Vehicle trajectory reconstruction method suitable for multiple scenes of mixed traffic flow

Abstract

The invention discloses a vehicle trajectory reconstruction method suitable for multiple scenes of mixed traffic flow. The method comprises the following steps: S1, collecting information such as positions and speeds of surrounding manually-driven vehicles; and S2, dividing the reconstruction scene into a full-perception scene, a semi-perception scene and a conventional traffic flow scene according to the types of adjacent floating vehicles, and generating research scene data by considering the permeability of the floating vehicles. And S3, generating a time sequence by adopting a sliding window technology. And S4, in a conventional traffic flow scene, using an intelligent driver model IDM to reconstruct the vehicle under different floating vehicle permeability. And S5, in a semi-perception scene, a fusion model IDM-LSTM is adopted to reconstruct a vehicle track with fragmented observation, and an IDM model is adopted to reconstruct the track of a vehicle with a large amount of track information missing. And S6, in a full-perception scene, obtaining fragmented observation trajectories of all vehicles, introducing a weight module to fuse the IDM model and the long-short-term memory neural network, and reconstructing a missing vehicle trajectory.

Description

Technical Field

[0001] The present invention belongs to the field of traffic perception and big data analysis, and specifically relates to a vehicle trajectory reconstruction method adapted to multiple scenarios of mixed traffic flows. Background Art

[0002] Vehicle trajectories contain rich spatiotemporal information about traffic flows and are widely used in driving pattern recognition, traffic safety, traffic state estimation, traffic flow models, signal optimization and control, and traffic energy consumption estimation. Obtaining complete vehicle trajectories is crucial for accurately describing traffic flow states and realizing the above applications. However, raw data usually contain a large number of missing segments, which poses a challenge to its integration into traffic flow models.

[0003] Currently, a large number of detection devices are installed to monitor traffic conditions. These detection devices are mainly divided into fixed detectors and mobile detectors. Studies have shown that by 2050, the proportion of connected autonomous vehicles will exceed 20%, which indicates that the future mixed traffic flow will be composed of manually driven vehicles, connected vehicles and autonomous vehicles. In mixed traffic flow, autonomous vehicles and connected vehicles can collect complete trajectory information of vehicles, overcoming the defect that fixed detectors can only obtain single-point traffic information. Based on the above advantages, autonomous vehicles and connected vehicles have promoted the development of applications such as following behavior simulation, basic map establishment, and overtaking behavior. However, due to the low penetration rate of autonomous vehicles and connected vehicles, they can only provide fragmented observation results rather than complete traffic trajectory information. The limited availability of observation data limits the improvement of the accuracy of trajectory reconstruction models.

[0004] The above analysis shows that reconstructing vehicle trajectories in mixed traffic flows is far from easy. The main challenges include: 1) Fixed detectors can only provide traffic status information at specific locations, and the low penetration rate of autonomous vehicles makes the trajectories of most manually driven vehicles undetectable; 2) The data formats obtained by fixed detectors and mobile detectors are inconsistent, and the differences between the two types of data need to be carefully considered to achieve irregular trajectory reconstruction.

[0005] Current trajectory reconstruction methods are mainly divided into data-driven methods and traffic flow theory-based methods. For data-based methods, the focus is on inferring the spatiotemporal distribution pattern of trajectory points based on the characteristics of trajectory data. Machine learning-based methods are one of the common methods and perform well in short-term trajectory prediction, but their prediction errors tend to increase as the length of the trajectory segment to be reconstructed increases. Traffic flow theory-based methods have better interpretability of driving behavior, but the reconstruction accuracy is not as high as that based on data-driven methods. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides a vehicle trajectory reconstruction method that is adaptable to multiple scenarios of mixed traffic flow. In a mixed traffic flow environment, the present invention takes highways as the research object, collects information such as the trajectory position, speed and passing time of surrounding manually driven vehicles through fixed roadside detectors and autonomous driving vehicles; divides the research scenarios into full perception scenarios, semi-perception scenarios and conventional traffic flow scenarios; and adopts different trajectory reconstruction methods for different trajectory reconstruction scenarios, so that the reconstructed vehicle trajectory has higher accuracy.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a vehicle trajectory reconstruction method adapted to multiple scenarios of mixed traffic flow, comprising the following steps:

[0008] S1. Collect the trajectory position, speed and passing time information of surrounding manually driven vehicles;

[0009] S2. Divide the target scene into three scenarios according to the type of adjacent floating vehicles: full perception scenario, semi-perception scenario and regular traffic flow scenario, consider the penetration rate of floating vehicles and generate trajectory data for the three scenarios;

[0010] S3, using a sliding window approach to divide vehicle trajectories, and dividing the data set into training set, validation set, and test set according to a specific ratio;

[0011] S4. In the conventional traffic flow scenario, the trajectory data of the floating vehicle itself is obtained, and the vehicle is reconstructed using the intelligent driver model under different floating vehicle penetration rates;

[0012] S5. In the semi-perception scenario, the vehicle with fragmented trajectory uses the fusion model IDM-LSTM to reconstruct the trajectory, and the vehicle with missing trajectory information uses the intelligent driver model to reconstruct the trajectory;

[0013] S6. In the full perception scenario, the fragmentary trajectory information of all vehicles is obtained, and a weight module is introduced to fuse the intelligent driver model with the long short-term memory neural network, namely IDM-LSTM, to reconstruct the missing vehicle trajectory.

[0014] Furthermore, in the aforementioned step S1, a vehicle passing through the target road section is detected by the fixed detector at the starting point of the road side and numbered as i, and the time when vehicle i passes the fixed detector position at the starting point is recorded as b i , the time it takes for vehicle i to pass the end of the study section is e i , retrieve the vehicle trajectory according to the vehicle number, and obtain the vehicle trajectory observation set Tra of vehicle i on the target road section i ={s(i,t),v(i,t),a(i,t),l(i,t),t}t∈[b i ,e i ],

[0015] The trajectory set of all vehicles Tra = {Tra1,…,Tra k ,…,Tra n}, where s(i,t) is the longitudinal distance traveled by vehicle i on the target section at time t, v(i,t) is the speed of vehicle i on the study section at time t, a(i,t) is the acceleration of vehicle i on the study section at time t, and l(i,t) is the lane number of vehicle i on the study section at time t.

[0016] Furthermore, in the aforementioned step S2, the target scene is divided into a full perception scene, a semi-perception scene and a regular traffic flow scene according to the types of adjacent floating vehicles. Specifically, the types of the two adjacent floating vehicles are determined in the target scene. If the two adjacent floating vehicles are both autonomous driving vehicles, the target scene is named a full perception scene; if one is an autonomous driving vehicle and the other is a manually driven vehicle that can only obtain its own vehicle driving status information, the target scene is named a semi-perception scene; if the two adjacent floating vehicles are both manually driven vehicles, the target scene is named a regular traffic flow scene.

[0017] Furthermore, in the aforementioned step S2, the penetration rate of the floating vehicle is considered and three scenario trajectory data are generated, as follows:

[0018] Based on the fact that the floating cars are evenly distributed, the interval d between the floating cars ij As follows:

[0019] d ij =100*p r (1)

[0020] In the formula, p r is the penetration rate of autonomous vehicles, d ij represents the interval between floating cars i and j, and introduces a random factor u to consider the non-uniform distribution between floating cars.

[0021]

[0022] In the formula, p i is a probability generated by a random program. If this probability is less than the random factor, the current vehicle is set as a floating vehicle. i =1, otherwise it is not a floating car f i =0.

[0023] Furthermore, the aforementioned step S3 includes the following sub-steps:

[0024] S3.1. Use sliding windows to divide vehicle trajectories, and select a history sequence with a length of l. h , the reconstruction length is lp , then the time segments of the historical sequence Tra(h,i) and the reconstructed sequence Tra(p,i) of the i-th sample are as follows:

[0025] Tra(h,i)∈[t start ,t start +l h ] (3)

[0026] Tra(p,i)∈[t start +l h +1,t start +l h +l p ] (4)

[0027] The vehicle's position, speed, acceleration, lane and other information are all within this time range, and the time series data All_Tra expression is obtained as follows:

[0028] All_Tra={Tra(h,1),Tra(p,1),…,Tra(h,k),Tra(p,k),…,Tra(h,n),Tra(p,n)} (5)

[0029] Where, t start is the start time of each sample;

[0030] S3.2. According to the stratified sampling principle, the time series data is divided into training set, validation set and test set according to the preset ratio, where Train_set, Val_set and Test_set represent the training set, validation set and test set respectively:

[0031] Train_set=m1*All_Tra (6)

[0032] Val_set=m2*All_Tra (7)

[0033] Test_set = m3*All_Tra (8)

[0034] Tra is the sample of all vehicle trajectories, and m1, m2, and m3 are the corresponding proportional coefficients.

[0035] Furthermore, the aforementioned step S4 includes the following steps:

[0036] S4.1. In the intelligent driver model, the leading vehicle and the following vehicle of vehicle i are determined by lane information, position information and time information, as shown in the following formula:

[0037]

[0038] Where is_front is used to indicate whether vehicle j is the leading vehicle in front of vehicle i, is_back is used to indicate whether vehicle j is the following vehicle of vehicle i, and a value of 1 indicates that it is a leading vehicle or a following vehicle, and a value of 0 indicates that it is not a leading vehicle or a following vehicle. S4.2. The intelligent driver model is used to estimate the missing vehicle trajectory information. The intelligent driver model IDM model is used to reconstruct other missing vehicle trajectory segments based on the observation segments of the leading vehicle. The intelligent driver model IDM is as follows:

[0039]

[0040] Δv′(i,t)=v′(i,t)-v′(i-1,t) (14)

[0041] d′(i,t)=s′(i-1,t)-s′(i,t)-L i-1 (15)

[0042] In the formula, s * (v′(i,t),Δv′(i,t)) represents the expected distance, d′(i,t) represents the headway between the front and rear vehicles, Δv′(i,t) represents the speed difference between the front and rear vehicles, and a max , b represent the maximum acceleration and comfortable deceleration respectively, T safe is the safe headway time, s0 is the headway distance at rest, v0 is the expected speed, L i-1 Indicates the vehicle length of the preceding vehicle;

[0043] S4.3. Adjust the reconstructed vehicle trajectory to ensure that the reconstructed vehicle trajectory is between the previous segment and the next segment, as shown in the following formula:

[0044] v′(i,t)=max(v′(i,t-1)+a′(i,t)Δt,0) (16)

[0045]

[0046] Furthermore, the aforementioned step S5 includes the following sub-steps:

[0047] S5.1. In the semi-perception scenario, some manually driven vehicle trajectories are reconstructed by the fusion model IDM-LSTM, and some vehicle trajectories are reconstructed by executing steps S4.1 to S4.3. The existing data is classified. When the historical sequence is missing, the intelligent driver model IDM is used for reconstruction; if the historical sequence data is complete, the fusion model IDM-LSTM is used for reconstruction;

[0048] S5.2. The fusion model IDM-LSTM is used for reconstruction. The LSTM model includes an input gate, a forget gate, and an output gate. The trajectory information is circulated between different gate structures, and the influence of the historical sequence on the reconstructed sequence is considered. The specific expressions are shown in equations (18) to (23):

[0049] f t =σ(W f ·[h t-1 ,x t ]+b f ) (18)

[0050] i t =σ(W i ·[h t-1 ,x t ]+b i ) (19)

[0051]

[0052] o t =σ(W o [h t-1 ,x t ]+b o ) (twenty two)

[0053] h t =o t *tanh(C t ) (twenty three)

[0054] where f t is the output of the forget gate, i t is the output of the input gate, is a candidate cell state, C t is the unit state at the current moment, h t is the hidden state at the current moment, x t Represents the input data; W f , W i , W C , W o Respectively represent the weight matrix of the forget gate, the weight matrix of the input gate, the weight matrix of the hidden state update, and the weight matrix of the output gate; b f , b i , b C , b o Respectively represent the bias vector of the forget gate, the bias vector of the input gate, the bias vector of the hidden state update, and the bias vector of the output gate;

[0055] S5.3 adopts the method of fusion model IDM-LSTM to realize the fusion of two models through the weight module. The weight module is shown in expressions (24)-(25). By fusing the reconstruction result of the IDM model and the reconstruction result of the LSTM model, the prediction result of the IDM-LSTM model is obtained. The prediction result is used to calculate the loss of the LSTM model, which is used to update the parameters in the training process and enhance the training process of the LSTM model.

[0056]

[0057] w IDM,i +w LSTM,i =1 (25)

[0058] Among them, s LSTM,i represents the reconstruction result of vehicle i by the LSTM model, s IDM,i represents the result of the intelligent driver model IDM reconstruction, w IDM,i ,w LSTM,i They represent the weight coefficients of the intelligent driver model IDM and LSTM model, representing the credibility of different models, w IDM,i ,w LSTM,i The sum of is 1;

[0059] S5.4. The model error is evaluated using three evaluation indicators: MAE, RMSE and MAPE, as shown in the following formula:

[0060]

[0061] Furthermore, the aforementioned step S6 includes the following sub-steps:

[0062] S6.1. In the full perception scenario, first use the intelligent driver model IDM to reconstruct the missing vehicle trajectory, complete the vehicle trajectory reconstruction based on the car-following rule, and execute steps S4.1 to S4.3;

[0063] S6.2, intelligent driver model IDM, the reconstructed result is used to input into the LSTM model to assist the training process of the LSTM model, and execute step S5.3;

[0064] S6.3, using the fusion model IDM-LSTM method, realize the fusion of the two models through the weight module, and execute step S5.4 to reconstruct the vehicle trajectory in the full perception scenario.

[0065] Compared with the prior art, the beneficial technical effects of the above technical solutions adopted in the present invention are as follows: Based on the mixed traffic flow environment, the present invention divides the trajectory reconstruction scenarios into three types: full perception scenarios, semi-perception scenarios and conventional traffic flow scenarios, and proposes a weighted module for combining the intelligent driver model and the long short-term memory network model, namely IDM-LSTM. In the conventional traffic flow scenario, the IDM model is used to reconstruct the vehicle trajectory; in the semi-perception scenario, the IDM model and the fusion model are used to reconstruct the vehicle trajectory respectively; in the full perception scenario, the fusion model is used to reconstruct the vehicle trajectory. The present invention proposes a vehicle trajectory reconstruction method suitable for multi-scenario mixed traffic flow. The method can accurately grasp the traffic flow state, which is of great significance for ensuring the smooth operation of highways and monitoring vehicle exhaust emissions. At the same time, the method also has significant academic value and application potential in the refinement of traffic state evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a multi-scenario diagram of mixed traffic flow; in the figure, (a) is a schematic diagram of a conventional traffic flow scenario, (b) is a schematic diagram of a semi-perception scenario, and (c) is a schematic diagram of a full-perception scenario.

[0067] Figure 2 It is a schematic diagram of the weight module.

[0068] Figure 3 It is a temporal and spatial distribution diagram of actual vehicle trajectories under the embodiment.

[0069] Figure 4 It is the vehicle trajectory reconstruction result diagram; in the figure, (a) is the vehicle trajectory reconstruction result diagram based on the intelligent driver model IDM, (b) is the vehicle trajectory reconstruction result diagram based on the LSTM model, and (c) is the vehicle trajectory reconstruction result diagram based on the IDM-LSTM fusion model.

[0070] Figure 5 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0071] In order to better understand the technical content of the present invention, specific embodiments are given and described as follows in conjunction with the accompanying drawings.

[0072] Various aspects of the invention are described herein with reference to the accompanying drawings, in which many illustrative embodiments are shown. The embodiments of the invention are not limited to those described in the accompanying drawings. It should be understood that the invention is implemented by any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the invention are not limited to any implementation. In addition, some aspects disclosed in the invention may be used alone or in any appropriate combination with other aspects disclosed in the invention.

[0073] refer to Figure 5 The present invention provides a vehicle trajectory reconstruction method adapted to multiple scenarios of mixed traffic flow, comprising the following steps:

[0074] S1. Collect the trajectory position, speed and passing time information of surrounding manually driven vehicles;

[0075] S2. Divide the target scene into three scenarios according to the type of adjacent floating vehicles: full perception scenario, semi-perception scenario and regular traffic flow scenario, consider the penetration rate of floating vehicles and generate trajectory data for the three scenarios;

[0076] S3, using a sliding window approach to divide vehicle trajectories, and dividing the data set into training set, validation set, and test set according to a specific ratio;

[0077] S4. In the conventional traffic flow scenario, the trajectory data of the floating vehicle itself is obtained, and the vehicle is reconstructed using the intelligent driver model under different floating vehicle penetration rates;

[0078] S5. In the semi-perception scenario, the vehicle with fragmented trajectory uses the fusion model IDM-LSTM to reconstruct the trajectory, and the vehicle with missing trajectory information uses the intelligent driver model to reconstruct the trajectory;

[0079] S6. In the full perception scenario, the fragmentary trajectory information of all vehicles is obtained, and a weight module is introduced to fuse the intelligent driver model with the long short-term memory neural network, namely IDM-LSTM, to reconstruct the missing vehicle trajectory.

[0080] like Figure 1 As shown in the figure, there are three trajectory reconstruction scenarios studied in this invention, in which (a) is a schematic diagram of a conventional traffic flow scenario, (b) is a schematic diagram of a semi-perception scenario, and (c) is a schematic diagram of a full-perception scenario. These scene divisions are the basis for the subsequent use of different trajectory reconstruction methods to reconstruct vehicle trajectories. Figure 2 The figure shows the model fusion module proposed in the present invention, which combines the IDM model and the LSTM model by weight. The combined model has a better trajectory reconstruction effect than a single model. By selecting appropriate trajectory reconstruction methods for different scenarios, the method proposed in the present invention can significantly reduce the error of trajectory reconstruction.

[0081] Further, as a preferred embodiment of the present invention, in step S1, a vehicle passing through the study section is detected by a fixed detector at the starting point of the road side and numbered as i, and the time when vehicle i passes the fixed detector position at the starting point is recorded as b i , the time it takes for vehicle i to pass the end of the study section is e i , retrieve the vehicle trajectory according to the vehicle number, and obtain the vehicle trajectory observation set of vehicle i on the research section:

[0082] Tra i ={s(i,t),v(i,t),a(i,t),l(i,t),t}t∈[b i ,e i ],

[0083] The trajectory set of all vehicles Tra = {Tra1,…,Tra k ,…,Tra n}, where s(i,t) is the longitudinal distance traveled by vehicle i on the study section at time t, v(i,t) is the speed of vehicle i on the study section at time t, a(i,t) is the acceleration of vehicle i on the study section at time t, and l(i,t) is the lane number of vehicle i on the study section at time t.

[0084] Furthermore, as a preferred embodiment of the present invention, the target scene is divided into a full perception scene, a semi-perception scene and a regular traffic flow scene according to the type of adjacent floating vehicles. Specifically, (1) If two adjacent floating vehicles are both autonomous vehicles that can perceive vehicles in the surrounding environment, all vehicles in this scene can obtain fragmented vehicle trajectory information, which is named a full perception scene. (2) If one of the two adjacent floating vehicles is an autonomous vehicle and the other is a manually driven vehicle that can only obtain its own vehicle driving status information, then in the area between the two vehicles, only some vehicles can obtain sufficient fragmented vehicle trajectory information, and the trajectory information of the other vehicles is largely missing, which is named a semi-perception scene. (3) If two adjacent floating vehicles are both manually driven vehicles, the trajectory information of other vehicles between the two floating vehicles is completely missing. In this scenario, it is more difficult to reconstruct the trajectory, which is named a regular traffic flow scene.

[0085] Furthermore, as a preferred embodiment of the present invention, the vehicle trajectory data of the full sample is obtained in the original data. Therefore, in order to simulate the data missing situation in the three trajectory reconstruction scenarios, it is necessary to generate trajectory data in the corresponding scenarios. In the present invention, the penetration rate of autonomous driving vehicles p r Set to 5%, 10% and 15%. Assuming that most of the floating cars are evenly distributed, the interval d between the floating cars ij As shown in expression (1). Taking randomness into consideration, a random factor u is introduced to consider the non-uniform distribution between floating cars, as shown in expression (2). Expressions (1) and (2) are parallel relationships, not inclusive relationships. Among them, d ij represents the interval between floating cars i and j, p i is a probability generated by a random program. If this probability is less than the random factor, the current vehicle is set as a floating vehicle. i =1, otherwise it is not a floating car f i =0;

[0086] d ij =100*p r (1)

[0087]

[0088] Further, as a preferred embodiment of the present invention, step S3 includes the following sub-steps:

[0089] S3.1. Use sliding windows to divide vehicle trajectories, and select a history sequence with a length of l. h , the reconstruction length is l p , then the time segments of the historical sequence Tra(h,i) and the reconstructed sequence Tra(p,i) of the i-th sample are as follows:

[0090] Tra(h,i)∈[t start ,t start +l h ] (3)

[0091] Tra(p,i)∈[t start +l h +1,t start +l h +l p ] (4)

[0092] The vehicle's position, speed, acceleration, lane and other information are all within this time range, and the time series data All_Tra expression is obtained as follows:

[0093] All_Tra={Tra(h,1),Tra(p,1),…,Tra(h,k),Tra(p,k),…,Tra(h,n),Tra(p,n)} (5)

[0094] Where, t start is the start time of each sample;

[0095] S3.2. According to the stratified sampling principle, the time series data is divided into training set, validation set and test set in a ratio of 7:1:2, where Train_set, Val_set and Test_set represent the training set, validation set and test set respectively:

[0096] Train_set=0.7*All_Tra (6)

[0097] Val_set=0.1*All_Tra (7)

[0098] Test_set=0.2*All_Tra (8)

[0099] Tra is all vehicle trajectory samples.

[0100] Further, as a preferred embodiment of the present invention, step S4 includes the following sub-steps:

[0101] S4.1. In the intelligent driver model, the following relationship of vehicles affects the input and output of the model. The leading vehicle and the following vehicle of vehicle i are determined by lane information, position information and time information, as shown in the following formula:

[0102]

[0103] Among them, is_front is used to characterize whether vehicle j is the leading vehicle in front of vehicle i, and is_back is used to characterize whether vehicle j is the following vehicle of vehicle i. The value is 1, indicating that it is the leading vehicle or the following vehicle, and the value is 0, indicating that it is not the leading vehicle or the following vehicle. S4.2. Since the trajectory information of all manually driven vehicles is missing in the conventional traffic flow scenario, only the vehicle trajectory information of the floating vehicle itself can be obtained. Therefore, the intelligent driver model is used to estimate the missing vehicle trajectory information. The intelligent driver model IDM model reconstructs other missing vehicle trajectory segments based on the observation segments of the front vehicle. The intelligent driver model IDM is as follows:

[0104]

[0105] Δv′(i,t)=v′(i,t)-v′(i-1,t) (14)

[0106] d′(i,t)=s′(i-1,t)-s′(i,t)-L i-1 (15)

[0107] Through the intelligent driver model IDM, the other missing vehicle trajectory segments can be reconstructed based on the observation segments of the front vehicle. * (v′(i,t),Δv′(i,t)) represents the expected distance, d′(i,t) represents the headway between the front and rear vehicles, Δv′(i,t) represents the speed difference between the front and rear vehicles, and a max , b represent the maximum acceleration and comfortable deceleration respectively, T safe is the safe headway time, s0 is the headway distance at rest, v0 is the expected speed, L i-1 Indicates the vehicle length of the preceding vehicle;

[0108] S4.3. The vehicle trajectory reconstructed by the intelligent driver model IDM may conflict with the trajectory of the next observed segment. Therefore, the reconstructed vehicle trajectory needs to be adjusted to ensure that the reconstructed vehicle trajectory is between the previous segment and the next segment, as shown in the following formula:

[0109] v′(i,t)=max(v′(i,t-1)+a′(i,t)Δt,0) (16)

[0110]

[0111] Further, as a preferred embodiment of the present invention, step S5 includes the following sub-steps:

[0112] S5.1. In the semi-perception scenario, some manually driven vehicle trajectories are reconstructed by the fusion model IDM-LSTM, and some vehicle trajectories are reconstructed by executing steps S4.1 to S4.3. The existing data is classified. When the historical sequence is missing, the intelligent driver model IDM is used for reconstruction; if the historical sequence data is complete, the fusion model IDM-LSTM is used for reconstruction;

[0113] S5.2. The fusion model IDM-LSTM is used for reconstruction. The LSTM model includes an input gate, a forget gate, and an output gate. The long-term and short-term dependencies in the time series data are considered through a special gate structure. The influence of the historical sequence on the reconstructed sequence is considered through the flow of trajectory information between different gate structures. The specific expressions are shown in equations (18) to (23):

[0114] f t =σ(W f ·[h t-1 ,x t ]+b f ) (18)

[0115] i t =σ(W i ·[h t-1 ,x t ]+b i ) (19)

[0116]

[0117] o t =σ(W o [h t-1 ,x t ]+b o ) (twenty two)

[0118] h t =o t *tanh(C t ) (twenty three)

[0119] where f t is the output of the forget gate, i t is the output of the input gate, is a candidate cell state, C t is the unit state at the current moment, h t is the hidden state at the current moment, x t Represents the input data; W f , W i , W C , W o Respectively represent the weight matrix of the forget gate, the weight matrix of the input gate, the weight matrix of the hidden state update, and the weight matrix of the output gate; b f , b i , b C , b o Respectively represent the bias vector of the forget gate, the bias vector of the input gate, the bias vector of the hidden state update, and the bias vector of the output gate;

[0120] S5.3 adopts the method of fusion model IDM-LSTM to realize the fusion of two models through the weight module. The weight module is shown in expressions (24)-(25). By fusing the reconstruction result of the IDM model and the reconstruction result of the LSTM model, the prediction result of the IDM-LSTM model is obtained. The prediction result is used to calculate the loss of the LSTM model, which is used to update the parameters in the training process and enhance the training process of the LSTM model.

[0121]

[0122] w IDM,i +w LSTM,i =1 (25)

[0123] Among them, s LSTM,i represents the reconstruction result of vehicle i by the LSTM model, s IDM,i represents the result of the intelligent driver model IDM reconstruction, w IDM,i ,w LSTM,i They represent the weight coefficients of the intelligent driver model IDM and LSTM model, representing the credibility of different models, w IDM,i ,w LSTM,i The sum of is 1;

[0124] S5.4. The model error is evaluated using three evaluation indicators: MAE, RMSE and MAPE, as shown in the following formula:

[0125]

[0126] Further, as a preferred embodiment of the present invention, step S6 includes the following sub-steps:

[0127] S6.1. In the full perception scenario, first use the intelligent driver model IDM to reconstruct the missing vehicle trajectory, complete the vehicle trajectory reconstruction based on the following relationship, and execute steps S4.1 to S4.3;

[0128] S6.2, intelligent driver model IDM, the reconstructed result is used to input into the LSTM model to assist the training process of the LSTM model, and execute step S5.3;

[0129] S6.3, using the fusion model IDM-LSTM method, realize the fusion of the two models through the weight module, and execute step S5.4 to reconstruct the vehicle trajectory in the full perception scenario.

[0130] The location selected for this embodiment is Highway 101 (also known as the Hollywood Freeway) in Los Angeles, California, USA, and the observation direction is from south to north. The road section is about 640 meters long, and the main road contains 5 lanes to ensure that a larger range of traffic flow dynamics can be covered. The data is recorded at a sampling frequency of 1 Hz to capture key features such as vehicle position information, speed changes, and lane occupancy in the area to support in-depth analysis of traffic flow characteristics on multi-lane highways. In this study, real data was used to calibrate the intelligent driver model IDM. The model calibration parameters are shown in Table 1, where v0 represents the free flow speed, S represents the minimum distance between vehicles at rest, and a represents the minimum distance between vehicles at rest. max represents the maximum acceleration, b represents the comfortable deceleration, T safe Indicates safe headway.

[0131] Table 1

[0132]

[0133] First, the trajectory data is processed and analyzed. The data is preprocessed, including the processing of missing values ​​and outliers. The data is converted to units. The distance unit in the original data is feet, which needs to be converted to meters, and the time is converted to seconds. The sampling interval in the original data is 1Hz. The present invention downsamples it to 10Hz to reduce the complexity of the model. Based on the data processing, the vehicle trajectory diagram is drawn. The vehicle trajectory diagram is as follows: Figure 3 As shown, since the data used are at the peak period, there is an obvious shock wave propagation path.

[0134] Then the scenario is divided into three trajectory reconstruction scenarios, and the corresponding trajectory data is generated under the three scenarios, taking into account the spatiotemporal distribution of autonomous driving vehicles at three penetration rates of 5%-15%.

[0135] Based on the constructed data set and input parameters, different trajectory reconstruction methods are used to reconstruct vehicle trajectories for different trajectory reconstruction scenarios. In the conventional traffic flow scenario, the IDM model is used to reconstruct vehicle trajectories. The penetration rate of autonomous driving vehicles is set to 5%, and the penetration rate of floating vehicles is set to vary from 5% to 15%. The corresponding experimental results are shown in Table 2 Trajectory reconstruction error table for conventional traffic flow scenarios. It can be found from Table 2 that the higher the penetration rate, the higher the accuracy of trajectory reconstruction.

[0136] Table 2

[0137]

[0138] In the semi-perception scenario, IDM and IDM-LSTM models are used to reconstruct vehicle trajectories. The corresponding model errors are shown in Table 3. The penetration rates of autonomous vehicles and floating vehicles are tested under three conditions: 5%, 10%, and 15%. It can be found that the penetration rate has a great influence on the accuracy of trajectory reconstruction. At the same time, the IDM-LSTM model is often more accurate than the IDM model under the same conditions.

[0139] Table 3 Trajectory reconstruction error in semi-perception scenario

[0140]

[0141]

[0142] In the full perception scenario, the fusion model is used to reconstruct the vehicle trajectory and compared with other models. The results shown in Table 4 are the trajectory reconstruction error table in the full perception scenario. It can be found that the IDM-LSTM model has higher accuracy under the same conditions, indicating that model fusion helps to improve the accuracy of trajectory reconstruction.

[0143] Table 4

[0144]

[0145] Combining the three trajectory reconstruction scenarios, the vehicle trajectory reconstruction method proposed in the present invention that is adaptable to multiple scenarios of mixed traffic flow still has good performance and plays an important role in the trajectory reconstruction of connected vehicles under low penetration rates in the future.

[0146] Furthermore, Figure 4 (a), (b), and (c) show the comparison of the trajectories reconstructed by the three algorithms. The reconstructed vehicle trajectory points by the method of the present invention are closer to the actual situation. These results further prove the superiority of the algorithm proposed in the present invention, and also confirm that in trajectory reconstruction, it is necessary to select a suitable method to reconstruct the vehicle trajectory according to different trajectory reconstruction scenarios.

[0147] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. A person skilled in the art of the present invention may make various modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the definition of the claims.

Claims

1. A vehicle trajectory reconstruction method adapted to multiple scenarios of mixed traffic flow, characterized in that: The following steps are involved: S1. Collect the trajectory position, speed and passing time information of surrounding manually driven vehicles; S2. Divide the target scene into three scenarios according to the type of adjacent floating vehicles: full perception scenario, semi-perception scenario and regular traffic flow scenario, consider the penetration rate of floating vehicles and generate trajectory data for the three scenarios; S3, using a sliding window approach to divide vehicle trajectories, and dividing the data set into training set, validation set, and test set according to a specific ratio; S4. In the conventional traffic flow scenario, the trajectory data of the floating vehicle itself is obtained, and the vehicle is reconstructed using the intelligent driver model under different floating vehicle penetration rates; S5. In the semi-perception scenario, the vehicle with fragmented trajectory uses the fusion model IDM-LSTM to reconstruct the trajectory, and the vehicle with missing trajectory information uses the intelligent driver model to reconstruct the trajectory; S6. In the full perception scenario, the fragmentary trajectory information of all vehicles is obtained, and a weight module is introduced to fuse the intelligent driver model with the long short-term memory neural network, namely IDM-LSTM, to reconstruct the missing vehicle trajectory.

2. According to claim 1, a vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios is characterized in that: In step S1, a vehicle passing through the target road section is detected by a fixed detector at the starting point of the road side and numbered as i. The time when vehicle i passes the fixed detector position at the starting point is recorded as b. i , the time it takes for vehicle i to pass the end of the study section is e i , retrieve the vehicle trajectory according to the vehicle number, and obtain the vehicle trajectory observation set Tra of vehicle i on the target road section i ={s(i,t),v(i,t),a(i,t),l(i,t),t} t∈[b i ,e i ], The trajectory set of all vehicles Tra = {Tra1,…,Tra k ,…,Tra n }, where s(i,t) is the longitudinal distance traveled by vehicle i on the target section at time t, v(i,t) is the speed of vehicle i on the study section at time t, a(i,t) is the acceleration of vehicle i on the study section at time t, and l(i,t) is the lane number of vehicle i on the study section at time t.

3. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 1 is characterized in that: In step S2, the target scene is divided into a full perception scene, a semi-perception scene and a regular traffic flow scene according to the types of adjacent floating vehicles. Specifically, the types of the two adjacent floating vehicles are determined in the target scene. If both of the two adjacent floating vehicles are autonomous driving vehicles, the target scene is named a full perception scene; if one is an autonomous driving vehicle and the other is a manually driven vehicle that can only obtain its own vehicle driving status information, the target scene is named a semi-perception scene; if both of the two adjacent floating vehicles are manually driven vehicles, the target scene is named a regular traffic flow scene.

4. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 1 is characterized in that: In step S2, the penetration rate of floating vehicles is considered and three scenario trajectory data are generated, as follows: Based on the fact that the floating cars are evenly distributed, the interval d between the floating cars ij As follows: d ij =100*p r (1) In the formula, p r is the penetration rate of autonomous vehicles, d ij represents the interval between floating car i and floating car j, At the same time, a random factor u is introduced to take into account the non-uniform distribution among floating cars. In the formula, p i is a probability generated by a random program. If this probability is less than the random factor, the current vehicle is set as a floating vehicle. i =1, otherwise it is not a floating car f i =0.

5. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 1 is characterized in that: Step S3 includes the following sub-steps: S3.

1. Use sliding windows to divide vehicle trajectories, and select a history sequence with a length of l. h , the reconstruction length is l p , then the time segments of the historical sequence Tra(h,i) and the reconstructed sequence Tra(p,i) of the i-th sample are as follows: Tra(h,i)∈[t start ,t start +l h ] (3) Between(p,i)∈[t start +l h +1,t start +l h +l p ] (4) The vehicle's position, speed, acceleration, lane and other information are all within this time range, and the time series data All_Tra expression is obtained as follows: All_Tra={Tra(h,1),Tra(p,1),…,Tra(h,k),Tra(p,k),…,Tra(h,n),Tra(p,n)} (5) Where, t start is the start time of each sample; S3.

2. According to the stratified sampling principle, the time series data is divided into training set, validation set and test set according to the preset ratio, where Train_set, Val_set and Test_set represent the training set, validation set and test set respectively: Train_set=m1*All_Tra (6) Val_set=m2*All_Tra (7) Test_set = m3*All_Tra (8) Tra is the sample of all vehicle trajectories, and m1, m2, and m3 are the corresponding proportional coefficients.

6. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 1 is characterized in that: Step S4 includes the following steps: S4.

1. In the intelligent driver model, the leading vehicle and the following vehicle of vehicle i are determined by lane information, position information and time information, as shown in the following formula: Among them, is_front indicates whether vehicle j is the leading vehicle in front of vehicle i, and is_back indicates whether vehicle j is the following vehicle of vehicle i. A value of 1 indicates that it is a leading vehicle or a following vehicle, and a value of 0 indicates that it is not a leading vehicle or a following vehicle. S4.

2. An intelligent driver model is used to estimate the missing vehicle trajectory information. The intelligent driver model IDM model is used to reconstruct other missing vehicle trajectory segments based on the observation segments of the preceding vehicle. The intelligent driver model IDM is as follows: Δv′(i,t)=v′(i,t)-v′(i-1,t) (14) d′(i,t)=s′(i-1,t)-s′(i,t)-L i-1 (15) In the formula, s * (v′(i,t),Δv′(i,t)) represents the expected distance, d′(i,t) represents the headway between the front and rear vehicles, Δv′(i,t) represents the speed difference between the front and rear vehicles, and a max , b represent the maximum acceleration and comfortable deceleration respectively, T safe is the safe headway, s0 is the headway at rest, v0 is the expected speed, L i-1 Indicates the vehicle length of the preceding vehicle; S4.

3. Adjust the reconstructed vehicle trajectory to ensure that the reconstructed vehicle trajectory is between the previous segment and the next segment, as shown in the following formula: v′(i,t)=max(v′(i,t-1)+a′(i,t)Δt,0) (16) 7. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 1 is characterized in that: Step S5 includes the following sub-steps: S5.

1. In the semi-perception scenario, some manually driven vehicle trajectories are reconstructed by the fusion model IDM-LSTM, and some vehicle trajectories are reconstructed by executing steps S4.1 to S4.

3. The existing data is classified. When the historical sequence is missing, the intelligent driver model IDM is used for reconstruction; if the historical sequence data is complete, the fusion model IDM-LSTM is used for reconstruction; S5.

2. The fusion model IDM-LSTM is used for reconstruction. The LSTM model includes an input gate, a forget gate, and an output gate. The trajectory information is circulated between different gate structures, and the influence of the historical sequence on the reconstructed sequence is considered. The specific expressions are shown in equations (18) to (23): f t =σ(W f ·[h t-1 ,x t ]+b f ) (18) i t =σ(W i ·[h t-1 ,x t ]+b i ) (19) the t =σ(W o [h t-1 ,x t ]+b o ) (22) h t =o t *tanh(C t ) (23) where f t is the output of the forget gate, i t is the output of the input gate, is a candidate cell state, C t is the unit state at the current moment, h t is the hidden state at the current moment, x t Represents the input data; W f , W i , W C , W o Respectively represent the weight matrix of the forget gate, the weight matrix of the input gate, the weight matrix of the hidden state update, and the weight matrix of the output gate; b f , b i , b C , b o Respectively represent the bias vector of the forget gate, the bias vector of the input gate, the bias vector of the hidden state update, and the bias vector of the output gate; S5.3 adopts the method of fusion model IDM-LSTM to realize the fusion of two models through the weight module. The weight module is shown in expressions (24)-(25). By fusing the reconstruction result of the IDM model and the reconstruction result of the LSTM model, the prediction result of the IDM-LSTM model is obtained. The prediction result is used to calculate the loss of the LSTM model, which is used to update the parameters in the training process and enhance the training process of the LSTM model. In IDM,i +in LSTM,i =1 (25) Among them, s LSTM,i represents the reconstruction result of vehicle i by the LSTM model, s IDM,i represents the result of the intelligent driver model IDM reconstruction, w IDM,i ,w LSTM,i They represent the weight coefficients of the intelligent driver model IDM and LSTM model, representing the credibility of different models, w IDM,i ,w LSTM,i The sum of is 1; S5.

4. The model error is evaluated using three evaluation indicators: MAE, RMSE and MAPE, as shown in the following formula:

8. The vehicle trajectory reconstruction method adapted to mixed traffic flow in multiple scenarios according to claim 7 is characterized in that: Step S6 includes the following sub-steps: S6.

1. In the full perception scenario, first use the intelligent driver model IDM to reconstruct the missing vehicle trajectory, complete the vehicle trajectory reconstruction based on the car-following rule, and execute steps S4.1 to S4.3; S6.2, intelligent driver model IDM, the reconstructed result is used to input into the LSTM model to assist the training process of the LSTM model, and execute step S5.3; S6.3, using the fusion model IDM-LSTM method, realize the fusion of the two models through the weight module, and execute step S5.4 to reconstruct the vehicle trajectory in the full perception scenario.

Images