A method and system for predicting vehicle following in mixed traffic flow for autonomous driving

Abstract

The present invention discloses a method and system for predicting vehicle following in an AV-Mixed Traffic Flow (AV-MTF) vehicle. An AV-MTF following vehicle classification model and a vehicle following prediction model are designed. The output vehicle following acceleration, speed or trajectory can be used for identifying vehicle following speed anomalies and managing and controlling vehicles in an AV-MTF environment. The following vehicle classification model is used to identify the following vehicle category based on vehicle following motion information. The vehicle following prediction model is used to combine vehicle category with vehicle motion information to predict the sequence of vehicle following acceleration, speed or trajectory for a period of time in the future. The anomaly recognition algorithm judges vehicle abnormal collection data by adaptively detecting whether the difference between the predicted speed of the following vehicle and its actual collection speed is within a reasonable range based on a one-time predicted acceleration or speed sequence. The method of the present invention is simple and easy to implement, and can better manage and control vehicles in an AV-MTF environment, thereby helping to improve road traffic efficiency and safety.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent transportation and autonomous driving vehicles, and in particular to a method and system for predicting car-following in autonomous driving mixed traffic flows. Background Art

[0002] The full development of autonomous vehicles (AVs) will take a long time. AV-Mixed Traffic Flow (AV-MTF), where AVs and human-driven vehicles (HVs) coexist, will continue to exist for a long time. However, existing data-driven car-following models based on deep learning algorithms mostly predict speed in traditional HV environments, ignoring the different car-following motion characteristics of AVs and HVs in AV-MTF environments. Accurately constructing an AV-MTF car-following prediction model can better manage and control vehicles in AV-MTF environments, thereby improving road traffic efficiency and safety. Summary of the Invention

[0003] Purpose of the invention: In response to the problems existing in the prior art, the purpose of the present invention is to establish a method and system for predicting vehicle following in autonomous driving mixed traffic flow, so as to accurately predict the vehicle following acceleration, speed or trajectory.

[0004] Technical solution: In order to achieve the purpose of the above invention, the present invention provides the following technical solution:

[0005] A method for predicting and identifying anomalies of the following speed of vehicles in mixed traffic flow under autonomous driving comprises the following steps:

[0006] (1) Obtaining the motion information of the lead vehicle and the following vehicle in the AV-MTF;

[0007] (2) After obtaining the following vehicle information over a period of time, a following vehicle classification model is used to discriminate the following vehicle category based on the following vehicle motion information, where the vehicle categories include AV and HV. The feature vectors used to discriminate the following vehicle category include the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments.

[0008] (3) combining the vehicle category with the vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a period of time in the future through the car-following prediction model; the car-following prediction model is used to predict the vehicle-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle category determined by the car-following vehicle classification model;

[0009] (4) Based on a one-time predicted acceleration or speed sequence, the abnormal speed of the vehicle is identified, or the vehicle is managed and controlled in the AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model; in the automatic driving mixed traffic flow following abnormal speed identification algorithm, the predicted following speed of the vehicle and the actual collected speed are compared, and the error value between the two is calculated and compared with the set threshold, thereby realizing the identification of abnormal collected data; wherein the predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

[0010] Preferably, the following vehicle classification model is a following vehicle classification model based on an Attention-LSTM model. The Attention-LSTM model is composed of multiple LSTM units connected together, and an Attention module is added before the model output.

[0011] Preferably, the method for constructing the car-following vehicle classification model comprises the following steps:

[0012] (1) Obtain the motion information of the leading car and the following car in the AV-MTF for T time steps from time t-T+1 to time t, including the speed v of the leading car l , following car speed v f , acceleration of the following car a f , the distance between the two vehicles Δx, and the relative speed Δv;

[0013] (2) Based on the Attention-LSTM neural network algorithm, a car-following vehicle classification model is constructed. Appropriate input information is selected from the relevant motion information of the lead vehicle and the following vehicle. After the model F1(·) is trained and calculated, the AV-MTF car-following vehicle category is determined:

[0014] Ω=F1(I1)

[0015] In the formula, the pilot car speed v is selected l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles constitutes the feature vector I1 of the following vehicle category, Ω represents the category of the following vehicle in AV-MTF, Ω∈{0,1}, where 0 represents AV and 1 represents HV, and F1(·) is the Attention-LSTM model.

[0016] Preferably, the input of the vehicle following prediction model is:

[0017] I3=Concat(I2,Ω)

[0018] In the formula, select the following car speed v f, the distance between the two vehicles Δx, and the relative speed Δv constitute the characteristic vector I2 of the following vehicle motion information, Concat(·) represents the connection, and I3 is the input information of the vehicle following prediction model.

[0019] Preferably, the vehicle following prediction model is implemented based on a Transformer model, and the model construction method includes the following steps:

[0020] (1) The AV-MTF vehicle following prediction model is constructed based on the Transformer algorithm. Appropriate input information is selected from the relevant motion information of the lead vehicle and the following vehicle to form the input information of the following speed prediction module. The input information must include historical information for a period of time before the selected parameter t:

[0021] I3=Concat(I2,Ω)

[0022] (2) After the Transformer model is trained and calculated, the predicted vehicle following speed is output. This study also considers the influence of the following vehicle category and inputs it into the model to establish the AV-MTF following speed vehicle following acceleration, speed and trajectory prediction model:

[0023] O1=F2(I3)

[0024] Where O1 is the predicted vehicle following acceleration, velocity, and trajectory sequence, and F2(·) is the Transformer model.

[0025] (3) The output of the prediction model is the acceleration, speed or trajectory sequence of the following vehicle. If the output is the following acceleration sequence, the speed and trajectory of the vehicle can be calculated based on the acceleration of the following vehicle:

[0026] v i+1 =v i +a i Δt

[0027] x i+1 =x i +v i Δt

[0028] Where a i is the vehicle following acceleration at time i, Δt is the time interval between data sampling and model output, v i is the vehicle following speed at time i, x i is the vehicle position at time i.

[0029] If the vehicle following speed is output, the vehicle trajectory can be calculated based on the vehicle following speed sequence:

[0030] x i+1 =x i +vi Δt

[0031] The model can also directly output the vehicle's future trajectory x i+1 ,...,x i+L .

[0032] Preferably, in step (4), based on the AV-MTF vehicle following speed anomaly recognition algorithm, the recognition and recording of abnormal collected data are realized, including the following steps:

[0033] (1) Based on a one-time prediction of acceleration or speed sequence, the predicted vehicle following speed and the actual collected speed are compared moment by moment, and the error between the two is calculated. The AV-MTF vehicle following speed anomaly identification algorithm uses the absolute percentage error between the predicted vehicle following speed and the actual collected speed as the measurement indicator for anomaly identification, and calculates the absolute percentage error between the predicted value and the actual value:

[0034]

[0035] Where, To predict the following speed of vehicle at time t, is the actual collected following speed at time t, and d(t) is the absolute percentage error at time t;

[0036] (2) During the recognition process, the deviation between the current speed prediction value and the actual collected value is calculated. When the absolute percentage error d(t) between the predicted value and the actual value exceeds the set threshold α, it means that the deviation between the predicted value and the actual value is large. The maximum number of iterations of the abnormal recognition algorithm is set to the length of the prediction sequence. When abnormal data is recognized, the iteration process ends and the time corresponding to the abnormal data is output. In this way, the abnormal collected data can be detected and recognized, and the time corresponding to the abnormal data is recorded. The judgment method is:

[0037]

[0038] Where α is the threshold for setting the average percentage error, R(t) is the abnormal data recognition result, 0 represents that the data at that moment is abnormal data, and 1 represents that the data at that moment is normal data.

[0039] Based on the same inventive concept, the present invention provides a system for predicting and identifying the following speed of vehicles in mixed traffic flows under autonomous driving, comprising:

[0040] The motion information acquisition module is used to obtain the motion information of the lead vehicle and the following vehicle in the AV-MTF;

[0041] A vehicle classification module is used to determine the following vehicle type based on the following vehicle motion information acquired over a period of time using a following vehicle classification model. The vehicle types include AVs and HVs. The feature vectors used to determine the following vehicle type include the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments.

[0042] a prediction module for combining vehicle classification with vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a future period of time through the car-following prediction model; the car-following prediction model for predicting the car-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle classification determined by the car-following vehicle classification model;

[0043] The post-processing module is used to identify abnormal vehicle speeds based on a one-time predicted acceleration or speed sequence, or to manage and control vehicles in an AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model. In the autonomous driving mixed traffic flow following abnormal speed identification algorithm, the vehicle's predicted following speed and actual collected speed are compared, and the error value between the two is calculated and compared with a set threshold to achieve the identification of abnormal collected data. The predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

[0044] The present invention also provides a computer system comprising a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor. When the computer program / instruction is executed by the processor, the steps of the autonomous driving mixed traffic flow vehicle following prediction method are implemented.

[0045] The present invention also provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the automatic driving mixed traffic flow vehicle following prediction method.

[0046] Beneficial Effects: Compared with existing technologies, this invention offers the following advantages: It establishes a car-following vehicle classification model and a car-following prediction model for AV-MTF scenarios involving both AVs and HVs, and proposes an algorithm for identifying abnormal car-following speeds. By combining the car-following vehicle classification model with the speed car-following prediction model, this invention effectively improves the accuracy of car-following acceleration, speed, or trajectory predictions. Furthermore, the abnormal car-following speed identification algorithm demonstrates a high accuracy rate for identifying abnormal speeds. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a method for predicting the following speed of vehicles in mixed traffic flow and identifying anomalies in an autonomous driving system, provided by an embodiment of the invention;

[0048] Figure 2 Flowchart of the AV-MTF vehicle following speed anomaly identification algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0050] Embodiments of the present invention provide a method for predicting vehicle following in mixed traffic flows for autonomous driving. This method, based on a deep learning algorithm, determines the following vehicle category and predicts the following vehicle's acceleration, speed, or trajectory sequence based on following vehicle motion information. This method enables the identification, management, and control of vehicle following speed anomalies. By correctly determining the following vehicle category and inputting this key information into the model, it improves the prediction of vehicle following acceleration, speed, or trajectory, thereby increasing the accuracy of speed anomaly identification and enabling better vehicle management and control in AV-MTF environments.

[0051] See also Figure 1 The method for predicting a vehicle following a mixed traffic flow in an autonomous driving system provided by an embodiment of the present invention specifically includes the following steps:

[0052] (1) Obtaining the motion information of the lead vehicle and the following vehicle in the AV-MTF;

[0053] (2) After obtaining the following vehicle information over a period of time, a following vehicle classification model is used to discriminate the following vehicle category based on the following vehicle motion information, where the vehicle categories include AV and HV. The feature vectors used to discriminate the following vehicle category include the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments.

[0054] (3) combining the vehicle category with the vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a period of time in the future through the car-following prediction model; the car-following prediction model is used to predict the vehicle-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle category determined by the car-following vehicle classification model;

[0055] (4) Based on a one-time predicted acceleration or speed sequence, the abnormal speed of the vehicle is identified, or the vehicle is managed and controlled in the AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model; in the automatic driving mixed traffic flow following abnormal speed identification algorithm, the predicted following speed of the vehicle and the actual collected speed are compared, and the error value between the two is calculated and compared with the set threshold, thereby realizing the identification of abnormal collected data; wherein the predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

[0056] The following describes the steps of an embodiment of the present invention in detail by using an Attention-LSTM-Transformer-based autonomous driving mixed traffic flow vehicle following prediction method, which is implemented by combining a following vehicle classification model based on the Attention-LSTM algorithm with a vehicle following prediction model based on the Transformer algorithm.

[0057] The Attention-LSTM-Transformer-based method for predicting car-following in mixed traffic flow for autonomous driving described in an embodiment of the present invention is applied to a car-following scenario and specifically includes the following steps:

[0058] (1) The motion information of the lead vehicle and the following vehicle in the AV-MTF is obtained through the roadside equipment, including the speed, acceleration, and distance between the two vehicles. The index numbers of the collected data are t-T+1,...,t in chronological order, with a total of T moments of motion information. The data collection interval is 0.1s.

[0059] (2) AVs and HVs have different following motion characteristics. After obtaining the following motion information over a period of time, the Attention-LSTM following vehicle classification model can be used to discriminate the AV-MTF following vehicle category based on the vehicle following motion information. The vehicle category includes AVs and HVs.

[0060] ΩΔF1(I1)

[0061] In the formula, the pilot car speed v is selected l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles constitutes the feature vector I1 of the following vehicle category, Ω represents the category of the following vehicle in AV-MTF, Ω∈{0,1}, where 0 represents AV and 1 represents HV, and F1(·) is the Attention-LSTM model.

[0062] LSTM is a special type of recurrent neural network (RNN) that is often used to process and predict time series data. By designing the memory unit and gating mechanism within the LSTM unit, it effectively solves the vanishing and exploding gradient problems of traditional RNNs when processing sequences.

[0063] Given the input x at time t t and the hidden state h at the previous moment t-1 , then the input gate i t 、Forget Gate t , output gate o tThe specific calculation method is as follows:

[0064] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )

[0065] f t =σ(W if x t +b if +W hf h t-1 +b hf )

[0066] o t =σ(W io x t +b io +W ho h t-1 +b ho )

[0067] Where W and b are weight and bias terms respectively, and σ is the sigmoid activation function.

[0068] The candidate g at time t in the model t , cell state c t and hidden state h t , which can be calculated by the following formula:

[0069] g t =σ(W ig x t +b ig +W hg h t-1 +b hg )

[0070] c t =f t ⊙c t-1 +i t ⊙g t

[0071] h t =o t ⊙tanh(c t )

[0072] Where tanh(·) is the hyperbolic tangent function, W and b are the weight and bias terms respectively, and ⊙ is the Hadamard product.

[0073] (3) The following vehicle category is combined with the vehicle motion information as the input of the Transformer following speed vehicle prediction model, and the vehicle following prediction model outputs the predicted acceleration, speed or trajectory sequence for a period of time in the future. This study also considers the influence of the following vehicle category and inputs it into the model:

[0074] I3=Concat(I2,Ω)

[0075] In the formula, select the following car speed v f , the distance between the two vehicles Δx, and the relative speed Δv constitute the characteristic vector I2 of the following vehicle motion information, Concat(·) represents the connection, and I3 is the input information of the speed vehicle following prediction model.

[0076] After the Transformer model training and calculation, the predicted vehicle following acceleration, speed or trajectory sequence is output.

[0077] The index numbers of the output sequence are t+1,...,t+L in chronological order, and the time interval of the output data is 0.1s.

[0078] O1=F2(I3)

[0079] Where O1 is the predicted vehicle following acceleration, velocity or trajectory sequence, and F2(·) is the Transformer model.

[0080] The Transformer architecture abandons recurrence and convolution and is based entirely on an attention mechanism. This allows it to simultaneously consider all positions in the input sequence, thereby better capturing long-range dependencies. The Transformer has an encoder-decoder architecture. The encoder consists of a multi-head attention layer and a feedforward neural network. Information passing through these two sub-layers is summed and normalized. The decoder also has a similar structure consisting of a multi-head attention layer and a feedforward neural network. However, the decoder receives not only the decoder input information but also the encoder output information.

[0081] The multi-head attention module is the most important mechanism of the Transformer. When processing sequential data, it assigns different attention weights to information at different positions, enabling it to extract important information from the input. The multi-head attention module is composed of h scaled dot-product attention (SDPA) modules connected together. Scaled dot-product attention is the transposed dot product of the query matrix Q and the key matrix K. It calculates the correlation between each vector and weights the value matrix V according to the matching degree between Q and K. Its output is as follows:

[0082]

[0083] Where Q(Query), K(Key) and V(Value) are query, key and value matrices respectively. Refers to the square root of the bond matrix.

[0084] Multi-head attention can calculate the relevance of any position in the sequence. Compared with SDPA, the multi-head attention module focuses on the global information of long sequences more effectively. After Q, K, and V are input to SDPA, this process is repeated h times. Finally, they are concatenated through the Concat layer to obtain a global self-attention feature map. The first multi-head attention module of the Transformer model decoder enables a mask to prevent the decoder from accessing the target sequence information after the position when generating the output of each position. The specific formula is:

[0085] Multihead(Q,K,V)=Concat(head1,…,head h )W 0

[0086] head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h

[0087] Where, head i Represents the self-attention of the i-th head, the parameter matrix where d model is the model dimension, d q is the dimension of the query vector, d k is the dimension of the key vector, d vis the dimension of the Value Vector.

[0088] Add&Norm consists of two parts: summation and normalization. It is used after both multi-head attention and feedforward neural networks. Add refers to adding the output result to the input information, and Norm refers to LayerNormalization. Taking the summation and normalization after the multi-head attention layer as an example, the calculation method is as follows:

[0089] Output=LayerNorm(X+MultiHeadAttention(X))

[0090] Where X represents the input of the multi-head attention layer, and MultiHeadAttention(X) is the output of X through the multi-head attention layer.

[0091] The inputs to the Transformer encoder and decoder must pass through an embedding module. This is typically achieved using a linear layer when building an AV-MTF car-following prediction model. Unlike RNNs, the Transformer model cannot utilize sequential input information, which is crucial for AV-MTF car-following speed prediction modeling. Therefore, the Transformer model uses positional encoding to implement sequential input information and adds the positional encoding to the input and output embeddings. The calculation formulas are:

[0092]

[0093]

[0094] In the formula, pos represents the position of the input information in the world, 2i refers to the even dimension, and 2i+1 refers to the odd dimension. Since the longest sequence length of input and output in this study is 50, d model A smaller value is better, such as 100. The input I3 of the AV-MTF vehicle following prediction model encoder is expressed as follows:

[0095]

[0096] If the sampling interval of the vehicle following motion features in the dataset is 0.1s, the length of the input vehicle motion feature sequence T is set to 50, that is, Δx, Δv, v and Ω within t-5 to t-0.1 are combined into a parameter matrix of dimension 50×4 as the input to the encoder of the Transformer model.

[0097] During the model training process, the decoder uses the following vehicle's future speed sequence v(t+1),...,v(t+L) as the model's output, and uses a masking mechanism to prevent the decoder from accessing information after that moment when outputting information. The decoder's output is also the predicted future vehicle following speed sequence. The input future speed sequence length L is set to 5. During the model prediction stage, the decoder no longer inputs the following vehicle's future speed sequence information, and uses an autoregressive mechanism to gradually decode and generate the future vehicle following acceleration, speed, or trajectory sequence. During the model testing process, assuming that t is the current moment, the input historical information is the parameter matrix of T moments from t-T+1 to t, and the output is the acceleration, speed, or trajectory of L time steps from t+1 to t+L. The formula is as follows:

[0098]

[0099] Where F(·) represents the Transformer model of data input and is also expressed as the mapping relationship between input and output variables. The length of the output vehicle speed sequence L is 5, that is, the model outputs the vehicle following acceleration sequence for the next 0.5 seconds as The vehicle following speed sequence is The vehicle following trajectory sequence is

[0100] If the output vehicle following acceleration sequence The speed and trajectory of the vehicle can be calculated based on the acceleration of the following vehicle; if the output vehicle following speed sequence The trajectory of the following vehicle can be calculated based on its speed; the model can also directly output the future trajectory of the vehicle The formula is as follows:

[0101] v i+1 =v i +a i Δt

[0102] x i+1 =x i +v i Δt

[0103] Where a i is the vehicle following acceleration at time i, Δt is the time interval between data sampling and model output, v i is the vehicle following speed at time i, x i is the vehicle position at time i.

[0104] (4) In one-time prediction of speed sequence Based on (can also be based on the acceleration sequence The AV-MTF vehicle following speed anomaly recognition algorithm will predict the vehicle following speed. and the actual collected following speed The absolute percentage error is used as a measure of anomaly identification, and the absolute percentage error between the predicted value and the actual value is calculated;

[0105]

[0106] Where, To predict the following speed of vehicle at time t, is the actual collected following speed at time t, and d(t) is the absolute percentage error at time t.

[0107] During the recognition process, the deviation between the current speed prediction value and the actual collected value is calculated. When the absolute percentage error d(t) between the predicted value and the actual value exceeds the set threshold, it means that the deviation between the predicted value and the actual value is large, indicating that abnormal data has been collected during the vehicle following process. The judgment method is as follows:

[0108]

[0109] Where α is the threshold for setting the average percentage error, R(t) is the abnormal data recognition result, 0 represents that the data at that moment is abnormal data, and 1 represents that the data at that moment is normal data.

[0110] The maximum number of iterations of the anomaly recognition algorithm is set to the length of the prediction sequence. The present invention sets the maximum number of iterations to 5. When abnormal data is identified, the iterative process ends and the time corresponding to the abnormal data is output. Figure 2 The AV-MTF vehicle following speed anomaly identification algorithm process is illustrated.

[0111] This embodiment of the present invention combines a car-following vehicle classification model with a car-following acceleration, speed, or trajectory prediction model. Based on the Attention-LSTM-Transformer model, it achieves short-term speed prediction errors on the order of centimeters per second, avoiding the performance degradation of traditional car-following models when applied to AV-MTF environments. Furthermore, the Transformer model is more effective than traditional recurrent neural networks in handling long-term dependencies, ensuring the stability and accuracy of the car-following prediction model. The prediction model can output different motion sequence information based on actual needs, specifically including the vehicle's car-following acceleration, speed, and trajectory. Furthermore, these information can be inferred from each other using formulas, thus providing the model with a richer range of application scenarios and more flexible usage. Based on this, a method for identifying car-following speed anomalies in autonomous driving mixed traffic flows can be constructed. This method adaptively detects and identifies abnormal collected data and records the time at which the abnormal data occurs, offering greater flexibility and efficiency than other methods. The output vehicle acceleration, speed, and trajectory sequences can be used for vehicle management and control research in AV-MTF environments.

[0112] Based on the same inventive concept, an embodiment of the present invention provides an autonomous driving mixed traffic flow vehicle following prediction system, comprising:

[0113] The motion information acquisition module is used to obtain the motion information of the lead vehicle and the following vehicle in the AV-MTF;

[0114] A vehicle classification module is used to determine the following vehicle type based on the following vehicle motion information acquired over a period of time using a following vehicle classification model. The vehicle types include AVs and HVs. The feature vectors used to determine the following vehicle type include the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments.

[0115] a prediction module for combining vehicle classification with vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a future period of time through the car-following prediction model; the car-following prediction model for predicting the car-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle classification determined by the car-following vehicle classification model;

[0116] The post-processing module is used to identify abnormal vehicle speeds based on a one-time predicted acceleration or speed sequence, or to manage and control vehicles in an AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model. In the autonomous driving mixed traffic flow following abnormal speed identification algorithm, the vehicle's predicted following speed and actual collected speed are compared, and the error value between the two is calculated and compared with a set threshold to achieve the identification of abnormal collected data. The predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

[0117] Based on the same inventive concept, an embodiment of the present invention provides a computer system, including a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor. When the computer program / instruction is executed by the processor, the steps of the autonomous driving mixed traffic flow vehicle following prediction method are implemented.

[0118] Based on the same inventive concept, an embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the autonomous driving mixed traffic flow vehicle following prediction method.

Claims

1. A method for predicting vehicle following in mixed traffic flow during autonomous driving, characterized in that: The steps include: (1) Obtaining motion information of the leading vehicle and the following vehicle in the autonomous driving mixed traffic flow (AV-MTF); (2) After obtaining the following information within a period of time, the following vehicle classification model is used to determine the following vehicle category based on the following vehicle motion information, wherein the vehicle category includes autonomous driving vehicles (AV) and manually driven vehicles (HV); the feature vector used to determine the following vehicle category includes the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments; the following vehicle classification model is a following vehicle classification model based on the Attention-LSTM model, which is composed of multiple LSTM units connected together, and an Attention module is added before the model output; the construction method includes: obtaining the motion information of the lead vehicle and the following vehicle in the AV-MTF of T time steps from time t-T+1 to time t, including the speed of the lead vehicle v l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles, and the relative speed Δv; a following vehicle classification model is constructed based on the Attention-LSTM neural network algorithm. Appropriate input information is selected from the relevant motion information of the lead vehicle and the following vehicle. After the model F1(·) is trained and calculated, the AV-MTF following vehicle category is determined: Ω=F1(I1); where the lead vehicle speed v is selected. l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles constitutes the feature vector I1 of the following vehicle category, Ω represents the category of the following vehicle in AV-MTF, Ω∈{0,1}, where 0 represents AV and 1 represents HV, and F1(·) is the Attention-LSTM model; (3) combining the vehicle category with the vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a period of time in the future through the car-following prediction model; the car-following prediction model is used to predict the vehicle-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle category determined by the car-following vehicle classification model; (4) Based on a one-time predicted acceleration or speed sequence, the abnormal speed of the vehicle is identified, or the vehicle is managed and controlled in the AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model; in the automatic driving mixed traffic flow following abnormal speed identification algorithm, the predicted following speed of the vehicle and the actual collected speed are compared, and the error value between the two is calculated and compared with the set threshold, thereby realizing the identification of abnormal collected data; wherein the predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

2. The method for predicting car-following in mixed traffic flow under autonomous driving according to claim 1, characterized in that: Inputs of the car-following prediction model: I3=Concat(I2,Ω) In the formula, select the following car speed v f , the distance between the two vehicles Δx, and the relative speed Δv constitute the feature vector I2 of the following vehicle motion information. Ω represents the category of the following vehicle in AV-MTF, Concat(·) represents the connection, and I3 is the input information of the vehicle following prediction model.

3. The method for predicting car-following in mixed traffic flow under autonomous driving according to claim 1, characterized in that: The vehicle following prediction model is implemented based on the Transformer model. The model construction method includes: The AV-MTF car-following prediction model is constructed based on the Transformer algorithm. Appropriate input information is selected from the relevant motion information of the lead vehicle and the following vehicle to form the input information of the car-following prediction model. The input information must include historical information for a period of time before the selected parameter t: I3=Concat(I2,Ω) In the formula, select the following car speed v f , the distance between the two vehicles Δx, and the relative speed Δv constitute the feature vector I2 of the following vehicle motion information, Ω represents the category of the following vehicle in AV-MTF, Concat(·) represents the connection, and I3 is the input information of the vehicle following prediction model; After the Transformer model is trained and calculated, the predicted vehicle following acceleration, speed, or trajectory is output to establish the AV-MTF vehicle following prediction model: O1=F2(I3) Where O1 is the sequence of predicted vehicle following acceleration, speed or trajectory, and F2(·) is the Transformer model; If the vehicle following acceleration sequence is output, the vehicle speed and trajectory can be calculated based on the acceleration of the following vehicle: v i+1 =v i +a i Δt x i+1 =x i +v i Δt Where a i is the vehicle following acceleration at time i, Δt is the time interval between data sampling and model output, v i is the vehicle following speed at time i, x i is the vehicle position at time i; If the vehicle following speed is output, the vehicle trajectory can be calculated based on the vehicle following speed sequence: x i+1 =x i +v i Δt The model can also directly output the future trajectory of the vehicle.

4. The method for predicting car-following in mixed traffic flow under autonomous driving according to claim 1, characterized in that: In step (4), the error value between the predicted following speed and the actual collected speed of the vehicle is calculated and compared with the set threshold value to realize the identification and recording of abnormal collected data, including the following steps: Based on a one-time prediction of acceleration or speed sequence, the predicted vehicle following speed is compared with the actual collected speed at each moment, and the error between the two is calculated. The absolute percentage error between the predicted vehicle following speed and the actual collected speed is used as a metric for anomaly identification, and the absolute percentage error between the predicted and actual values ​​is calculated. Where, To predict the vehicle following speed at time t, is the actual collected following speed at time t, and d(t) is the absolute percentage error at time t; When the error value exceeds the set threshold, the abnormal collected data is detected and identified, and the time when the abnormal data appears is recorded; the judgment method is: Where α is the threshold for setting the average percentage error, R(t) is the abnormal data recognition result, 0 represents that the data at that moment is abnormal data, and 1 represents that the data at that moment is normal data.

5. The method for predicting car-following in mixed traffic flow under autonomous driving according to claim 4, characterized in that: On the basis of a one-time prediction of the acceleration or velocity sequence, the maximum number of iterations is set to the length of the predicted acceleration or velocity sequence. When abnormal data is identified, the iteration process is terminated and the time corresponding to the abnormal data is output.

6. An autonomous driving mixed traffic flow vehicle following prediction system, characterized by: include: The motion information acquisition module is used to obtain the motion information of the lead vehicle and the following vehicle in the autonomous driving mixed traffic flow AV-MTF; The vehicle category discrimination module is used to discriminate the category of the following vehicle using a following vehicle classification model based on the vehicle following motion information after obtaining the following information within a period of time, wherein the vehicle categories include autonomous driving vehicles (AV) and manually driven vehicles (HV); the feature vector used to discriminate the category of the following vehicle includes the speed of the lead vehicle, the speed of the following vehicle, the acceleration of the following vehicle, and the distance between the two vehicles at multiple moments; the following vehicle classification model is a following vehicle classification model based on the Attention-LSTM model, which is composed of multiple LSTM units connected together, and an Attention module is added before the model output; the construction method includes: obtaining the motion information of the lead vehicle and the following vehicle in the AV-MTF for T time steps from time t-T+1 to time t, including the speed of the lead vehicle v l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles, and the relative speed Δv; a following vehicle classification model is constructed based on the Attention-LSTM neural network algorithm. Appropriate input information is selected from the relevant motion information of the lead vehicle and the following vehicle. After the model F1(·) is trained and calculated, the AV-MTF following vehicle category is determined: Ω=F1(I1); where the lead vehicle speed v is selected. l , following car speed v f , acceleration of the following car a f , the distance Δx between the two vehicles constitutes the feature vector I1 of the following vehicle category, Ω represents the category of the following vehicle in AV-MTF, Ω∈{0,1}, where 0 represents AV and 1 represents HV, and F1(·) is the Attention-LSTM model; a prediction module for combining vehicle classification with vehicle motion information as input to a car-following prediction model, and outputting a predicted acceleration, speed, or trajectory sequence for a future period of time through the car-following prediction model; the car-following prediction model for predicting the car-following acceleration, speed, or trajectory sequence based on the vehicle motion information over a period of time and the vehicle classification determined by the car-following vehicle classification model; The post-processing module is used to identify abnormal vehicle speeds based on a one-time predicted acceleration or speed sequence, or to manage and control vehicles in an AV-MTF environment based on the vehicle trajectory output by the vehicle following prediction model. In the autonomous driving mixed traffic flow following abnormal speed identification algorithm, the vehicle's predicted following speed and actual collected speed are compared, and the error value between the two is calculated and compared with a set threshold to achieve the identification of abnormal collected data. The predicted following speed is directly predicted by the vehicle following prediction model or calculated from the following acceleration predicted by the vehicle following prediction model.

7. A computer system comprising a memory, a processor, and a computer program / instruction stored in the memory and executable on the processor, wherein: When the computer program / instructions are executed by a processor, the steps of the automatic driving mixed traffic flow vehicle following prediction method are implemented according to any one of claims 1 to 5.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the automatic driving mixed traffic flow vehicle following prediction method are implemented according to any one of claims 1 to 5.

Images