A vehicle driving trajectory prediction method for an automatic driving system

Abstract

The present invention discloses a vehicle driving trajectory prediction method for an automatic driving system, which includes the steps of 1, determining the position and boundary of surrounding static obstacles; 2, predicting the trajectory of the dynamic obstacles of interest by a long short-term memory network method; 3, planning a collision-free local path by a dynamic programming optimization method; 4, calculating a suitable preview distance according to the current positioning and speed of the vehicle itself, and determining the corresponding preview point on the local path; 5, calculating the wheel deflection angle required to reach the preview point according to the difference between the vehicle heading and the heading of the preview point; 6, calculating the vehicle's own position in the next step according to the vehicle kinematics; 7, determining whether the index value of the preview point exceeds the index value of the local path end point, and if so, outputting the final result as a collision-free predicted trajectory; otherwise, returning to step 4. The technical effect of the present invention is that an automatic driving vehicle can accurately predict the driving trajectory on a road without lane lines, reduce computing power, and avoid collision accidents.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to a driving trajectory prediction method for autonomous driving. Background Art

[0002] Rural roads or intersections without lane lines have a high degree of randomness among traffic participants and it is difficult to predict vehicle behavior, which makes it difficult for autonomous vehicles to make accurate decisions. This has always been a pain point and difficulty in the field of autonomous driving.

[0003] Most current decision-making methods make decisions on the behavior of the vehicle based on whether the reference trajectory intersects with the predicted trajectory of the traffic participants. See patent document CN110751847B. According to the determined reference path, the traffic participants within the autonomous driving scene where the vehicle is located are screened to obtain key traffic participants, and then decisions are made for the key traffic participants. The reference trajectory itself is not the actual driving trajectory of the vehicle, especially for the reference trajectory of unstructured roads (unstructured roads refer to roads with a relatively low degree of structure, such as urban non-main roads and rural roads, and the roads have no clear boundaries). Some path curvatures do not meet the vehicle kinematics. In some scenarios, such as when the vehicle detours around obstacles, it may miss key obstacle targets if only the reference path is selected. Therefore, accurate prediction of the vehicle's own trajectory is the key to accurate decision-making of autonomous vehicles.

[0004] Common trajectory prediction methods include physical model prediction, machine learning prediction, and lane line prediction. The physical model prediction method does not take into account the interaction with the surrounding environment, and the prediction error is large. The machine learning-based prediction method extracts the vehicle's own motion characteristics for prediction, which can achieve better prediction accuracy, but the computing resource requirements are high. The lane line-based prediction method is relatively simple, but it cannot predict the trajectory on roads without lane lines. Summary of the invention

[0005] In view of the problems existing in the prior art, the technical problem to be solved by the present invention is to provide a vehicle driving trajectory prediction method for an automatic driving system, which can accurately predict the driving trajectory on a road without lane lines, reduce computing power, and avoid collision accidents.

[0006] The technical problem to be solved by the present invention is achieved through such a technical solution, which comprises the following steps:

[0007] Step 1: Determine the location and boundaries of surrounding static obstacles through the sensor system;

[0008] Step 2: Use the long short-term memory network method to predict the trajectory of the dynamic obstacle of interest and give the position of the dynamic obstacle in the future;

[0009] Step 3: Based on the position and boundary of static obstacles, the position of dynamic obstacles in the future, and the reference path provided by the high-precision map, a collision-free local path is planned using a dynamic programming optimization method;

[0010] Step 4: Calculate the appropriate preview distance based on the vehicle's current position and speed, and determine the corresponding preview point on the local path;

[0011] The preview distance is calculated as: L dis =L int +k*V x

[0012] L int is the set initial preview distance, Vx is the current speed of the vehicle, and k is the preview coefficient;

[0013] The method of matching the local path with the preview point is: starting from the starting point of the local path, the distances between the local path target points are continuously accumulated in sequence to obtain the cumulative length L point , until L point >L dis Get a point on the local path, use this point as the preview point, and record the index value of this point;

[0014] Step 5: Calculate the wheel deflection angle required to reach the preview point based on the difference between the vehicle heading and the preview point heading;

[0015] The calculation method of wheel deflection angle is:

[0016]

[0017] Where, L wheelbase is the wheelbase of the vehicle itself, α is the difference between the vehicle heading and the heading of the preview point;

[0018] Step 6: Calculate the next vehicle position based on vehicle kinematics. The specific position calculation method is:

[0019] x i+1 =x i +D*cos(θ i )

[0020] y i+1 =y i +D*sin(θ i )

[0021]

[0022] In the formula, x i ,y i ,θ iare the x-coordinate, y-coordinate and heading angle of the current vehicle position calculated by the step length; δ i is the discrete value of the wheel deflection angle δ; x i+1 ,y i+1 ,θ i+1 are the x-coordinate, y-coordinate and vehicle heading angle of the next step towards the preview point, and D is the distance step length;

[0023] Step 7: Determine whether the index value of the preview point exceeds the index value of the local path end point. If so, output the final result as a collision-free predicted trajectory; otherwise, return to step 4.

[0024] The technical effects of the present invention are:

[0025] The invention of this method requires less computing power, is easy to implement in engineering, and is independent of lane lines, and can accurately predict the vehicle's driving trajectory. It not only overcomes the disadvantage of large prediction errors in physical model prediction methods, but also avoids the engineering problems caused by high computing resources in machine learning prediction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings of the present invention are as follows:

[0027] Figure 1 is a flow chart of the present invention;

[0028] Figure 2 A trajectory prediction diagram of an autonomous driving vehicle according to an embodiment. DETAILED DESCRIPTION

[0029] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0030] like Figure 1 As shown, the process of the present invention comprises the following steps:

[0031] Step 1: Determine the location and boundaries of surrounding static obstacles:

[0032] According to the output of the vehicle vision sensor, target feature recognition is performed, and regular obstacles including cars, motorcycles, bicycles, and pillars are output as target boxes (BoundingBox), and irregular obstacles including walls and doors are outlined and boundary coordinate points (FreeSpace) are output to form a planning space. The position and boundary of static obstacles in one embodiment are as follows: Figure 2 shown.

[0033] Step 2: Predict the trajectory of the dynamic obstacle of interest and give the position of the dynamic obstacle in the future:

[0034] According to the document "Uncertainty-Aware Decision-Making for Autonomous Driving at Uncontrolled Intersections", X. Tang et al, IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 9, pp. 9725-9735, Sept. 2023, the long short-term memory network (LSTM) recorded in the document is used for trajectory prediction:

[0035] Record the historical trajectory of dynamic obstacles: Input t =[h t-m ,…h t ]

[0036] In the formula, h t-m It is a representation of the trajectory of historical moments. is the trajectory point at the current time t, is the coordinate of the starting point of the predicted trajectory at the current moment, is the coordinate of the end point of the predicted trajectory point at the current moment.

[0037] Use LSTM encoder to encode historical trajectories:

[0038] PV t =FC(LSTM enc (Emb(Input t )))

[0039] Where, Emb() is the fully connected layer that embeds x and y, and LSTM enc () is a shared LSTM encoder, and FC() is the fully connected layer of a certain obstacle PV.

[0040] First, use Emb() to embed the historical trajectory of PV and send it to the LSTM network. Then, the LSTM encoded features are processed by the FC() function as the dynamic features of the dynamic obstacle PV.

[0041] The convolutional pooling layer is used to extract the interaction signs SV between the predicted dynamic obstacles and the surrounding obstacles according to the relative positions of the surrounding obstacles and the predicted dynamic obstacles. t :

[0042] SV t =MaxP(Soc(LSTM enc(Emb(Input t ))))

[0043] In the formula, Soc() is a two-layer convolutional neural network and MaxP is the maximum pooling layer.

[0044] The predicted trajectory of the predicted dynamic obstacle is obtained by decoding with the LSTM decoder:

[0045] Output t =LSTM dec [PV t ,SV t ]

[0046] In the formula, LSTM dec Characterize the decoder.

[0047] The output of the dynamic obstacle prediction model follows a Gaussian distribution, and the decoded output can be represented as

[0048] Output t =[p t ,p t+1 ,…p t+f ]

[0049] In the formula, N is Gaussian distribution, is the mean, δ t is the variance, t+1 is the next time step, and f is the total step length of the prediction.

[0050] Step 3: Plan a collision-free local path based on the location and boundaries of static obstacles, the location of dynamic obstacles in the future, and the reference path provided by the high-precision map:

[0051] The reference path refers to the path that the autonomous driving vehicle needs to travel from the starting point to the end point as a reference. This path is called the reference path. The reference path is planned based on the path, starting point and end point information collected by the high-precision map.

[0052] The specific planning method for collision-free local paths adopts dynamic programming, such as Figure 2 As shown, the frenet coordinate system is established according to the reference path, and the total length L of the local path is determined according to the current speed. loacalpat=V*t, where V is the current vehicle speed and t is the predicted time interval. Discretize the planning space according to the road width. The specific method is: rasterize with a lateral accuracy of 0.2m and a longitudinal accuracy of 0.3m, and then use spline curves to link the spatial points (for specific linking methods, please refer to "Zhu Wei, Yang Shibo, Teng Fan, etc. Research on trajectory planning of unmanned vehicles in unstructured scenes [J / OL]. Computer Science: 1-14"). When linking, the vehicle dynamics should be met. The candidate paths are screened with the goal of being closest to the reference path and without collision, and the local path that meets the conditions is obtained. The local path information contains the x, y coordinate information of the discrete points of the local path and the index value of the point (that is, the serial number of each path point), such as Figure 2 As shown, the form of the local path discrete points and the positions of the starting point and the end point are marked.

[0053] Step 4: Calculate the appropriate preview distance based on the vehicle's current position and speed, and determine the corresponding preview point on the local path:

[0054] The preview point refers to: selecting a point a certain distance away from the vehicle on the planned path (i.e., a collision-free local path) as the preview point, and controlling the vehicle to drive along the trajectory to the preview point by adjusting the steering wheel angle (front wheel angle) according to manual driving habits.

[0055] The vehicle's own position is expressed as: P veh =[X, Y, H];

[0056] X, Y, and H are the x, y coordinates and the vehicle heading angle, respectively.

[0057] The preview distance is calculated as: L dis =L int +k*V x

[0058] Where, L int is the set initial preview distance, Vx is the current speed of the vehicle, and k is the preview coefficient.

[0059] The method of matching the local path with the preview point is: starting from the starting point of the local path, the distances between the local path coordinate points are continuously accumulated in sequence to obtain the cumulative length L point , until L point >L dis Get a point on the local path and use it as the preview point, denoted as [x ahead ,y ahead , H ahead ], where x ahead ,y ahead is the x, y coordinates of the preview point, H ahead is the heading angle of the preview point, and records the index value of the point (i.e., the number of points on the local path where the preview point is located).

[0060] Step 5: Calculate the wheel deflection angle required to reach the preview point based on the difference between the vehicle heading and the preview point heading:

[0061] The difference between the vehicle heading and the preview point heading: α = |H ahead -H|;

[0062] The calculation method of wheel deflection angle is:

[0063] Where, L wheelbase The wheelbase of the vehicle.

[0064] Step 6: Calculate the next vehicle position based on vehicle kinematics. The specific position calculation method is:

[0065] x i+1 =x i +D*cos(θ i )

[0066] y i+1 =y i +D*sin(θ i )

[0067]

[0068] In the formula, x i ,y i ,θ i are the x-coordinate, y-coordinate and heading angle of the current vehicle position calculated by the step length; δ i is the discrete value of the wheel deflection angle δ; x i+1 ,y i+1 ,θ i+1 They are the x-coordinate, y-coordinate and vehicle heading angle of the next step towards the preview point, and D is the distance step.

[0069] Step 7: Determine whether the index value of the preview point exceeds the index value of the local path end point. If so, output the final result as a collision-free predicted trajectory, that is, the predicted vehicle driving trajectory; otherwise, return to step 4.

[0070] like Figure 2 The predicted trajectory shown is expressed as follows:

[0071]

[0072] In the formula, t is the current time, is the next distance step position of the vehicle.

[0073] The present invention is characterized in that it uses a preview method to predict the trajectory, which requires less computing power, is easy to implement in engineering, and does not rely on lane lines, and can accurately predict the vehicle's driving trajectory. It not only overcomes the disadvantage of the physical model prediction method that the prediction error is large due to the lack of interaction with the external environment, but also avoids the engineering problems caused by the high computing resources of the machine learning prediction method.

Claims

1. A vehicle driving trajectory prediction method for an automatic driving system, characterized in that: The following steps are involved: Step 1: Determine the location and boundaries of surrounding static obstacles through the sensor system; Step 2: Use the long short-term memory network method to predict the trajectory of the dynamic obstacle of interest and give the position of the dynamic obstacle in the future; Step 3: Based on the position and boundary of static obstacles, the position of dynamic obstacles in the future, and the reference path provided by the high-precision map, a collision-free local path is planned using a dynamic programming optimization method; Step 4: Calculate the appropriate preview distance based on the vehicle's current position and speed, and determine the corresponding preview point on the local path; The preview distance is calculated as: L dis =L int +k*V x L int is the set initial preview distance, Vx is the current speed of the vehicle, and k is the preview coefficient; The method of matching the local path with the preview point is: starting from the starting point of the local path, the distances between the local path target points are continuously accumulated in sequence to obtain the cumulative length L point， Until L point >L dis Get a point on the local path, use this point as the preview point, and record the index value of this point; Step 5: Calculate the wheel deflection angle required to reach the preview point based on the difference between the vehicle heading and the preview point heading; The calculation method of wheel deflection angle is: Where, L wheelbase is the wheelbase of the vehicle itself, α is the difference between the vehicle heading and the heading of the preview point; Step 6: Calculate the next vehicle position based on vehicle kinematics. The specific position calculation method is: x i+1 =x i +D*cos(θ i ) and i+1 =and i +D*sin(θ i ) In the formula, x i ,y i ,θ i are the x-coordinate, y-coordinate and heading angle of the current vehicle position calculated by the step length; δ i is the discrete value of the wheel deflection angle δ; x i+1 ,y i+1 ,θ i+1 are the x-coordinate, y-coordinate and vehicle heading angle of the next step towards the preview point, and D is the distance step length; Step 7: Determine whether the index value of the preview point exceeds the index value of the local path end point. If so, output the final result as a collision-free predicted trajectory; otherwise, return to step 4.

2. The vehicle driving trajectory prediction method for an automatic driving system according to claim 1 is characterized in that: In step 1, target feature recognition is performed based on the output of the vehicle vision sensor, and regular obstacles including cars, motorcycles, bicycles, and pillars are output as target boxes. The boundaries of irregular obstacles including walls and doors are outlined, and the boundary coordinate points are output to form a planning space.

3. The vehicle driving trajectory prediction method for an automatic driving system according to claim 2 is characterized in that: In step 2, the position of the dynamic obstacle in the future period of time is expressed as: Output t =[p t ,p t+1 ,…p t+f ] In the formula, N is Gaussian distribution, is the mean, is the coordinate of the starting point of the predicted trajectory at the current moment, δ t is the variance, t+1 is the next time step, and f is the total step length of the prediction.

4. The vehicle driving trajectory prediction method for an automatic driving system according to claim 3 is characterized in that: In step 3, the collision-free local path is planned as: The Frenet coordinate system is established according to the reference path, the total length of the local path is determined according to the current speed, the planning space is discretized according to the road width, and then the spatial points are linked with spline curves. The linking should meet the vehicle dynamics and the candidate paths are screened with the goal of being closest to the reference path and without collision.

Images