Vehicle control delay prediction method based on hypergraph model

Abstract

The vehicle control delay prediction method based on the hypergraph model is applicable to the autonomous driving technology in the field of intelligent transportation. First, unlike the traditional graph model, the hyperedge defined by the hypergraph can connect two or more trajectory points to encode the high-order relationship between the trajectory points; secondly, the hyperedge is constructed by dense neighborhood search based on multi-core trajectory similarity to more accurately characterize the connection strength between trajectory points; by adding road network constraints, the predicted trajectory is constrained within the feasible domain of the road, further improving the accuracy of the prediction results. The main technical problems solved by the present invention include: (a) improving the accuracy of trajectory prediction results; (b) reducing parameter training costs; (c) improving the accuracy and computational efficiency of control delay prediction results.

Description

Technical Field

[0001] The present invention is applicable to the automatic driving technology in the field of intelligent transportation. Background Art

[0002] Estimation of vehicle delays that often occur at signalized intersections on a road network is an important task for transportation departments. It not only optimizes signal timing at intersections, but also helps alleviate traffic congestion and effectively manage traffic throughout a city. Delay prediction methods can generally be divided into two steps: trajectory prediction and delay estimation.

[0003] Vehicle trajectory prediction is to use the current and historical location information of the vehicle, as well as other auxiliary information (such as road network status, neighboring vehicles, weather, environment, etc.) to dynamically predict the future location of the vehicle. Vehicle trajectory prediction technology plays a vital role in the research and development of related fields. For example, vehicle trajectory prediction technology promotes the development of intelligent transportation, wireless network deployment, location-based service applications and other fields. Vehicle trajectory prediction methods can be roughly divided into the following two categories: statistical learning-based methods and deep learning-based methods.

[0004] (1) Trajectory prediction method based on statistical learning

[0005] Trajectory prediction methods based on statistical learning include hidden Markov models, Bayesian models, Gaussian processes, etc. These methods generally consider the motion trajectories of different mobile targets from the statistical distribution characteristics of the data, analyze the frequent motion patterns in the historical trajectories of mobile targets, and perform computational reasoning based on traditional probability statistics methods to obtain the future motion trends and trajectories of the targets.

[0006] For non-independent time series data such as mobile trajectories and videos, the prediction method based on the hidden Markov model (HMM) can reasonably classify and regress such data through the information provided by the context. However, the existing low-order Markov model method only considers current information and ignores historical information, which is inefficient. In contrast, high-order Markov models can improve accuracy, but increase the time and space complexity of prediction. The Bayesian Network-based method proposes to infer the distribution of all vehicle driving operations in the driving scene by using causal relationships and diagnostic evidence in the Bayesian network for modeling, and then evaluate the probability of the self-vehicle colliding at least once in a time interval through Monte Carlo simulation as a function of the prediction range. Based on the Gaussian process, a long-term vehicle trajectory prediction method is proposed, in which road and traffic data can be obtained through communication between vehicles and between vehicles and infrastructure such as roads. Gaussian process regression is used to learn motion patterns, data collected from vehicles and the Gaussian process regression model, and finally predict the trajectory of vehicles in future scenarios from the infrastructure. The advantage of this method is that it is not limited to being able to predict the future trajectory of the vehicle, but can also reliably predict the distribution of future vehicle trajectories.

[0007] (2) Trajectory prediction method based on deep learning

[0008] In the research field of trajectory prediction, deep learning has also been widely used, such as trajectory prediction methods based on long short-term memory networks and trajectory prediction methods based on generative adversarial networks.

[0009] Long-Short Term Memory (LSTM) is a special type of recurrent neural network (RNN) that can learn the interdependence between data at different times in long-term time series data. The advantage of the LSTM model is that it solves the problem of vanishing gradients in RNN, so it is conducive to training long-term time series and predicting long-term vehicle trajectories. However, the disadvantage is that the LSTM model cannot simultaneously describe the spatial interactions between different vehicles and the spatial relationships between trajectories. Generative Adversarial Networks (GAN) is a machine learning architecture proposed by Ian Goodfellow of the University of Montreal in 2014. GAN includes two independent networks: the generator and the discriminator. The generator is used to generate random samples similar to real samples and use them as fake samples, and the discriminator is used to distinguish the authenticity of data. However, the trajectory prediction method based on the GAN network has the problem of difficult GAN training.

[0010] The research on vehicle delay prediction mainly includes the following aspects: control delay, approach delay and parking delay. Control delay is defined as the difference between the travel time when the vehicle is not affected by the controlled intersection and when it is affected by the controlled intersection. Approach delay is defined as the difference between the fixed distance that any vehicle travels from a predetermined point upstream of the intersection to the intersection stop pole and the free flow time associated with that distance. Parking delay is the delay caused by a vehicle being stationary for some reason, which is equal to the sum of the parking time and the driver's reaction time from the vehicle stopping to the vehicle starting again. Since control delay is caused by traffic facilities, it is regarded as a measure of the service level of signalized intersections, so control delay prediction is one of the current research hotspots.

[0011] The accuracy of vehicle delay prediction depends greatly on the trajectory prediction accuracy, which generally has the following three requirements: the prediction results have high accuracy; parameter training has low spatial and temporal complexity; the prediction model has strong interpretability. Traditional vehicle trajectory prediction methods can generally only meet one or two of the above conditions, and it is difficult to meet all three conditions at the same time. Considering that the trajectory prediction problem can be transformed into a trajectory point similarity problem, graph models are widely used in trajectory prediction. However, for multiple vehicle trajectories, traditional graph models can only establish a single trajectory point similarity, which easily leads to its difficulty in handling the correlation of multiple trajectory points in the global time domain of long-term trajectories. To solve this problem, the present invention proposes a vehicle trajectory prediction method based on a hypergraph model, using hyperpoints to represent the trajectory points of each vehicle trajectory, and hyperedges to represent the high-order relationship between the trajectory points of different vehicles that integrates position, speed, and direction constraints, so as to ensure that the predicted trajectory points have a high degree of similarity with the historical trajectory point set of the corresponding trajectory, that is, in the global time domain of the historical trajectory, the predicted points can belong to the historical trajectory point set of the same trajectory to a large extent. The trajectory prediction method proposed in the present invention can better meet the three requirements of vehicle trajectory prediction. Summary of the invention

[0012] The present invention proposes a vehicle control delay prediction method based on a hypergraph model. First, unlike the traditional graph model, the hyperedge defined by the hypergraph can connect two or more trajectory points to encode the high-order relationship between the trajectory points; secondly, the present invention proposes a new method for constructing a hypergraph, which constructs hyperedges by dense neighborhood search based on multi-core trajectory similarity to more accurately characterize the connection strength between trajectory points; by adding road network constraints, the predicted trajectory is constrained within the feasible domain of the road, further improving the accuracy of the prediction results. The main technical problems solved by the present invention include: (a) improving the accuracy of trajectory prediction results; (b) reducing parameter training costs; (c) improving the accuracy and computational efficiency of control delay prediction results.

[0013] The vehicle control delay prediction method based on the hypergraph model proposed in the present invention can effectively encode the high-order relationship between the trajectory points of different vehicles, improving the shortcomings of the trajectory prediction method based on the graph model; the hyperedge is constructed by using the multi-core trajectory similarity measurement method and dense neighborhood search to effectively characterize the connection strength between the trajectory points. By adding the feasible direction constraint of the road network, the predicted trajectory is constrained within the feasible domain of the road, further improving the accuracy of the prediction result. Finally, the vehicle control delay is estimated based on the obtained predicted trajectory. Figure 1 A flow chart of vehicle control delay prediction based on hypergraph model is given.

[0014] The present invention is achieved through the following technical solutions:

[0015] Firstly, a fixed-length curved trajectory is extracted from the trajectory database. Then, the extracted trajectory is divided into historical trajectory and future trajectory. Each trajectory point of the historical trajectory corresponds to a node in the graph. The trajectory matching algorithm is used to generate candidate prediction points for the next moment. With each candidate point as the center, a dense neighborhood search algorithm is used to construct a hyperedge. The hypergraph is constructed by traversing all candidate points to construct a hyperedge. A calculation algorithm for the hyperedge weight is designed, and the prediction point is determined according to the maximum hyperedge weight. The prediction is iterated in sequence. Finally, the control delay is estimated based on the obtained predicted trajectory.

[0016] The specific steps are as follows:

[0017] Step 1: Curved trajectory extraction

[0018] The curved trajectory extraction algorithm is used. First, the inflection point search is performed to search for all the inflection points contained in each trajectory (the position where the direction of the reaction trajectory suddenly changes), the slope of each trajectory point is calculated, and the trajectory points greater than the threshold are screened by the set corner threshold η1, and the trajectory points are regarded as inflection points; secondly, the inflection point filtering is performed to ensure that there are no repeated trajectory points between the generated curved trajectories, and then the inflection points that do not meet the conditions are filtered out by setting the interval threshold η2; finally, the curved trajectory is generated based on the filtered inflection points. The I curved trajectories extracted from the database are defined as Where I is the total number of curved trajectories, and the i-th curved trajectory is defined as: Its length is T′ i , the trajectory is divided into historical trajectory and future trajectory, and the historical trajectory is expressed as: Its length is T i , the future trajectory is expressed as: Its length is T′ i -T i , where T′ i =2T i .

[0019] Step 2: Trajectory representation and prediction based on hypergraph

[0020] 1. Selection of candidate points

[0021] First, according to the historical trajectory of the extracted curved trajectory If at least one trajectory point in the trajectory in the database falls within the range of longitude and latitude determined by the algorithm, the trajectory is regarded as a rough matching trajectory. Secondly, the historical trajectory is matched with the rough matching trajectory, that is, the Euclidean distance of each point is calculated, and then the matching trajectory is screened out according to the set error threshold η3. Finally, the trajectory point at the next moment that is successfully matched in the matching trajectory is regarded as a candidate point, and the i-th trajectory is defined as i The candidate point set at time +1 for:

[0022]

[0023] Where q is a trajectory point and N′ is the number of candidate points. Redefine a trajectory, which includes historical trajectories and candidate points, as:

[0024]

[0025] The historical trajectory is

[0026] 2. Similarity measurement of trajectory points

[0027] A trajectory point is represented by three attributes (p, v, r), where Represents the position coordinates of the trajectory point, represents the velocity of the trajectory point, represents the local feasible direction of the road where the trajectory point is located, is the dimension identifier. The similarity measure between trajectory points uses the Gaussian kernel function to comprehensively consider the similarity between corresponding attributes. Specifically, N is the number of historical trajectory points, and N′ is the number of predicted candidate points. The multi-kernel similarity matrix is ​​defined for historical trajectory points and predicted candidate points The matrix is ​​obtained by weighting the position, velocity, and direction similarity matrices of the trajectory points, namely:

[0028] A=ω1A pos +ω2A vel +ω3A dir (3)

[0029] Where ω1+ω2+ω3=1, the corresponding elements of the similarity matrix of the position, velocity and feasible direction of the trajectory point are defined as:

[0030]

[0031] in,[·] ijrepresents the i-th row and j-th column element of the corresponding matrix, ||·||2 represents the bi-norm of the corresponding matrix, A pos represents the position similarity, p i ,p j are the positions of the i-th and j-th trajectory points respectively; A vel represents the speed similarity, v i ,v j are the velocities of the i-th and j-th trajectory points respectively; A dir represents the local feasible direction similarity, r i ,r j are respectively the feasible direction vectors of the i-th and j-th trajectory points. The feasible direction vector is obtained by selecting the position with the highest passing rate of historical trajectory points in the local neighborhood of the predicted candidate point as the end point. The similarity of this feature constrains the predicted point within the feasible range of the road as much as possible, and the historical road network data can be used as a priori to improve the accuracy of the prediction result; α 2 , β 2 , γ 2 They are respectively the covariance of the trajectory point position, the covariance of the trajectory point speed and the covariance of the local feasible direction of the road network.

[0032] 3. Hyperedge construction based on dense neighborhood search algorithm

[0033] Defining a Hypergraph The node set is: The hyperedge set is defined as ε, and the hyperedge weight is defined as For each candidate point q n ,N+1≤n≤N+N′, and regard it as the search center to search for its If the candidate point belongs to the future trajectory, the similarity between the candidate point and its neighboring points will be large, and vice versa. Then according to the similarity matrix obtained above in The hyperedge construction of dimension identification will be solved by optimizing formula (5):

[0034]

[0035] Among them, K is the number of nodes contained in the hyperedge, is the number of neighborhood points, N is the number of historical trajectory points, j1j2 is the index set Any two indices in .

[0036] Formula (5) is a typical nonlinear optimization problem with constraints. After calculation and solution, the hyperedge of can be expressed as:

[0037]

[0038] Among them, en-N is a hyperedge centered at the nth candidate point, q k For this candidate point and its The point set consists of neighborhood points.

[0039] 4. Hypergraph Construction

[0040] Next, we construct a hypergraph consisting of historical trajectory points and predicted candidate points. Use the current and historical trajectory points to generate the predicted candidate points for the next moment, and use the historical trajectory points and predicted candidate points as the node set of the hypergraph Construct a hyperedge set ε with each candidate point as the center. Specifically, with each candidate point as the center, search for neighboring points based on the dense neighborhood search algorithm, and form a hyperedge with each candidate point and these neighboring points. Each hyperedge corresponds to a predicted candidate point, which connects points. For each hyperedge e n-N , define its weight w en-N for:

[0041]

[0042] in,[·] ij Represents the element in the i-th row and j-th column of the corresponding matrix.

[0043] 5. Select prediction points

[0044] According to the above hypergraph construction method, the trajectory prediction at this time is to construct a hyperedge for each predicted candidate point through dense neighborhood search, adjust the similarity weight and the number of neighboring points K, and minimize the distance loss between the candidate point corresponding to the maximum hyperedge weight and the actual trajectory point. Correspondingly, the predicted point position p at time t+1 t+1 It is given by selecting the central node with the largest hyperedge weight, that is:

[0045]

[0046] where i * is the index of the candidate point with the maximum hyperedge weight.

[0047] In this way, the trajectory points at the next moment are iteratively predicted to obtain the predicted trajectory.

[0048] Step 3: Vehicle Delay Estimation

[0049] The predicted trajectory and the historical trajectory form a complete trajectory It is expressed as:

[0050]

[0051] in For the historical trajectory, To predict the trajectory.

[0052] According to the calculation definition of vehicle control delay estimation, it is equal to the difference between the time it takes a vehicle to pass the intersection and the time it takes to travel at the same speed. Therefore, for each trajectory:

[0053]

[0054] According to the segmentation strategy of the curved trajectory, we know that the turning point of the trajectory is Compare the speeds of the trajectory points before and after the turning point, set a fluctuation threshold η4, calculate the difference in speed between the trajectory points at both ends of the turning point, and take the time less than the fluctuation threshold and with the largest interval as the estimated value of the control delay, which can be expressed as:

[0055]

[0056] Among them, t(·) represents the timestamp, ||·|| represents the modulus of the corresponding vector, express The speed of D i Represents an estimate of control delay. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 :Vehicle delay prediction flow chart

[0058] Figure 2 : Visualization results of curved trajectory extraction

[0059] Figure 3 :Box plot of prediction accuracy based on hypergraph model

[0060] Figure 4(a): Trajectory prediction ADE comparison line chart

[0061] Figure 4(b): Trajectory prediction FDE comparison line chart

[0062] Figure 5 : Control delay estimation ADE comparison line chart DETAILED DESCRIPTION

[0063] The present invention experimentally verifies the above method and achieves significant results. The trajectory data used in this experiment are all desensitized data from the Didi Gaia data open plan in Chengdu in November 2016. The data trajectory points include longitude, latitude, and timestamp. Figure 2 This is the visualization result of the original data set after curved trajectory extraction (the black area is the trajectory point distribution of the sub-trajectory in the data set).

[0064] 1. Quantitative evaluation of trajectory prediction

[0065] This experiment uses two evaluation criteria to evaluate the trajectory prediction algorithm, namely, Average Displacement Error (ADE) and Final Displacement Error (FDE), which are defined as:

[0066]

[0067] FDE=||p cur+step -g cur+step || (13)

[0068] where p t+1 is the predicted trajectory point at time t+1, g t+1 is the actual trajectory point at time t+1, p cur+step is the predicted trajectory end point, g cur+step For p cur+step Corresponding to the real trajectory point at the same time, step represents the number of predicted time intervals.

[0069] The parameter settings involved in the experiment are:

[0070] ω1=0.6,ω2=0.25,K=4

[0071]

[0072] The model external parameters are: ω1, ω2, K. In the experiment, the setting range of ω1 and ω2 is [0, 1], and the step size is 0.1; the setting range of K is [2, 6], and the step size is 1. The threshold parameters η1, η2, η3, and η4 are set according to experience.

[0073] In order to analyze and verify the effectiveness and accuracy of our hypergraph model, we used three cutting-edge comparison methods in trajectory prediction: prefix projection coding model, long short-term memory network, and graph model. We made predictions for Chengdu every day in November 2016 and drew a box plot of the hypergraph model prediction accuracy. Figure 3 As shown in the figure, it can be seen that the prediction method of the hypergraph model proposed in this patent has a certain accuracy.

[0074] At the same time, compared with the other three methods, we counted the average ADE error and average FDE error of each method every day in November, and drew a line graph as shown in Figure 4 (a) (b). It can be seen from the figure that the prediction method of the hypergraph model proposed in this patent has higher accuracy than the three control methods.

[0075] It can be seen from the experimental data compared in the figure that the method proposed in the present invention has a lower prediction error than the cutting-edge method, and effectively improves the accuracy of trajectory prediction.

[0076] 2. Quantitative evaluation of delay estimation

[0077] We estimate the control delay based on the predicted trajectory and draw the following Figure 5 A line chart of the average ADE of delay estimates is shown.

[0078] Figure 5 As can be seen from the figure, our method achieves the highest estimation accuracy on more than 11 days in a month, and effectively verifies the accuracy of our trajectory prediction. Among them, our prediction results on the 10th and 22nd days are not optimistic, which is attributed to the presence of some difficult-to-predict "special trajectories" in the trajectories of these two days, such as: sudden acceleration, abnormal lane change, etc. The prediction of such trajectories requires some other data as prior support, such as traffic rules, weather, driver's driving motivation, driving style, etc. On the other hand, our prediction model has a lower time complexity. For about 2,000 trajectory data in a day, our total prediction time is about 4 hours, that is, the prediction time for each trajectory is about 7 seconds, which effectively improves the computational efficiency of the control delay prediction results.

Claims

1. The vehicle control delay prediction method based on the hypergraph model has the following steps: Step 1: Curved trajectory extraction The curved trajectory extraction algorithm is used; first, the inflection point search is performed to search for all the inflection points contained in each trajectory, the slope of each trajectory point is calculated, and the trajectory points greater than the threshold are screened by the set corner threshold η1, and the trajectory points are regarded as inflection points; secondly, the inflection point filtering is performed to ensure that there are no repeated trajectory points between the generated curved trajectories, and then the inflection points that do not meet the conditions are filtered out by setting the interval threshold η2; finally, the curved trajectory is generated according to the filtered inflection points; the I curved trajectories extracted from the database are defined as Where I is the total number of curved trajectories, and the i-th curved trajectory is defined as: Its length is T i ′, the trajectory is divided into historical trajectory and future trajectory, and the historical trajectory is expressed as: Its length is T i , the future trajectory is expressed as: Its length is T i ′-T i , where T i ′=2T i ; Step 2: Trajectory representation and prediction based on hypergraph 1) Selection of candidate points First, according to the historical trajectory of the extracted curved trajectory If at least one trajectory point in the trajectory in the database falls within the range of longitude and latitude determined by the algorithm, the trajectory is regarded as a rough matching trajectory. Secondly, the historical trajectory is matched with the rough matching trajectory, that is, the Euclidean distance of each point is calculated, and then the matching trajectory is screened out according to the set error threshold η3. Finally, the trajectory point at the next moment that is successfully matched in the matching trajectory is regarded as a candidate point, and the i-th trajectory is defined as i The candidate point set at time +1 for: Where q is a trajectory point and N′ is the number of candidate points. Redefine a trajectory, which includes historical trajectories and candidate points, as: The historical trajectory is 2) Similarity measurement of trajectory points A trajectory point is represented by three attributes (p, v, r), where Represents the position coordinates of the trajectory point, represents the velocity of the trajectory point, represents the local feasible direction of the road where the trajectory point is located, is the dimension identifier; the similarity measure between trajectory points uses the Gaussian kernel function to comprehensively consider the similarity between corresponding attributes; specifically, N is the number of historical trajectory points, N′ is the number of predicted candidate points; a multi-kernel similarity matrix is ​​defined for historical trajectory points and predicted candidate points The matrix is ​​obtained by weighting the position, velocity, and direction similarity matrices of the trajectory points, namely: A=ω1A pos +ω2A vel +ω3A dir (3) Where ω1+ω2+ω3=1, the corresponding elements of the similarity matrix of the position, velocity and feasible direction of the trajectory point are defined as: in,[·] ij represents the i-th row and j-th column element of the corresponding matrix, ||·||2 represents the bi-norm of the corresponding matrix, A pos represents the position similarity, p i ,p j are the positions of the i-th and j-th trajectory points respectively; A vel represents the speed similarity, v i ,v j are the velocities of the i-th and j-th trajectory points respectively; A dir represents the local feasible direction similarity, r i ,r j are respectively the feasible direction vectors of the i-th and j-th trajectory points. The feasible direction vector is obtained by selecting the position with the highest passing rate of historical trajectory points in the local neighborhood of the predicted candidate point as the end point. The similarity of this feature constrains the predicted point within the feasible range of the road as much as possible, and the historical road network data can be used as a priori to improve the accuracy of the prediction result; α 2 , β 2 , γ 2 They are the covariance of the trajectory point position, the covariance of the trajectory point speed and the covariance of the local feasible direction of the road network; 3) Hyperedge construction based on dense neighborhood search algorithm Defining a Hypergraph The node set is: The hyperedge set is defined as ε, and the hyperedge weight is defined as For each candidate point q n ,N+1≤n≤N+N′, and regard it as the search center to search for its dense neighborhood; if the candidate point belongs to the future trajectory, the similarity between the candidate point and its neighborhood points will be large, and vice versa; then according to the similarity matrix obtained above in For dimension identification, hyperedge construction will be solved by optimizing formula (5): Among them, K is the number of nodes contained in the hyperedge, is the number of neighborhood points, N is the number of historical trajectory points, j1j2 is the index set Any two indices in ; Formula (5) is a typical nonlinear optimization problem with constraints. After calculation and solution, the hyperedge is expressed as: Among them, e n-N is a hyperedge centered at the nth candidate point, q k For this candidate point and its The point set composed of neighborhood points; 4) Construction of Hypergraph Next, we construct a hypergraph consisting of historical trajectory points and predicted candidate points. Use the current and historical trajectory points to generate the predicted candidate points for the next moment, and use the historical trajectory points and predicted candidate points as the node set of the hypergraph Construct a hyperedge set ε with each candidate point as the center in turn; specifically, with each candidate point as the center, search for neighboring points based on the dense neighborhood search algorithm, and form a hyperedge with each candidate point and these neighboring points. Each hyperedge corresponds to a predicted candidate point, which connects points; for each hyperedge e n-N , define its weight for: in,[·] ij Represents the i-th row and j-th column element of the corresponding matrix; 5) Select prediction points According to the above hypergraph construction method, the trajectory prediction at this time is to construct a hyperedge for each predicted candidate point through dense neighborhood search, adjust the similarity weight and the number of neighboring points K, and minimize the distance loss between the candidate point corresponding to the maximum hyperedge weight and the true trajectory point; accordingly, the predicted point position p at time t+1 t+1 It is given by selecting the central node with the largest hyperedge weight, that is: where i * is the index of the candidate point with the maximum hyperedge weight; In this way, the trajectory points at the next moment are iteratively predicted to obtain the predicted trajectory; Step 3: Vehicle Delay Estimation The predicted trajectory and the historical trajectory form a complete trajectory It is expressed as: in For the historical trajectory, To predict the trajectory; According to the calculation definition of vehicle control delay estimation, it is equal to the difference between the time it takes a vehicle to pass through the intersection and the time it takes to travel at the same speed; therefore, for each trajectory: According to the segmentation strategy of the curved trajectory, we know that the turning point of the trajectory is Compare the speeds of the trajectory points before and after the turning point, set a fluctuation threshold η4, calculate the difference in speed between the trajectory points at both ends of the turning point, and take the time that is less than the fluctuation threshold and has the largest interval as the estimated value of the control delay, expressed as: Among them, t(·) represents the timestamp, ||·|| represents the modulus of the corresponding vector, express The speed of D i Express an estimate of control delays; The parameter settings involved are:

Images