four different algorithms and dividing by the total number of algorithms. All variables are calculated sequentially and ranked. Based on the rankings, select the seven most important variables as input to the model.

[0087] Step 4: Establish an Informer-FDR vehicle speed prediction model to predict vehicle speed. The Informer-FDR vehicle speed prediction model includes: an Informer time series prediction model, an expanded causal convolution module, and a residual connection module.

[0088] The Informer time series prediction model is an improvement on the Transformer model. It uses a multi-head probabilistic sparse self-attention mechanism to give greater weight to important features, reducing the time complexity of the network model. At the same time, it uses distillation operations between self-attention layers to reduce the input length of the next layer of the model, and uses generative inference decoding to improve the prediction speed and solve the long-term dependency problem of time series data.

[0089] The Informer-FDR model of the embodiment of the present invention is an improvement on the Informer model. It uses dilated causal convolution to replace the conventional convolution layer in the distillation layer between the attention blocks in the original encoder, and introduces an average pooling layer to improve the model performance. The activation function of the distillation layer is set to the rectified linear unit PRELU with parameters to speed up the convergence of the model. In addition, a residual connection module is added to enable information to be transmitted across the depth of the layer. This design not only improves the stability of training, but also alleviates the problem of deep network degradation. The overall network structure is as follows Figure 2 shown.

[0090] Step 4.1: Build the Informer deep learning model.

[0091] The Informer time series prediction model uses a multi-head probabilistic sparse self-attention mechanism to give greater weight to important features, reducing the time complexity of the network model. At the same time, it uses distillation operations between self-attention layers to reduce the input length of the next layer model, and uses generative reasoning decoding to improve the prediction speed and solve the long-term dependency problem of time series data.

[0092] In the rolling forecast setting with a fixed window size, the input at time point t is:

[0093]

[0094] Where L is the length of the sample sequence, D is the feature dimension, and in X, a single data sample is:

[0095]

[0096] Where, It represents the input data of the i-th model at the j-th time step. The vehicle speed is usually used as the output indicator, and the label Y is recorded as:

[0097]

[0098] Where y i represents the output of the i-th model at each time step.

[0099] The encoder in Informer is composed of a multi-head sparse probabilistic self-attention mechanism stacked with a distillation operation. Traditional self-attention mechanisms require quadratic dot product calculations, which is the primary factor contributing to their slow prediction training speed. The sparse probabilistic self-attention mechanism reduces time complexity and memory usage by selecting a small number of dot product pairs with high scores and ignoring other pairs with low scores. The specific operation is as follows:

[0100]

[0101] in, Is a sparse matrix containing a sparsity measurement value; Q and K generated during the encoder training process represent the query vector and key vector, respectively. In addition, V represents the value vector and d represents the input dimension.

[0102] The sparse probabilistic self-attention mechanism is defined as:

[0103]

[0104] Where, L K 1 / L K =q(k j |q i ); define the i queries of sparsity measurement as q i ;q i and k j from Q and K, respectively. In practice, the input lengths of the query vector and the key vector are usually equal in the self-attention computation, making the total time complexity and space complexity of ProbSparse Self-Attention both o(LlnL). Sparse ProbSparse Self-Attention selects the result with a higher score in the sparsity metric and calculates the corresponding score.

[0105] Informer adds a convolutional layer and a max pooling layer between each pair of self-attention layers to reduce the input length. After the previous self-attention layer, a convolutional layer with a kernel size of 3 and a stride of 1 is used to enhance the perception of contextual information features. Subsequently, a max pooling layer with a kernel size of 3 and a stride of 2 is used to prioritize the main features, thereby generating smaller but more focused feature maps for the subsequent self-attention layer:

[0106]

[0107] Where, sparse probabilistic self-attention is contained in [·] AB In the figure, 1-D convolution filtering is implemented by Conv(·) with ELU(·) as the activation function. The encoder combines sparse probabilistic self-attention and distillation operations to generate feature maps.

[0108] The time series input to the decoder It is divided into two parts, namely the known sequence before the predicted point and forecast series that require masking future data The formula is as follows:

[0109]

[0110] Where, L token is the length of the starting tag sequence; L y is the length of the predicted sequence; dmodel is the model dimension.

[0111] In the vehicle speed prediction problem, time series information represents the vehicle's operating status at every moment, and the order of the data is crucial. Since the Informer model does not have iterative operations like the RNN model, it is necessary to positionally encode each input information in the sequence. The specific process is shown in the formula:

[0112]

[0113] Where: pos represents the position of the feature vector, d model Represents the dimension of the feature vector. Given the importance of time information for time series prediction, the time information in the time series is encoded, and the position information and time information are integrated into the input model.

[0114] To address several key shortcomings of the Informer model when applied to vehicle speed prediction in car-following scenarios: traditional convolutional layers can only review linearly large amounts of historical information; stacking self-attention blocks with standard convolutional layers can lead to repeated and meaningless computations and leakage of future information. This embodiment of the present invention addresses these issues by introducing dilated causal convolution.

[0115] Step 4.2: Replace the maximum pooling layer of the distillation layer in the decoder of the Informer model with an average pooling layer, and change the conventional convolution layer to a dilated causal convolution. The structure of the dilated causal convolution is as follows: Figure 3 shown.

[0116] Dilated causal convolution consists of two parts: causal convolution and dilated convolution. Unlike traditional convolution, causal convolution is subject to time constraints. It convolves the output at time t only with the elements at time t and earlier in the previous layer to ensure causality in time series prediction, so that the network's output information is only affected by past input information. The dilated convolution operation is as follows:

[0117]

[0118] Where K is the convolution kernel size, s is the element of X, and F i is the weight of the convolution kernel, t is the time step, and td·i represents the past direction.

[0119] On this basis, the parameterized rectified linear unit (PRELU) is introduced as the activation function of the distillation layer. Compared with the exponential linear unit (ELU), the parameterized rectified linear unit (PRELU) pushes the output mean toward zero without significantly increasing computational complexity or overfitting risk, thereby promoting faster model convergence. The improved distillation operation is shown in the formula:

[0120]

[0121] Where, the sparse probability ProbSparse self-attention is contained in [·] AB Middle; Dilated causal convolution is implemented by DCConv(·), the activation function is PReLU(·), and the final output is generated by AvgPool(·).

[0122] Step 4.3: Add a residual connection module. The residual connection module consists of three sparse probabilistic self-attention layers and two distillation layers. To ensure that the feature maps output by these two layers have the same shape when added together, this embodiment of the present invention uses a 1x1 convolution to merge the output of the first sparse probabilistic self-attention layer in the main encoder with the output of the third sparse probabilistic self-attention layer, ensuring efficient transmission of feature information.

[0123] In order to verify the speed prediction effect of the method of the present invention in the vehicle following scenario, the vehicle following dataset constructed by the embodiment of the present invention was used for training to predict the vehicle speed 2 seconds later. The absolute error between the actual value and the predicted value of each frame was calculated. The test results and the error are shown in Figure 2. Figure 4 As shown in the figure, it can be seen that the error predicted by the method of the embodiment of the present invention is always kept at a low level, with only a few significant errors, and the fit with the actual vehicle speed is high.

[0124] Furthermore, the method is compared with other mainstream time series prediction methods to verify the superiority of the present invention. Figure 5 The results of a comparison of the method described in this paper and other vehicle speed prediction methods for a specific vehicle following segment are presented. As can be seen from the figure, the GRU's prediction performance is poor, while the Autoformer, Transformer, and LSTM models can capture data fluctuation trends and some extreme values, but there is a certain degree of bias in their predictions. Compared with other models, the Informer-FDR model achieves the best prediction results, more accurately capturing the fluctuation trends of the data series and achieving a higher degree of fit with the actual data. In particular, the Informer-FDR model has a smaller prediction bias at extreme points, indicating that it can more accurately capture the randomness of traffic conditions.

[0125] Another embodiment of the present invention further provides a vehicle speed prediction system in a dynamic following scenario taking into account a traffic environment, comprising:

[0126] A car-following case data extraction module is used to extract car-following case data by setting standards and thresholds for the car-following process in the natural driving dataset;

[0127] The feature fusion and normalization processing module is used to use the target detection algorithm to obtain the traffic environment features related to the car following case data, extract the vehicle dynamics parameters in the car following case data, fuse the vehicle dynamics parameters with the traffic environment features, and obtain the fused features after normalization.

[0128] Redundant variable removal module, used to remove redundant variables in fusion features using combined feature selection method;

[0129] The vehicle speed prediction model training and prediction output module is used to divide the fused features after removing redundant variables into data sets, train the pre-established Informer-FDR vehicle speed prediction model, and use the trained Informer-FDR vehicle speed prediction model to obtain vehicle speed prediction results.

[0130] Another embodiment of the present invention further proposes an electronic device, comprising: a memory storing at least one instruction; and a processor executing the instruction stored in the memory to implement the vehicle speed prediction method in the dynamic following scenario considering the traffic environment.

[0131] Another embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the vehicle speed prediction method in the dynamic following scenario considering the traffic environment is implemented.

[0132] For example, the instructions stored in the memory may be divided into one or more modules / units, which are stored in a computer-readable storage medium and executed by the processor to implement the vehicle speed prediction method for a dynamic vehicle following scenario taking into account the traffic environment as described in the present invention. The one or more modules / units may be a series of computer-readable instruction segments capable of performing specific functions, which describe the execution process of the computer program on the server.

[0133] The electronic device may be a computing device such as a smartphone, laptop, PDA, or cloud server. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the electronic device may include more or fewer components, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0134] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0135] The memory may be an internal storage unit of the server, such as a hard disk or memory of the server. The memory may also be an external storage device of the server, such as a plug-in hard disk equipped on the server, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Furthermore, the memory may include both an internal storage unit of the server and an external storage device. The memory is used to store the computer-readable instructions and other programs and data required by the server. The memory may also be used to temporarily store data that has been output or is about to be output.

[0136] It should be noted that the information interaction, execution process, etc. between the above-mentioned module units are based on the same concept as the method embodiment. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0137] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0138] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above-mentioned various method embodiments. Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk.

[0139] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0140] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A vehicle speed prediction method considering a dynamic following scenario in a traffic environment, characterized by: The following steps are involved: By setting standards and thresholds for the car-following process in the natural driving dataset, we extract car-following case data; Use the target detection algorithm to obtain the traffic environment characteristics related to the car following case data, and extract the vehicle dynamics parameters in the car following case data. Then, fuse the vehicle dynamics parameters with the traffic environment characteristics and obtain the fused features after normalization. Combined feature selection method is used to remove redundant variables in fusion features; The fused features after removing redundant variables are divided into data sets, and the pre-established Informer-FDR vehicle speed prediction model is trained. The trained Informer-FDR vehicle speed prediction model is used to obtain the vehicle speed prediction results. The Informer-FDR vehicle speed prediction model replaces the maximum pooling layer of the distillation layer in the decoder of the Informer deep learning model with an average pooling layer, and replaces the convolution layer with a dilated causal convolution. The dilated causal convolution consists of two parts: causal convolution and dilated convolution. Causal convolution is subject to time constraints. The output at time t is convolved only with the time t and earlier elements of the previous layer to ensure causality in time series prediction, so that the network output information is only affected by past input information. The dilated convolution operation is as follows: Where K is the convolution kernel size, s is the element of X, and F i is the weight of the convolution kernel, t is the time step, and td·i represents the past direction; The Informer-FDR vehicle speed prediction model introduces PRELU as the activation function of the distillation layer. PRELU pushes the output mean to zero. The improved distillation operation is as follows: Where, sparse probabilistic self-attention is contained in [·] AB Middle; Dilated causal convolution is implemented by DCConv(·), the activation function is PReLU(·), and the final output is generated by AvgPool(·); The Informer-FDR vehicle speed prediction model adds a residual connection module, which consists of three sparse probabilistic self-attention layers and two distillation layers. To ensure that the feature maps output by these two layers have the same shape when added together, a 1*1 convolution is used to merge the output of the first sparse probabilistic self-attention layer in the main encoder with the output of the third sparse probabilistic self-attention layer to ensure effective transmission of feature information.

2. The vehicle speed prediction method in a dynamic following scenario taking into account the traffic environment according to claim 1 is characterized in that: In the step of extracting the car-following case data by setting the standards and thresholds for the car-following process in the natural driving dataset, the standards and thresholds for the car-following process are as follows, and the car-following case data is selected when the following conditions are met: The relative speed of the front and rear vehicles is less than 2.5m / s; The lateral offset between the front and rear vehicles is less than 2.5m; The following time is greater than 10 seconds; The extracted car-following case data are divided into two road types: urban roads and highways.

3. The vehicle speed prediction method in a dynamic following scenario taking into account the traffic environment according to claim 1 is characterized in that: The step of using the target detection algorithm to obtain traffic environment characteristics related to the vehicle following case data includes: The yolov5 target detection algorithm is used to detect other vehicles and traffic participants in adjacent lanes. The traffic density is defined according to the number of vehicles traveling in the same direction as the vehicle as detected by the camera. The case where the number of vehicles traveling in the same direction is 1-2 is defined as low traffic density, the case where the number of vehicles traveling in the same direction is 3-4 is defined as medium traffic density, and the case where the number of vehicles traveling in the same direction is greater than 4 is defined as high traffic density. Pedestrians and cyclists identified by the target detection algorithm are combined into a unified category, defined as vulnerable traffic participants, which are used as input features of the Informer-FDR vehicle speed prediction model.

4. The vehicle speed prediction method in a dynamic following scenario taking into account the traffic environment according to claim 1 is characterized in that: When extracting the vehicle dynamics parameters from the vehicle following case data, the vehicle's own driving information is read through the CAN bus. The vehicle's own driving information includes speed, acceleration, relative distance, accelerator pedal opening, and headway.

5. The vehicle speed prediction method in a dynamic following scenario taking into account the traffic environment according to claim 1 is characterized in that: The combined feature selection method for removing redundant variables in the fusion feature includes: Use message integrity check (MIC), classification and regression tree (CART), random forest (RF), and gradient boosting ensemble learning (XGBoost) algorithms to calculate the correlation between candidate input features and vehicle speed; The average value of each candidate variable is calculated by adding the relevant results of the four different algorithms and dividing it by the total number of algorithms. All variables are calculated and ranked in turn. Based on the ranking, the top few variables with the highest importance are selected as the input of the model.

6. The vehicle speed prediction method in a dynamic following scenario taking into account traffic environment according to claim 1 is characterized in that: The Informer-FDR vehicle speed prediction model builds the Informer deep learning model as follows: In the rolling forecast setting with a fixed window size, the input at time point t is: Where L is the length of the sample sequence, D is the feature dimension, and in X, a single data sample is: Where, Represents the data of the input of the i-th model at the j-th time step; The vehicle speed is used as the output indicator: Where y i represents the output of the i-th model at each time step; The encoder of the Informer deep learning model is composed of a multi-head sparse probabilistic self-attention mechanism and a stack of distillation operations: in, is a sparse matrix containing a sparsity measurement value; Q and K generated during encoder training represent the query vector and key vector respectively, V represents the value vector, and d represents the input dimension; the sparse probabilistic self-attention mechanism is defined as: Where, L K 1 / L K =q(k j |q i ); define the i queries of sparsity measurement as q i ;q i and k j from Q and K respectively; The Informer deep learning model adds a convolutional layer and a max pooling layer between each pair of self-attention layers to reduce the input length. After the previous self-attention layer, a convolutional layer with a kernel size of 3 and a stride of 1 is used to enhance the perception of contextual information features. Subsequently, a max pooling layer with a kernel size of 3 and a stride of 2 is used to prioritize the main features, thereby generating smaller but more focused feature maps for the subsequent self-attention layer: Where, sparse probabilistic self-attention is contained in [·] AB In the figure, 1-D convolution filtering is implemented by Conv(·) with ELU(·) as the activation function. The encoder combines sparse probabilistic self-attention and distillation operations to generate feature maps. Time series input to the decoder of the Informer deep learning model It is divided into two parts, namely the known sequence before the predicted point and forecast series that require masking future data Where, L token is the length of the starting tag sequence; L y is the length of the predicted sequence; d model is the model dimension; The position encoding of each input information of the sequence is performed as follows: Where pos represents the position of the feature vector, d model Represents the dimension of the feature vector. Given the importance of time information for time series prediction, the time information in the time series is encoded, and the position information and time information are integrated into the input model.

7. A vehicle speed prediction system considering a dynamic following scenario in a traffic environment, characterized by: include: A car-following case data extraction module is used to extract car-following case data by setting standards and thresholds for the car-following process in the natural driving dataset; The feature fusion and normalization processing module is used to use the target detection algorithm to obtain the traffic environment features related to the car following case data, extract the vehicle dynamics parameters in the car following case data, fuse the vehicle dynamics parameters with the traffic environment features, and obtain the fused features after normalization. Redundant variable removal module, used to remove redundant variables in fusion features using combined feature selection method; The vehicle speed prediction model training and prediction output module is used to divide the fused features after removing redundant variables into data sets, train the pre-established Informer-FDR vehicle speed prediction model, and use the trained Informer-FDR vehicle speed prediction model to obtain vehicle speed prediction results; The Informer-FDR vehicle speed prediction model replaces the maximum pooling layer of the distillation layer in the decoder of the Informer deep learning model with an average pooling layer, and replaces the convolution layer with a dilated causal convolution. The dilated causal convolution consists of two parts: causal convolution and dilated convolution. Causal convolution is subject to time constraints. The output at time t is convolved only with the time t and earlier elements of the previous layer to ensure causality in time series prediction, so that the network output information is only affected by past input information. The dilated convolution operation is as follows: Where K is the convolution kernel size, s is the element of X, and F i is the weight of the convolution kernel, t is the time step, and td·i represents the past direction; The Informer-FDR vehicle speed prediction model introduces PRELU as the activation function of the distillation layer. PRELU pushes the output mean to zero. The improved distillation operation is as follows: Where, sparse probabilistic self-attention is contained in [·] AB Middle; Dilated causal convolution is implemented by DCConv(·), the activation function is PReLU(·), and the final output is generated by AvgPool(·); The Informer-FDR vehicle speed prediction model adds a residual connection module, which consists of three sparse probabilistic self-attention layers and two distillation layers. To ensure that the feature maps output by these two layers have the same shape when added together, a 1*1 convolution is used to merge the output of the first sparse probabilistic self-attention layer in the main encoder with the output of the third sparse probabilistic self-attention layer to ensure effective transmission of feature information.

Images