Method and device for autonomous driving trajectory planning considering perception uncertainty, and medium

Abstract

The application relates to an automatic driving track planning method and device considering perception uncertainty and a medium; the method comprises the following steps: constructing a probabilistic target detection model to perform environment perception, and constructing an obstacle buffer zone; a track planning algorithm based on a soft constraint convex feasible set is constructed to generate an optimal planning track; the application models obstacles in a driving environment by using a multivariate Gaussian distribution, establishes a probabilistic perception model, performs high-precision perception and variance inference on the surrounding environment; the application constructs a maximum shaking area of the perceived obstacles by using envelope box attribute variance information, establishes an obstacle buffer zone, can quantize and transfer the perception uncertainty, and provides more reliable perception data support for downstream modules; a relaxation factor is added to a track planning objective function, so that the planning process can adapt to the uncertain environment of perception, and the smoothness of the planning track and the adaptability and driving comfort of the vehicle in the uncertain environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent networked vehicles, in particular to an automatic driving trajectory planning method and device considering perception uncertainty and a medium. BACKGROUND

[0002] In recent years, automatic driving technology has made significant progress and has shown excellent performance in many typical driving tasks. However, when dealing with complex and dynamic traffic scenes, the performance of the automatic driving system still has a lot of room for improvement. Among them, factors such as bad weather, sensor noise and limitations of the algorithm itself often cause the perception module to produce jitter and even false detection uncertainty in many scenarios, making the system's understanding of the surrounding environment uncertain and unstable. In actual engineering development, the decision planning module often completely trusts the reasoning results of the perception module, so when dealing with the uncertainty of perception, it tends to adopt an overly conservative strategy to reduce potential risks and ensure driving safety.

[0003] However, when the vehicle is driving in an obstacle-dense area, frequent conservative avoidance not only brings significant computational burden to the decision planning algorithm, increases the computation time and resource consumption, but also may cause the driving decision to be too cautious, significantly reducing the passenger's driving comfort. At the same time, this strategic conservatism may even cause a contradiction between safety and efficiency in certain situations, especially in crowded urban streets or complex interaction scenarios, affecting the overall performance and practicality of the system.

[0004] Therefore, how to deal with perception uncertainty in complex environments, improve the flexibility of planning strategies and the smoothness of planned trajectories while ensuring safety is one of the key problems to be solved in the development of current automatic driving technology. In view of the above problems, the present application proposes an automatic driving trajectory planning method considering perception uncertainty, which effectively solves the multi-objective optimization between safety and comfort in trajectory planning, more intelligently handles the perception uncertainty problem of the automatic driving system, and improves its adaptability and passenger experience in actual scenarios. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the above-mentioned problems existing in the prior art, and provide an automatic driving trajectory planning method, device and medium considering perception uncertainty.

[0006] To solve the above technical problems, the present application is implemented by adopting the following technical solutions, which are described in combination with the accompanying drawings as follows:

[0007] It should be noted that, in this document, relational terms such as first and second and the like can only be used to distinguish one entity or action from another entity or action, and do not necessarily require or imply that these entities or actions exist in any actual relationship or order. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles, or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or other elements inherent in such processes, methods, articles, or devices.

[0008] An automatic driving trajectory planning method considering perception uncertainty, comprising:

[0009] A probabilistic object detection model is constructed for environment perception, and an obstacle buffer zone is constructed; a trajectory planning algorithm based on soft constraint convex feasible set is constructed to generate an optimal planning trajectory.

[0010] Further, the probabilistic object detection model includes four parts: an environment encoder, a category decoder, an envelope box decoder, and a variance decoder; when the environment information X is input into the probabilistic object detection model, the environment encoder performs feature extraction to obtain high-order features of the environment information, and the category decoder, the envelope box decoder, and the variance decoder obtain obstacle categories, envelope box attributes, and variance information, respectively; a data flow is represented as:

[0011]

[0012] wherein, represents high-level semantic information of the environment information, represents a predicted obstacle category, represents predicted obstacle envelope box information, f encD(·; θ) is the encoding process at parameter θ, D cls D(·; θ) is the category decoding process at parameter θ, D reg D(·; θ) is the bounding box decoding process at parameter θ, D var D(·; θ) is the variance decoding process at parameter θ, and a softplus layer is added at the output end to ensure that the output is always positive. is the covariance matrix of the obstacles predicted by the model;

[0013] Further, each attribute of the bounding box is independent of each other, and the covariance matrix is a 7x7 diagonal matrix.

[0014] Further, the probabilistic object detection model is constructed based on a multivariate Gaussian distribution.

[0015] Assuming that the bounding box attribute information of the obstacles in the environment obeys a multivariate Gaussian distribution, the probabilistic object detection model outputs the parameters of the obstacle distribution, and the mean vector of the multivariate Gaussian distribution is used as the attribute information of the bounding box The variance information constitutes the covariance matrix of the distribution is expressed as:

[0016]

[0017] where θ represents the network parameters, B represents the true value of the obstacle attribute, |·| represents the determinant of the matrix, and ·T represents the transpose operation of the matrix.

[0018] The label information of the data set is regarded as a multivariate Gaussian distribution with a group of covariance matrices being all 0 matrices, and is regarded as a Dirac distribution, which is expressed as:

[0019]

[0020] where P G represents the distribution of the label, and δ(·) represents the Dirac function.

[0021] By optimizing the network parameters θ, the difference between the distribution predicted by the model and the true distribution is minimized, so that the model has the ability of environment perception and variance reasoning, which is expressed as:

[0022]

[0023] where represents the Kullback-Leibler (KL) divergence between the two distributions;

[0024] The KL divergence is used as an index to measure the difference between the two distributions, and the difference is used as the learning and optimization target of the neural network, and a loss function is constructed The regression task and the variance inference task of each variance inference neural network are optimized; the loss function is represented as:

[0025]

[0026] wherein, B represents the number of obstacles detected by the model; B n B represents the true value of the nth envelope box, and respectively represent the nth envelope box attribute information and the covariance matrix predicted by the model, and n represents the serial number index of the obstacle.

[0027] Further, the constructed probabilistic target detection model is trained using a data set, and the loss function includes three parts, namely the category loss, the regression loss and the direction loss; Focal Loss is selected as the category loss:

[0028]

[0029] wherein, represents the classification loss function, p a is the softmax output of the anchor in the category decoder, and a and g are empirical parameters in the loss function;

[0030] The loss function of the probabilistic target detection model is represented as:

[0031]

[0032] wherein, represents the total loss function of the model, β cls , β KL and β dir are the balance coefficients of the three loss categories, is a softmax classification loss.

[0033] After training the probabilistic target detection model for multiple rounds, a probabilistic target detection model with environmental perception and variance inference capabilities is constructed, and when environmental information X is input into the model, a probabilistic perception result is obtained:

[0034]

[0035] wherein, represents the set of perception results predicted by the model;

[0036] The encoder and the decoder of the model gradually have the ability to extract and predict the features of the point cloud information.

[0037] Further, an obstacle buffer zone is constructed, and the specific content is as follows:

[0038] Based on the distribution assumption of the probabilistic target detection model and the variance information of the output, a buffer zone of the obstacle is set;

[0039] Get the probabilistic target detection result

[0040]

[0041] By It can be seen that the standard deviation of each attribute of the nth obstacle envelope box is:

[0042]

[0043] Wherein, n represents the ID of the obstacle; σ x , σ y and σ z are the standard deviations of the obstacle center point in the x, y and z axes respectively; σ l , σ w and σ h are the standard deviations of the length, width and height of the obstacle envelope box respectively; σ θ is the standard deviation of the heading angle of the obstacle.

[0044] The envelope box attributes of all perceived obstacles should follow the 3σ principle, and the jitter range of each attribute information should be between ±3σ value, that is:

[0045] x∈[x-3σ x , x+3σ x ] l∈[l-3σ l , l+3σ l ]

[0046] y∈[y-3σ y , y+3σ y ] w∈[w-3σ w , w+3σ w ] θ∈[θ-3σ θ , θ+3σ θ ]

[0047] z∈[z-3σ z , z+3σ z ] h∈[h-3σ h , h+3σ h ]

[0048] Select the envelope box, translate it in the coordinate system under the BEV perspective, and select the union as the obstacle expansion envelope box. The length and width of the expansion envelope box are respectively:

[0049]

[0050] where l and w represent the length and width of the obstacle envelope predicted by the perception module, respectively, l e and w e represent the length and width of the expanded obstacle envelope, respectively.

[0051] Further, a trajectory planning algorithm based on soft constraint convex feasible set is constructed, and the specific content is as follows:

[0052] Define the planning trajectory point sequence as

[0053]

[0054] wherein, represents the coordinates of the vehicle planning trajectory, H represents the planning step of preview, p0 represents the current position of the vehicle;

[0055] Solve the optimal trajectory of the vehicle, denoted as:

[0056]

[0057] s.t.Γ(S)=∩ n Γ n (S)=∩ n {P:φ n (P)≥-S}

[0058] wherein J(P) represents the objective function in the trajectory planning process, ω||S|| 2 is the penalty term for invading the soft constraint, P * , S * represents the optimal state in the state space, φ n (P) represents the distance from the nth obstacle, Γ n (·) represents the constraint condition for the nth obstacle, and Γ(·) represents the union of all ;

[0059] The state variable is expanded as

[0060]

[0061] The feasible set is denoted as:

[0062]

[0063] After sorting, we get

[0064]

[0065] wherein Z represents an intermediate variable in the reasoning process, represents the feasible set of the algorithm under the nth obstacle, Let L represent the feasible set in the entire environment, k represent the k-th iteration, and L and S represent the coefficients in the derivation process.

[0066] In each iteration, the algorithm starts from the current solution and finds a new solution in the globally convex feasible set, causing the state variables to move in the optimal direction.

[0067]

[0068] The optimal state Z is obtained by iteratively solving the state variables using the convex feasible set algorithm, where the difference between the number of iterations is less than a threshold. * Thus, the optimal output trajectory P is obtained. * .

[0069] Furthermore, before constructing a probabilistic target detection model for environmental perception, the following steps are required:

[0070] Step 1: The vehicle uses onboard sensors to collect environmental information during the driving process;

[0071] Step 2: Label the data collected in Step 1 to form a labeled dataset;

[0072] Step 3: Divide the dataset from Step 2 into a training set, a validation set, and a test set;

[0073] In steps one, two, and three, the vehicle uses onboard sensors to collect and label environmental information during its journey. The specific details are as follows:

[0074] Intelligent vehicles are equipped with various onboard sensors that can perceive the surrounding environment and acquire environmental information, denoted as X. Each piece of environmental information X contains N obstacles, denoted as:

[0075]

[0076] in, Represents a list of obstacles, o n Let N represent the nth obstacle, and N represent the number of obstacles.

[0077] Dataset D includes information on the type of each obstacle. n and 3-D envelope information B n ={x n y n , z n w n , l n h n θ n}; where, (x n y n , zn ) represents the position information of the obstacle o n n , l n , h n , theta n ) represents the attribute and direction of the 3-D bounding box of the obstacle o n

[0078] The obstacle label is defined as:

[0079]

[0080] The above data set is divided into a training set, a validation set and a test set.

[0081] An apparatus comprising one or more processors;

[0082] a memory for storing one or more programs;

[0083] When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0084] A computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the method as described above.

[0085] The beneficial effects of the present application compared with the prior art are:

[0086] 1. The present application proposes a trajectory planning method considering perception uncertainty. The method models the obstacles in the driving environment using multivariate Gaussian distribution, establishes a probabilistic perception model, and performs high-precision perception and variance inference on the surrounding environment.

[0087] 2. The present application is based on the 3σ principle, uses the envelope box attribute variance information to construct the maximum shaking area of the perceived obstacle, thereby establishing the obstacle buffer zone, which can quantify and transfer the perception uncertainty, provide more reliable perception data support for the downstream modules, and help improve the safety and stability of vehicle decision-making in complex environments.

[0088] 3. The present application uses a trajectory planning algorithm that fuses soft constraint convex feasible set, adds a relaxation factor to the trajectory planning objective function, adapts to the uncertain environment of perception during planning, and improves the smoothness of the planned trajectory and the adaptability and driving comfort of the vehicle in uncertain environments. BRIEF DESCRIPTION OF DRAWINGS

[0089] The present application will be further described below in conjunction with the accompanying drawings:

[0090] Figure 1 The technical roadmap of the present application​​

[0091] Figure 2 schematic diagram of an obstacle buffer zone DETAILED DESCRIPTION

[0092] For the purposes of the present application, the technical solutions and advantages will be more apparent, the technical solutions in the embodiments of the present application will be described in more detail below with reference to the drawings of the embodiments of the present application. In the drawings, the same or similar reference numbers represent the same or similar elements or elements having the same or similar functions throughout. The described embodiments are part of the embodiments of the present application, not all of the embodiments. The embodiments described below with reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. The embodiments of the present application will be described in detail below with reference to the drawings.

[0093] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the scope of protection of the present application.

[0094] The present application will be described in detail below with reference to the drawings:

[0095] As Figure 2 shown, the present application proposes an automatic driving trajectory planning method considering perception uncertainty, to deal with the multi-objective solution problem between safety and comfort of trajectory planning under the condition of uncertain perception. First, based on the multivariate Gaussian distribution, a probabilistic target detection model is constructed to perceive and infer the variance of obstacles in the environment with high precision. Second, based on the obstacle perception results, the obstacle perception jitter range is constructed in space, and the obstacle buffer zone is constructed. Finally, a soft constraint convex feasible set planning algorithm is used to consider the perception uncertainty of obstacles in space, to introduce a relaxation variable in the objective function, to solve the optimal trajectory, to ensure the safety and smoothness of the planned trajectory. Specifically, it includes:

[0096] Step one: the vehicle uses the vehicle-mounted sensor to collect the environmental information in the driving process;

[0097] Step two: label the data collected in step one to form a labeled data set;

[0098] Step three: divide the data set in step two into training set, validation set and test set;

[0099] In steps one, two and three, the intelligent car is equipped with various vehicle-mounted sensors, which can perceive the surrounding environment, obtain environmental information, denoted as X, and each environmental information X contains N obstacles, denoted as:

[0100]

[0101] Among them, represents the obstacle list, o n represents the nth obstacle, and N represents the number of obstacles.

[0102] The data set D includes the category information c n of each obstacle, such as pedestrians, bicycles and cars, and 3-D bounding box information B n = {x n , y n , z n , w n , l n , h n , θ n}. Among them, (xn ， y n , z n ) represents the position information of the obstacle o n , (w n , l n , h n , θ n ) represents the attributes and directions of the 3-D bounding box of the obstacle o n .

[0103] Therefore, in each X, the obstacle label is defined as:

[0104]

[0105] The above data set is divided into training set, validation set and test set according to a certain proportion (for example, 8:1:1).

[0106] Step four: based on multivariate Gaussian distribution, construct a probabilistic target detection model:

[0107] A probabilistic target detection model with variance inference capability is constructed. The model can be simplified into the following four parts: environment encoder, class decoder, bounding box decoder and variance decoder. When the environment information X is input into the model, the environment encoder extracts the high-order features of the environment information, and the decoder further obtains the obstacle category, bounding box attribute and variance, etc. The data flow can be represented as:

[0108]

[0109] where, is the high-level semantic information representing the environment information, is the predicted obstacle category, is the predicted obstacle bounding box information, f enc (·; θ) is the encoding process under parameter θ, D cls (·; θ) is the category decoding process under parameter θ, D reg (·; θ) is the bounding box decoding process under parameter θ, D var (·; θ) is the variance decoding process under parameter θ, and a softplus layer is added at the output end to ensure that the output result is always positive. is the covariance matrix of the predicted obstacle by the model, and in the present application, it is assumed that each attribute of the bounding box is independent of each other, so the covariance matrix is a 7x7 diagonal matrix.

[0110] The probabilistic target detection model represents the target detection model being constructed in step four, which is different from the ordinary target detection model and can output the probability distribution of the detected target. The output is also the data basis for the subsequent steps.

[0111] Since there is a lack of label information on uncertainty in the existing data set, the model cannot be directly trained in parameters by some supervised learning algorithm. Therefore, it is assumed that the bounding box attribute information of the obstacle in the environment obeys a multivariate Gaussian distribution, the mean vector of the multivariate Gaussian distribution is taken as the attribute information of the bounding box and the variance information constitutes the covariance matrix of the distribution which can be expressed as:

[0112]

[0113] where θ represents the network parameters, which are a set of learnable variables, and B represents the true value of the obstacle attribute. |·| represents the determinant of the matrix, and · T represents the transpose operation of the matrix.

[0114] Obviously, the label information of the data set can be regarded as a multivariate Gaussian distribution with a set of covariance matrices being all 0 matrices, which is regarded as a Dirac distribution, and can be expressed as:

[0115]

[0116] where P G represents the distribution of the label, and δ(·) represents the Dirac function.

[0117] By optimizing the network parameters θ, the difference between the model predicted distribution and the real distribution is minimized, which enables the model to have environmental perception and variance inference capabilities, which can be represented as:

[0118]

[0119] wherein, represents the Kullback-Leibler (KL) divergence between the two distributions. The "model" in "minimizing the model predicted distribution" refers to the probabilistic object detection model being constructed.

[0120] Using KL divergence as an indicator to measure the difference between two distributions, the difference is used as the learning optimization target of the neural network, and the loss function is constructed Optimize the regression task and variance inference task of each variance inference neural network. The loss function can be represented as:

[0121]

[0122] wherein, represents the number of obstacles detected by the model. B n represents the true value of the nth envelope box, and respectively represent the nth envelope box attribute information and the covariance matrix predicted by the model, and n represents the serial number index of the obstacle. The "model" in "the number of obstacles detected by the model" specifically refers to the probabilistic object detection model being constructed.

[0123] Step five: use the data set to train the target detection model:

[0124] Use the data set to train the probabilistic object detection model constructed in step four, and the loss function contains three parts, which are class loss, regression loss and direction loss. Here, Focal Loss is selected as the class loss:

[0125]

[0126] wherein, represents the classification loss function, p a is the softmax output of the anchor in the class decoder, and α, γ are empirical parameters in the loss function.

[0127] In summary, the loss function of the model can be represented as:

[0128]

[0129] wherein, represents the total loss function of the model, βcls , β KL and β dir are balance coefficients of three loss categories, is a softmax classification loss.

[0130] After training the model for multiple rounds, a probabilistic target detection model is successfully constructed, which has the ability of environmental perception and variance inference. When the environmental information X is input into the model, the probabilistic perception result can be obtained:

[0131]

[0132] wherein, represents the set of perception results predicted by the model.

[0133] The encoder and decoder of the model gradually have the ability of feature extraction and prediction of point cloud information, and can achieve good perception effect.

[0134] Step six: using the probabilistic target detection model constructed in step five to perform environmental perception and constructing an obstacle buffer zone:

[0135] In real traffic scenarios, there are many traffic participants, and they are dynamic and time-varying, and there are complex topological relationships between each other. The perception model of intelligent vehicles often appears jittering or even jumping in time sequence, and other perception uncertainty phenomena. In order to ensure the safety of vehicle driving, based on the distribution assumption of the probabilistic target detection model in step five and the variance information output by the model, an obstacle buffer zone is set.

[0136] In step five, the probabilistic target detection result obtained is

[0137]

[0138] It can be seen that the standard deviation of each attribute of the nth obstacle envelope box is:

[0139]

[0140] wherein, n represents the ID of the obstacle; σ x , σ y and σ z are the standard deviations of the obstacle center point in the x, y and z axes respectively; σ l , σ w and σ h are the standard deviations of the length, width and height of the obstacle envelope box respectively; σ θ is the standard deviation of the heading angle of the obstacle.

[0141] The envelope box attributes of all perceived obstacles should follow the 3σ principle, and the jitter range of each attribute information should be between ±3σ values, that is:​

[0142] x∈[x-3σ x ,x+3σ x l∈[l-3σ] l ,l+3σ l ]

[0143] y∈[y-3σ y y+3σ y w∈[w-3σ] w w+3σ w ] θ∈[θ-3σ θ ,θ+3σ θ ]

[0144] z∈[z-3σ z ,z+3σ z h∈[h-3σ] h h+3σ h ]

[0145] Therefore, appendix Figure 1 As shown, the largest bounding box (dashed box in the figure) is selected, and a translation transformation is performed in the coordinate system under the BEV perspective. The union of these transformations is then selected as the obstacle-extended bounding box. For simplicity, a new rectangle B is constructed. e (The outer solid-line rectangle in the image) serves as the new expanded envelope. Based on geometric reasoning, the length and width of the expanded envelope are:

[0146]

[0147] Where l and w represent the length and width of the obstacle envelope predicted by the perception module, respectively. e and w e These represent the length and width of the expanded obstacle envelope, respectively. B e It is input into the planning algorithm of the downstream planning module.

[0148] Step 7: Construct a trajectory planning algorithm based on soft-constrained convex feasible sets:

[0149] Step 8: Generate the optimal planning trajectory.

[0150] In traditional planning algorithms, obstacles are usually treated as a kind of hard constraint, and the vehicle must avoid any spatial overlap with them. In this invention, step six designs a dynamic buffer zone for perceiving obstacles, and outputs the perception results to the planning module in a very conservative paradigm. If it is completely treated as a hard constraint, it will greatly reduce the drivable range of the autonomous vehicle and the comfort of the passengers. Therefore, a trajectory planning algorithm based on soft constraint convex feasible set is adopted in step seven, and the obstacle buffer zone is treated as a kind of soft constraint, so that the vehicle can slightly invade the buffer zone without affecting safety, so as to more calmly cope with the jitter changes of the perceived obstacles, and significantly improve the smoothness of the trajectory.

[0151] In the trajectory planning problem of autonomous vehicles, define the planning trajectory point sequence as

[0152]

[0153] wherein, represents the coordinates of the vehicle planning trajectory, H represents the planning step of preview, p0 represents the current position of the vehicle.

[0154] In order to solve the optimal trajectory of the vehicle, this problem is modeled as a convex feasible set optimization problem with relaxation variables, which can be represented as:

[0155]

[0156] s.t.Γ(S)=∩ n Γ n (S)=∩ n {P:φ n (P)≥-S}

[0157] wherein, J(P) represents the objective function in the trajectory planning process, ω||S|| 2 is the penalty term for invading the soft constraint, P * , S * represents the optimal state in the state space, φ n (P) represents the distance from the nth obstacle, Γ n (·) represents the constraint condition for the nth obstacle, Γ(·) represents the union of all

[0158] Compared with the traditional convex optimization feasible set algorithm, in this problem, the state variable is expanded to

[0159]

[0160] Therefore, the feasible set is represented as:

[0161] ​

[0162] The arrangement is obtained

[0163]

[0164] Wherein, Z represents an intermediate variable in the reasoning process, represents the feasible set of the algorithm under the nth obstacle, represents the feasible set in the entire environment, k represents the kth iteration process, and L and S represent the coefficients in the derivation process.

[0165] In each iteration, the algorithm starts from the current solution and finds a new solution in the global convex feasible set, so that the state variable moves in the optimal direction, that is

[0166]

[0167] The convex feasible set algorithm is used to solve the state variable multiple times, and when the difference between the iteration times is less than the threshold value, the optimal state Z is considered to be obtained * , so as to obtain the optimal output trajectory P * . The solving method is mainly through multiple iterations of the above formula to obtain the optimal z, that is, the optimal [P, S], and the solving process is based on the convex feasible set method, which is a relatively common solving method. Due to the addition of the penalty term, the buffer area is regarded as a region that can slightly intrude in extreme cases, which improves the smoothness of the planned trajectory and the comfort of the passengers.

[0168] Based on the above consideration of the automatic driving trajectory planning method considering perception uncertainty, the application provides another device. The device includes but is not limited to one or more processors and memories.

[0169] The memory, as a computer readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions corresponding to the automatic driving trajectory planning method considering perception uncertainty in the embodiment of the application. The processor executes various functional applications and data processing of the vehicle by running the software programs, instructions and modules stored in the memory, that is, realizes the above-mentioned automatic driving trajectory planning method considering perception uncertainty.

[0170] The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; the data storage area can store data created according to the use of the terminal. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage device.

[0171] The application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize an autonomous driving trajectory planning method considering perception uncertainty, and the autonomous driving trajectory planning method considering perception uncertainty comprises the following steps.

[0172] Step one: a vehicle collects environmental information in a driving process by using a vehicle-mounted sensor;

[0173] Step two: the data collected in step one is labeled to form a labeled data set;

[0174] Step three: the data set in step two is divided into a training set, a verification set and a test set;

[0175] Step four: a probabilistic target detection model is constructed based on a multivariate Gaussian distribution;

[0176] Step five: the target detection model is trained by using the data set;

[0177] Step six: the constructed probabilistic target detection model is used for environmental perception, and an obstacle buffer zone is constructed;

[0178] Step seven: a trajectory planning algorithm based on a soft constraint convex feasible set is constructed;

[0179] Step eight: an optimal planning trajectory is generated.

[0180] The computer executable instructions of the computer readable storage medium provided by the application are not limited to the method operations described above, and can also perform the related operations in the autonomous driving trajectory planning method considering perception uncertainty provided by any embodiment of the application.

[0181] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0182] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. Computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (for example, coaxial cable, optical fiber, digital subscriber line DSL) or wireless (for example, infrared, wireless, microwave, etc.) manner. The computer readable storage medium can be any available medium that the computer can access or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (for example, floppy disk, hard disk, magnetic tape), optical media (for example, DVD), or semiconductor media (for example, Solid State Disk, SSD) and the like.

[0183] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any modification, equivalent replacement and improvement made by those skilled in the art within the technical range disclosed by the present application, as long as it is within the spirit and principle of the present application, should be covered within the protection scope of the present application. Meanwhile, the contents not described in detail in the specification are all prior art known to those skilled in the art.

Claims

1. An autonomous driving trajectory planning method considering perception uncertainty, characterized in that, Comprise: construct a probabilistic target detection model for environment perception, and construct an obstacle buffer zone; construct a trajectory planning algorithm based on a soft constraint convex feasible set to generate an optimal planning trajectory; The probabilistic target detection model comprises four parts: an environment encoder, a category decoder, an envelope box decoder, and a variance decoder; When the environment information X is input into the probabilistic target detection model, the environment encoder performs feature extraction to obtain high-order features of the environment information, and the category decoder, the envelope box decoder, and the variance decoder obtain obstacle categories, envelope box attributes, and variance information, respectively; The data flow is represented as: wherein, represents high-level semantic information of environmental information, represents a predicted obstacle category, represents predicted obstacle bounding box information, f enc (·; θ) is an encoding process under parameter θ, D cls (·; θ) is a category decoding process under parameter θ, D reg (·; θ) is a bounding box decoding process under parameter θ, D var (·; θ) is a variance decoding process under parameter θ, and a softplus layer is added at the output end to ensure that the output result is always positive; is the covariance matrix of the model-predicted obstacle; The probabilistic target detection model is constructed based on a multivariate Gaussian distribution; Assuming that the envelope box attribute information of the obstacles in the environment obeys a multivariate Gaussian distribution, the probabilistic target detection model outputs parameters of obstacle distribution, and a mean vector of the multivariate Gaussian distribution is used as the attribute information of the envelope box The variance information constitutes a covariance matrix of the distribution is represented as: Wherein, θ represents network parameters, B represents the true value of obstacle attribute; |·| represents taking the determinant of the matrix, · T represents the transpose operation of the matrix; The label information of the data set is regarded as a multivariate Gaussian distribution with a group of covariance matrices being all-zero matrices, and is regarded as a Dirac distribution, which is represented as: where P G denotes the distribution of labels, and δ(·) denotes the Dirac function. By optimizing the network parameters θ, the difference between the model prediction and the true distribution is minimized, so that the model has the ability of environment perception and variance inference, which is represented as: wherein, denotes the Kullback-Leibler (KL) divergence between two distributions; A KL divergence is used as an index for measuring the difference between two distributions, and the difference is used as an optimization target for learning of the neural network to construct a loss function The regression task and the variance inference task of each variance inference neural network are optimized; the loss function is represented as: wherein, B represents the number of obstacles detected by the model; n B represents the true value of the nth envelope frame, and respectively represent the nth envelope frame attribute information and the covariance matrix predicted by the model, and n represents the serial number index of the obstacle. The probabilistic target detection model is trained using the data set, and the loss function comprises three parts, namely, a category loss, a regression loss, and a direction loss; the Focal Loss is selected as the category loss: wherein, represents a classification loss function, p a is the softmax output of the anchor in the class decoder, and a, g are empirical parameters in the loss function. The loss function of the probabilistic target detection model is represented as: wherein, represents the model total loss function, β cls , β KL and β dir are three loss category balancing coefficients, is a softmax classification loss; After training the probabilistic target detection model for multiple rounds, a probabilistic target detection model with environment perception and variance inference capabilities is constructed, and when the environment information X is input into the model, a probabilistic perception result is obtained: wherein, a set of perception results predicted by the model; The encoder and the decoder of the model gradually have the ability of feature extraction and prediction on point cloud information; The obstacle buffer zone is constructed, and the specific content is as follows: Based on the distribution assumption of the probabilistic target detection model and the output variance information, the obstacle buffer zone is set; The probabilistic target detection result is obtained By It can be known that the standard deviation of each attribute of the nth obstacle envelope frame is: where n represents the ID of the obstacle; σ x , σ y , and σ z are the standard deviations of the obstacle center point on the x, y, and z axes, respectively; σ l , σ w , and σ h are the standard deviations of the obstacle bounding box length, width, and height, respectively; and σ θ is the standard deviation of the obstacle heading angle. All the envelope box attributes of the perceived obstacles should follow the 3σ principle, and the jitter range of each attribute information should be between ±3σ values, that is: x e [x - 3σ x , x + 3σ] x ] l e [l - 3σ l , l + 3σ] l ] y e [y - 3σ y , y + 3σ] y ] w e [w - 3σ w , w + 3σ] w ] θ e [θ - 3σ θ , θ + 3σ] θ ​ z e [z - 3σ z , z + 3σ] z ] h e [h - 3σ h , h + 3σ] h ] The envelope box is selected, which is translated in the coordinate system under the BEV perspective, and the union is selected as the obstacle expansion envelope box. The length and width of the expansion envelope box are: where l and w represent the length and width of the obstacle envelope frame predicted by the perception module, respectively, and l e and w e represent the length and width of the expanded obstacle envelope frame, respectively.

2. The automatic driving trajectory planning method considering perception uncertainty according to claim 1, characterized in that: The covariance matrix between each attribute of the bounding box is a 7x7 diagonal matrix. is a 7x7 diagonal matrix.

3. The automatic driving trajectory planning method considering perception uncertainty according to claim 1, characterized in that: The trajectory planning algorithm based on the soft constraint convex feasible set is constructed, and the specific content is as follows: The planning trajectory point sequence is defined as wherein, coordinates representing a planned trajectory of the vehicle, H denotes a planning step of the preview, and p0 denotes a current position of the vehicle; The optimal trajectory of the vehicle is solved, which is represented as: s.t. Γ(S) = ∩ n Γ n (S) = ∩ n {P: φ n (P) ≥ -S} where J(P) denotes the objective function in the trajectory planning process, ω||S|| is a penalty term for violating soft constraints, P 2 is a penalty term for violating soft constraints, P * , S * denotes the optimal state in the state space, φ n (P) denotes the distance to the nth obstacle, Γ n (·) denotes the constraint condition for the nth obstacle, and Γ(·) denotes the union of all the constraint conditions for all obstacles. The state variable is expanded as The feasible set is represented as: The result is obtained after sorting where Z represents an intermediate variable in the reasoning process, represents the feasible set under the nth obstacle of the algorithm, represents the feasible set in the entire environment, k represents the kth iteration process, and L and S represent coefficients in the derivation process; In each iteration, the algorithm starts from the current solution and finds a new solution in the global convex feasible set, so that the state variable moves in the optimal direction, that is The optimal state Z is obtained by using the convex feasible set algorithm to solve the state variable multiple times, and when the difference between the iteration times is less than a threshold value * , thereby obtaining the optimal output trajectory P * . 4.The method of claim 1, wherein, Before constructing the probabilistic target detection model for environment perception, the following steps need to be performed: Step one: the vehicle uses the vehicle-mounted sensor to collect the environment information during driving; Step two: label the data collected in step one to form a labeled data set; Step three: divide the data set in step two into a training set, a validation set, and a test set; In the step one, the step two and the step three, the ego vehicle collects the environmental information in the process of driving by using the vehicle-mounted sensor, and performs labeling and division, and the specific content is as follows: The intelligent vehicle is provided with various vehicle-mounted sensors, which can perceive the surrounding environment, obtain environmental information, denoted as X, and each environmental information X contains N obstacles, denoted as: wherein, represents a list of obstacles, o n represents the nth obstacle, N represents the number of obstacles; a category information c of each obstacle included in the dataset D n and 3-D bounding box information B n = {x n ,y n ,z n ,w n ,l n ,h n ,θ n}; wherein (x n ,y n ,z n ) represents the position information of the obstacle o n , and (w n ,l n ,h n ,θ n ) represents the attributes and direction of the 3-D bounding box of the obstacle o n ​ In each X, the obstacle label is defined as: The above data set is divided into a training set, a validation set and a test set.

5. An apparatus, characterized by: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-4.

6. A computer readable storage medium having stored thereon a computer program, characterized in that: The program is executed by the processor to implement the method of any one of claims 1-4.

Images