A path planning system and method for intelligent driving vehicles based on generative adversarial networks and imitation learning

Abstract

The present invention discloses a path planning system and method for intelligent driving vehicles based on generative adversarial networks and imitation learning, comprising two parts: a trajectory point generative adversarial network and a corresponding trajectory generation part. The trajectory point generative adversarial network takes scene features and random noise as input, and the driving trajectory of an experienced driver as the corresponding sample, and imitates and generates the horizontal and vertical trajectory point states of the final state of the experienced trajectory; the trajectory generation part uses the generated horizontal and vertical final states of the trajectory and the current horizontal and vertical final states of the vehicle to fit a quintic polynomial of the horizontal and vertical trajectory, and merges the horizontal and vertical trajectories. The present invention solves the problem that a single sample may correspond to multiple data labels due to the changing driving styles of drivers, making imitation learning difficult to train; in addition, by fitting the horizontal and vertical trajectories separately through quintic polynomials, the difficulty of imitation learning to learn the entire driving trajectory is reduced, and the smoothness of the generated trajectory is ensured. At the same time, the use of trajectory evaluation avoids the potential risks of generated trajectories.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving of intelligent vehicles, and relates to a path planning system and method for intelligent driving vehicles based on generative adversarial networks and imitation learning. Background Art

[0002] Autonomous driving, a major research area in intelligent traffic control, integrates key technologies such as perception, prediction, planning, and control. Autonomous driving technology has developed rapidly in recent years. It not only holds enormous economic potential but also offers significant advantages in improving traffic efficiency and driving safety. Path planning, as an essential technical module in autonomous driving systems, is of vital importance to the overall research of autonomous vehicles. Accurately avoiding surrounding obstacles and ensuring safe and efficient driving based on upper-layer perception and prediction results is a fundamental requirement for autonomous vehicles. A reliable autonomous driving path planning algorithm can safely avoid surrounding obstacles in real time, ensuring high safety and comfort, and significantly improving travel efficiency. Existing planning algorithms are mostly based on sampling and search methods based on artificial rules. General sampling methods such as grid methods struggle to achieve complete sampling and can only sample relatively optimal trajectories. Complete search methods, on the other hand, struggle to account for the dynamic constraints of autonomous vehicles and place high demands on the computing power of onboard computers. Therefore, most autonomous driving researchers, both in academia and industry, are currently focusing on developing more intelligent, safe, and reliable path planning methods.

[0003] Imitation learning uses machine learning to train behavioral models based on expert experience. It is primarily used to solve control planning problems in complex scenarios. Currently, a growing number of researchers are applying this approach to autonomous driving. Leveraging the empirical data of experienced drivers, a model is trained that directly maps the traffic environment to their trajectories, making it a viable approach to intelligent trajectory planning. However, experienced drivers' behaviors are often complex and diverse, and the same environmental features often correspond to multiple trajectories, posing significant challenges to conventional regression models. Summary of the Invention

[0004] To address the above issues, the present invention designs a path planning system and method for intelligent driving vehicles based on generative adversarial networks and imitation learning. By using generative adversarial networks, a trajectory planning model that imitates various types of experienced drivers is trained, which improves the driving safety and efficiency of autonomous vehicles while retaining the diversity of planned trajectories, making the driving trajectory of autonomous vehicles more intelligent.

[0005] The present invention proposes an intelligent driving car path planning system based on generative adversarial networks and imitation learning, which includes two parts: a generative adversarial network module and a corresponding trajectory generation module.

[0006] The generative adversarial network module includes a trajectory point generator and a trajectory point discriminator. The input of the trajectory point generator is random noise z and scene features y extracted from the scene dependency graph, and the output is the horizontal and vertical final states of the trajectory after time t. and The input of the trajectory point discriminator is the scene feature y extracted from the scene dependency graph and the horizontal and vertical final states generated by the randomly selected generator Or the final state label of the original experienced driver's driving trajectory The output is the judgment result True / False.

[0007] Furthermore, the random noise z is Gaussian noise. Through training, the driving styles of various experienced drivers can be mapped to Gaussian distribution. By sampling the random noise z multiple times, the generator can generate different horizontal and vertical final states.

[0008] Furthermore, the scene feature y is obtained from the environmental information O t Extracted, while environmental information O t and the final state label of the experienced driver's driving trajectory Obtained from the collected data set;

[0009] Furthermore, the specific training process of the generative adversarial network module consists of two parts. The first is the training of the discriminator, which is performed using real samples and generated data. When the discriminator can correctly distinguish between samples and generated data, the generator's ability is improved. The second is the training of the generator, which updates the generator by feeding the pseudo-final state trajectory points output by the generator into the discriminator and using the error of the discrimination results for backpropagation. When the generator cannot correctly distinguish between samples and generated data, the discriminator's ability is further improved. During the training process, the trajectory point discriminator strives to improve its ability to distinguish between the labels of experienced driver data and the pseudo-data generated by the generator. Meanwhile, the trajectory point generator strives to generate realistic lateral and vertical final states in an attempt to deceive the trajectory point generator.

[0010] The loss function of the generative adversarial network is:

[0011] min G max D loss = E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]

[0012] Among them, G is the trajectory point generator, D is the trajectory point discriminator, and x is the final state trajectory point y is the scene feature, z is the random Gaussian noise; E x～p(x)(f(x)) is the expectation of f(x) under probability p(x), where f(x) = logD(x|y); E z～p(z) (g(z)) is the expectation of g(z) under probability p(z), where g(z) = log(1-D(G(z|y))).

[0013] The corresponding trajectory generation module includes two parts: a trajectory bundle generation module and a trajectory evaluation module;

[0014] Furthermore, the trajectory bundle generation module includes a transverse trajectory generation module and a longitudinal trajectory generation module;

[0015] Furthermore, the lateral trajectory generation module uses the lateral state of the current vehicle and the final state lateral state quantity generated As a boundary condition, establish a quintic polynomial l with respect to time t trajectory :

[0016]

[0017] Furthermore, the coefficients of the quintic polynomial of the lateral trajectory solved using the boundary conditions are:

[0018]

[0019] Furthermore, the longitudinal trajectory generation module uses the longitudinal state of the current vehicle and the resulting final longitudinal state quantity As a boundary condition, establish a quintic polynomial s with respect to time t trajectory :

[0020]

[0021] Furthermore, the coefficients of the quintic polynomial of the longitudinal trajectory solved using the boundary conditions are:

[0022]

[0023] Furthermore, the fifth-order polynomial l corresponding to the plurality of lateral trajectories trajectory The fifth-order polynomial s corresponding to the longitudinal trajectory trajectory One-to-one correspondence is stored in the trajectory bundle <l trajectory , s trajectory >

[0024] Traverse all the quintic polynomial pairs of the horizontal and vertical trajectories in the trajectory bundle <l trajectory , s trajectory >, traverse the transverse trajectory quintic polynomial l at time interval Δt trajectory All track points and the longitudinal trajectory quintic polynomial s trajectory All track points Merge horizontal and vertical track points into track points Finally, all trajectory points are merged at time interval Δt to generate the final trajectory i And save it into the trajectory bundle <trajectory i >

[0025] Furthermore, the trajectory evaluation module is used to perform online evaluation on all trajectories in the trajectory bundle: traverse the trajectory bundle <trajectory i > All trajectories in the trajectory bundle are sorted by value using the set value function R, and trajectories that may cause collisions and trajectories containing trajectory points whose lateral and longitudinal speeds do not conform to the physical characteristics of the vehicle are deleted, and finally the trajectory with the best value is selected. max The system sends the trajectory to the downstream control module. When none of the trajectories in the trajectory bundle pass the evaluation, that is, there is no legal trajectory, the driver takes over to ensure safety.

[0026] The present invention proposes a method for intelligent driving vehicle path planning based on generative adversarial networks and imitation learning. The specific implementation steps of the method are described in detail in the following specific implementation manner.

[0027] Beneficial effects of the present invention:

[0028] (1) The present invention generates the planned trajectory of the autonomous vehicle through imitation learning, making the planned trajectory more intelligent and improving the path planning ability of the autonomous vehicle in complex scenarios;

[0029] (2) This invention uses a generative adversarial network to imitate experienced drivers to generate trajectory points, preserving the driving styles of various experienced drivers and increasing the diversity of generated trajectories. It also solves the problem that due to the diverse driving styles of drivers, a single sample may correspond to multiple data labels, making imitation learning difficult to train.

[0030] (3) The present invention uses a method of fitting the horizontal and vertical trajectories separately using quintic polynomials, which reduces the difficulty of learning the entire driving trajectory through imitation learning and ensures the smoothness of the generated trajectory. At the same time, the corresponding trajectory evaluation method avoids the potential risks of the generated trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flowchart of the path planning method based on generative adversarial networks and imitation learning;

[0032] Figure 2 This is the corresponding generative adversarial network structure diagram;

[0033] Figure 3 Generate flow charts for corresponding horizontal and vertical trajectories; DETAILED DESCRIPTION

[0034] The implementation of the present invention involves designing a trajectory point generative adversarial network and a corresponding trajectory generation method. The trajectory point generative adversarial network uses scene features and random noise as input, and uses the driving trajectories of experienced drivers as corresponding samples to simulate the lateral and longitudinal trajectory point states of the experienced trajectory. The trajectory generation method uses the generated lateral and longitudinal final states of the trajectory and the vehicle's current lateral and longitudinal final states to fit a quintic polynomial of the lateral and longitudinal trajectories and then merges the lateral and longitudinal trajectories.

[0035] The present invention will be further described below with reference to the accompanying drawings.

[0036] Figure 1 This is a flow chart of a path planning method for an intelligent driving car based on generative adversarial networks and imitation learning. The specific invention method process specifically includes the following steps:

[0037] 1) Through manual driving by experienced drivers of various styles, the vehicle-mounted camera, lidar and GPS navigation system are used to collect environmental information at every moment during the driving process. t , and collect the corresponding lateral trajectory point state of the vehicle after time t and longitudinal trajectory point status

[0038] 2) Process the collected environmental information and construct a planning dependency graph, which includes various factors that affect path planning, including lane lines, obstacles, pedestrians, and traffic lights;

[0039] 3) Train the generative adversarial network, with Gaussian noise z and scene features as input, and generate the state of the lateral trajectory points after time t in each scene. and longitudinal track point status

[0040] 4) Using the trained generative adversarial network, use Gaussian noise to uniformly sample the horizontal final state after time t, and according to the generated horizontal sampling points and the vehicle's current lateral state Generate the lateral trajectory quintic polynomial l trajectory ; Uniformly sample the longitudinal final state after time t, and according to the generated longitudinal sampling points and the vehicle's current longitudinal state Generate the longitudinal trajectory quintic polynomial s trajectory ; Finally, the generated horizontal and vertical trajectories are stored in the trajectory bundle one by one <l trajectory , strajectory >

[0041] 5) Horizontal and vertical track merging: traverse all horizontal and vertical track pairs in the track bundle <l trajectory , s trajectory >, traverse the lateral trajectory l at time intervals Δt trajectory All track points and longitudinal trajectory s trajectory All track points Merge horizontal and vertical track points into track points Finally, all trajectory points are merged at time interval Δt to generate the final trajectory i And save it into the trajectory bundle <trajectory i >

[0042] 6) Sort all the traces in the trace bundle by value: traverse the trace bundle <trajectory i >All trajectories in the trajectory bundle are sorted using the set value function R;

[0043] The value function R includes the following parts: 1～5 The proportional coefficient corresponding to each part of the cost:

[0044] R=k1cost speed +k2cost jerk +k3cost lateral +k4cost comfort +k5cost var

[0045] Where cost speed is the speed cost, the goal is to keep the vehicle speed at the target speed, v target is the desired target speed, t total is the number of trajectory points corresponding to the trajectory in time units, v t is the vehicle speed at time point t:

[0046]

[0047] Where cost jerk The longitudinal comfort cost is to keep the longitudinal acceleration small. t is the longitudinal acceleration at each time point:

[0048]

[0049] Where cost lateralis the lateral deviation cost, the goal is to maintain a small lateral deviation from the reference line, l t is the horizontal deviation of each time point from the reference line:

[0050]

[0051] Where cost comfort is the lateral comfort cost, the goal is to keep the lateral acceleration small, j_lateral t is the lateral acceleration at each time point:

[0052]

[0053] Where cost var is the trajectory change cost, the goal is to reduce the rate of change between the trajectory of the previous frame and the current frame, is the lateral displacement of the current trajectory at time t, is the lateral displacement of the previous frame trajectory at time t+1, is the longitudinal displacement of the current trajectory at time t, is the longitudinal displacement of the previous frame trajectory at time t+1:

[0054]

[0055] 7) Traverse the trajectories in the evaluation trajectory bundle in order of value, determine whether a collision is likely to occur and whether the lateral and longitudinal velocities of all trajectory points in the trajectory meet the dynamic characteristics of the vehicle, until the optimal legal trajectory is selected max ;

[0056] 8) When the optimal legal trajectory is selected max , and send it to the downstream control module; when none of the trajectories in the trajectory bundle pass the collision and dynamics evaluation, the driver takes over.

[0057] Figure 2 This is the corresponding generative adversarial network structure diagram, the specific structure includes a trajectory point generator and a trajectory point discriminator.

[0058] The input of the trajectory point generator is random Gaussian noise z and the scene feature y extracted from the scene dependency graph, and the output is the horizontal and vertical final state of the trajectory after time t and The input of the trajectory point discriminator is the scene feature y extracted from the scene dependency graph and the horizontal and vertical final states generated by the randomly selected generator Or the original experienced driver label The output is the judgment result True / False.

[0059] Experienced drivers have diverse driving styles, and the same driver's emotions can lead to different routes in the same scenario. This diversity is hidden within random Gaussian noise z. By sampling z, the resulting lateral and longitudinal states for the same scenario feature y also vary. Furthermore, the resulting trajectories are also diverse.

[0060] Randomly change the horizontal and vertical final states generated by the generator And the horizontal and vertical final state labels collected by experienced drivers during driving The track point discriminator is sent to determine whether the current horizontal and vertical final state is the horizontal and vertical final state label collected by the experienced driver when driving.

[0061] During training, the trajectory point discriminator strives to improve its ability to distinguish between labels from experienced driver data and pseudo-data generated by the generator. The trajectory point generator, on the other hand, attempts to create realistic lateral and longitudinal final states in an attempt to deceive the trajectory point generator. The specific training process consists of two parts: first, training the discriminator using real samples and generated data; and second, training the generator, which feeds the pseudo-final state trajectory points output by the generator into the discriminator and uses the error from the discrimination results for backpropagation to update the trained generator. Therefore, the loss function of the entire network consists of two parts, as follows:

[0062] min G max D loss = E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))]

[0063] Among them, G is the trajectory point generator, D is the trajectory point discriminator, and x is the final state trajectory point y is the scene feature, z is the random Gaussian noise; E x～p(x) (f(x)) is the expectation of f(x) under probability p(x), f(x) = logD(x|y); E z～p(z) (g(z)) is the expectation of g(z) under probability p(z), g(z) = log(1-D(G(z|y))).

[0064] Figure 3 Generate a flow chart for the corresponding horizontal and vertical trajectories, which specifically includes the following steps:

[0065] 1) The surrounding environment information is extracted in the form of an image, and the feature y is fed into the generative adversarial network;

[0066] 2) Generate the generator of the adversarial network, which generates the final state trajectory points based on the samples uniformly sampled in Gaussian noise z and the current traffic scene features y

[0067] 3) Use the generator to generate several final state trajectory points Generate horizontal and vertical trajectories. For the horizontal trajectory, use the current vehicle's lateral state and the final state lateral state quantity generated As a boundary condition, there is a quintic polynomial l with respect to time t trajectory , where t0 is the starting time point:

[0068]

[0069] The boundary conditions are:

[0070]

[0071] According to the fifth-order polynomial of the lateral trajectory and the boundary conditions:

[0072]

[0073] According to the obtained a1a2a3a4a5, the lateral trajectory fifth-order polynomial l can be obtained trajectory . Multiple hidden states uniformly sampled from the Gaussian distribution z can generate multiple different Further, multiple different lateral trajectory quintic polynomials l can be generated trajectory ;

[0074] For the longitudinal trajectory, the longitudinal state of the current vehicle is used and the resulting final longitudinal state quantity As a boundary condition, there is a fifth-degree polynomial s with respect to time t trajectory :

[0075]

[0076] The boundary conditions are:

[0077]

[0078] According to the fifth-order polynomial of the longitudinal trajectory and the boundary conditions:

[0079]

[0080] According to the obtained b1b2b3b4b5, the longitudinal trajectory quintic polynomial s can be obtained trajectory . Multiple hidden states uniformly sampled from the Gaussian distribution z can generate multiple different Further, multiple different longitudinal trajectory quintic polynomials s can be generated trajectory ;

[0081] The generated quintic polynomials of several horizontal trajectories l trajectory and the longitudinal trajectory quintic polynomial s trajectory One-to-one correspondence is stored in the trajectory bundle <l trajectory , s trajectory >.

[0082] The series of detailed descriptions listed above are only specific descriptions of feasible implementation methods of the present invention. They are not intended to limit the scope of protection of the present invention. Any equivalent methods or changes that do not deviate from the technology of the present invention should be included in the scope of protection of the present invention.

Claims

1. A path planning system for intelligent driving vehicles based on generative adversarial networks and imitation learning, characterized by: Includes a generative adversarial network module and a corresponding trajectory generation module; The generative adversarial network module includes a trajectory point generator and a trajectory point discriminator; The input of the trajectory point generator is random noise z and the scene feature y extracted from the scene dependency graph, and the output is the horizontal and vertical final state of the trajectory after time t and The input of the trajectory point discriminator is the scene feature y extracted from the scene dependency graph and the horizontal and vertical final states generated by the randomly selected generator Or the final state label of the original experienced driver's driving trajectory The output is the judgment result True / False; The corresponding trajectory generation module includes a trajectory bundle generation module and a trajectory evaluation module; the trajectory bundle generation module includes a lateral trajectory generation module and a longitudinal trajectory generation module; The lateral trajectory generation module: uses the lateral state of the current vehicle and the final state lateral state quantity generated As a boundary condition, establish a quintic polynomial l with respect to time t trajectory : The longitudinal trajectory generation module uses the longitudinal state of the current vehicle and the resulting final longitudinal state quantity As a boundary condition, establish a quintic polynomial s with respect to time t trajectory : The fifth-order polynomial l of the transverse trajectory trajectory and the fifth-order polynomial s of the longitudinal trajectory trajectory One-to-one correspondence is stored in the trajectory bundle <l trajectory ,s trajectory > Traverse all horizontal and vertical track pairs in the track bundle <l trajectory ,s trajectory >, traverse the lateral trajectory at time interval Δt by the fifth-order polynomial l trajectory All track points and the fifth-order polynomial s of the longitudinal trajectory trajectory All track points Merge horizontal and vertical track points into track points Finally, all trajectory points are merged at time interval Δt to generate the final trajectory i And save it into the trajectory bundle <trajectory i > The trajectory evaluation module: traverses the trajectory bundle <trajectory i > All trajectories in the trajectory bundle are sorted by value using the set value function R, and trajectories that may cause collisions and trajectories containing trajectory points whose lateral and longitudinal speeds do not conform to the physical characteristics of the vehicle are deleted, and finally the trajectory with the best value is selected. max The system sends the trajectory to the downstream control module. When none of the trajectories in the trajectory bundle pass the evaluation, that is, there is no legal trajectory, the driver takes over.

2. The intelligent driving vehicle path planning system based on generative adversarial networks and imitation learning according to claim 1 is characterized in that: The random noise z is Gaussian noise. Through training, the driving styles of various experienced drivers can be mapped to Gaussian distribution. By sampling the random noise z multiple times, the generator can generate different lateral and longitudinal final states.

3. The intelligent driving vehicle path planning system based on generative adversarial networks and imitation learning according to claim 1 is characterized in that: The scene feature y is derived from the environmental information O t Extracted, while environmental information O t and the final state label of the experienced driver's driving trajectory Obtained from the collected dataset.

4. The intelligent driving vehicle path planning system based on generative adversarial networks and imitation learning according to claim 1, characterized in that: The training of the generative adversarial network module consists of two parts: first, the training of the discriminator, which uses real samples and generated data for training. When the discriminator can correctly distinguish between samples and generated data, the generator's ability is improved; The second is the training of the generator. The pseudo-final state trajectory output by the generator is input to the discriminator, and the error of the discrimination result is used for backpropagation to update the training generator. When the generator cannot correctly distinguish between samples and generated data, the discriminator's ability is further improved. During training, the trackpoint discriminator strives to improve its ability to distinguish between labels from experienced driver data and pseudo data generated by the generator; the trackpoint generator attempts to generate realistic lateral and longitudinal final states in an attempt to deceive the trackpoint generator. The loss function of the generative adversarial network is: min G max D loss=E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))] Among them, G is the trajectory point generator, D is the trajectory point discriminator, and x is the final state trajectory point y is the scene feature, and z is random Gaussian noise.

5. The intelligent driving vehicle path planning system based on generative adversarial networks and imitation learning according to claim 1 is characterized in that: The coefficients of the quintic polynomial of the transverse trajectory and the quintic polynomial of the longitudinal trajectory are solved using the boundary conditions:

6. A path planning method for intelligent driving vehicles based on generative adversarial networks and imitation learning, characterized in that: include: S1 uses various styles of experienced drivers to manually drive and uses on-board cameras, lidar and GPS navigation to collect environmental information at every moment during the driving process. t , and collect the corresponding lateral trajectory point state of the vehicle after time t and longitudinal trajectory point status S2 processes the collected environmental information and constructs a planning dependency graph, which includes various factors that affect path planning, including lane lines, obstacles, pedestrians, and traffic lights. S3 trains a generative adversarial network with Gaussian noise z and scene features as input. The generated object is the state of the lateral trajectory point after time t in each scene. and longitudinal track point status S4 uses the trained generative adversarial network to uniformly sample the lateral final state after time t using Gaussian noise, and then generates the lateral sampling points according to the generated and the vehicle's current lateral state Generate the lateral trajectory quintic polynomial l trajectory ; Uniformly sample the longitudinal final state after time t, and according to the generated longitudinal sampling points and the vehicle's current longitudinal state Generate the longitudinal trajectory quintic polynomial s trajectory ; Finally, the generated horizontal and vertical trajectories are stored in the trajectory bundle one by one <l trajectory ,s trajectory > S5 horizontal and vertical track merging: traverse all horizontal and vertical track pairs in the track bundle <l trajectory ,s trajectory >, traverse the transverse trajectory quintic polynomial l at time interval Δt trajectory All track points and the longitudinal trajectory quintic polynomial s trajectory All track points Merge horizontal and vertical track points into track points Finally, all trajectory points are merged at time interval Δt to generate the final trajectory i And save it into the trajectory bundle <trajectory i > S6 sorts all the traces in the trace bundle by value: traverse the trace bundle <trajectory i >All trajectories in the trajectory bundle are sorted using the set value function R; S7 traverses the trajectories in the evaluation trajectory bundle in order of value, determines whether a collision is possible and whether the lateral and longitudinal speeds of all trajectory points in the trajectory conform to the dynamic characteristics of the vehicle, until the optimal legal trajectory is selected. max ; S8 selects the best legal trajectory max , and send it to the downstream control module; when none of the trajectories in the trajectory bundle pass the collision and dynamics evaluation, the driver takes over.

7. The intelligent driving vehicle path planning method based on generative adversarial network and imitation learning according to claim 6 is characterized in that: In S3, the generative adversarial network includes a trajectory point generator and a trajectory point discriminator; The input of the trajectory point generator is random Gaussian noise z and the scene feature y extracted from the scene dependency graph, and the output is the horizontal and vertical final state of the trajectory after time t and The input of the trajectory point discriminator is the scene feature y extracted from the scene dependency graph and the horizontal and vertical final states generated by the randomly selected generator Or the original experienced driver label The output is the judgment result True / False; Randomly generate the horizontal and vertical final states of the generator And the horizontal and vertical final state labels collected by experienced drivers during driving The track point discriminator is sent to determine whether the current horizontal and vertical final state is the horizontal and vertical final state label collected by the experienced driver when driving.

8. The intelligent driving vehicle path planning method based on generative adversarial network and imitation learning according to claim 7 is characterized in that: During the training process of the generative adversarial network, the trajectory point discriminator strives to improve its ability to distinguish between the labels of experienced driver data and the pseudo data generated by the generator; while the trajectory point generator strives to generate realistic lateral and longitudinal final states in an attempt to deceive the trajectory point generator. The specific training process includes two parts: the first is the training of the discriminator, which is trained using real samples and generated data; the second is the training of the generator, which sends the pseudo final state trajectory points output by the generator to the discriminator and uses the error of the discrimination result for backpropagation to update the trained generator. The loss function of the entire network consists of two parts, as follows: min G max D loss=E x～p(x) [logD(x|y)]+E z～p(z) [log(1-D(G(z|y)))] Among them, G is the trajectory point generator, D is the trajectory point discriminator, and x is the final state trajectory point y is the scene feature, and z is random Gaussian noise.

9. The intelligent driving vehicle path planning method based on generative adversarial network and imitation learning according to claim 6, characterized in that: The lateral trajectory quintic polynomial l is generated in S4 trajectory and the longitudinal trajectory quintic polynomial s trajectory The method is as follows: S4.1 extracts features y from surrounding environmental information in the form of images and feeds them into the generative adversarial network; S4.2 Generates the generator of the adversarial network, which generates the final state trajectory points based on the samples uniformly sampled in Gaussian noise z and the current traffic scene features y S4.3 Using the generator to generate several final state trajectory points Generate horizontal and vertical trajectories: For the lateral trajectory, the lateral state of the current vehicle is used and the final state lateral state quantity generated As a boundary condition, there is a quintic polynomial l with respect to time t trajectory : The boundary conditions are: According to the fifth-order polynomial of the lateral trajectory and the boundary conditions: According to the obtained a1a2a3a4a5, the lateral trajectory fifth-order polynomial l can be obtained trajectory , multiple hidden states uniformly sampled from the Gaussian distribution z can generate multiple different Further, multiple different lateral trajectory quintic polynomials l can be generated trajectory ; For the longitudinal trajectory, the longitudinal state of the current vehicle is used and the resulting final longitudinal state quantity As a boundary condition, there is a fifth-degree polynomial s with respect to time t trajectory : The boundary conditions are: According to the fifth-order polynomial of the longitudinal trajectory and the boundary conditions: According to the obtained b1b2b3b4b5, the longitudinal trajectory fifth-order polynomial s can be obtained trajectory , multiple hidden states uniformly sampled from the Gaussian distribution z can generate multiple different Further, multiple different longitudinal trajectory quintic polynomials s can be generated trajectory ; The generated quintic polynomials of several horizontal trajectories l trajectory and the longitudinal trajectory quintic polynomial s trajectory One-to-one correspondence is stored in the trajectory bundle <l trajectory ,s trajectory >.

10. The intelligent driving vehicle path planning method based on generative adversarial network and imitation learning according to claim 6, characterized in that: The value function R in S6 is designed as follows: R=k1cost speed +k2cost jerk +k3cost lateral +k4cost comfort +k5cost var in cost speed is the speed cost, the goal is to keep the vehicle speed at the target speed, v target is the desired target speed, v t The speed of the vehicle at each time point: Where cost jerk The longitudinal comfort cost is to keep the longitudinal acceleration small. t is the longitudinal acceleration at each time point: Where cost lateral is the lateral deviation cost, the goal is to maintain a small lateral deviation from the reference line, l t is the horizontal deviation of each time point from the reference line: Where cost comfort is the lateral comfort cost, the goal is to keep the lateral acceleration small, j_lateral t is the lateral acceleration at each time point: Where cost var is the trajectory change cost, the goal is to reduce the rate of change between the trajectory of the previous frame and the current frame, is the lateral displacement of the current trajectory at time t, is the lateral displacement of the previous frame trajectory at time t+1, is the longitudinal displacement of the current trajectory at time t, is the longitudinal displacement of the previous frame trajectory at time t+1:

Images