A method for optimizing the deployment of autonomous driving perception sensors

Abstract

This invention provides a method for optimizing the deployment of autonomous driving perception sensors, belonging to the field of autonomous driving technology. Based on a macro- and micro-layered perception information quantification model based on traffic flow statistics, this method enables deployment schemes to possess spatiotemporal coupling optimization characteristics, focusing more on the areas where dynamic occupancy events occur within a scene. Furthermore, by introducing macro- and micro-layered information indicators, information quantification not only focuses on overall information changes caused by micro-grids, but also on information changes in specific macro-regions, making sensor deployment schemes more adaptable to real-world traffic scenarios.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving technology, and specifically relates to a method for optimizing the deployment of autonomous driving perception sensors. Background Art

[0002] Perception sensors, typically including visible light cameras, lidar, and millimeter-wave radar, are the primary hardware components used by autonomous vehicles (AVs) to acquire environmental information and complete key autonomous driving tasks such as object detection, localization, and mapping. Previous research has shown that the deployment of perception sensors impacts the effectiveness of a vehicle's environmental perception. Appropriate sensor placement provides reliable input to the vehicle's decision-making system, enabling it to make safe and accurate judgments and responses in complex and changing traffic scenarios.

[0003] Existing sensor deployment methods for autonomous vehicles focus on optimizing the deployment of a single sensor type (such as omnidirectional lidar or cameras) to achieve improvements in sensing range and point cloud density. However, existing sensor deployment methods suffer from relatively homogeneous sensor types, a single multi-sensor collaborative perception relationship, backward evaluation metrics, and difficulty adapting across vehicle models. These issues make sensor solutions inadequate for complex perception tasks and higher levels of autonomous driving.

[0004] Therefore, it is necessary to provide an optimization method for the deployment of autonomous driving perception sensors to solve the above problems. Summary of the Invention

[0005] The present invention provides a method for optimizing the deployment of autonomous driving perception sensors. Based on a macro-micro layered perception information quantification model of traffic flow statistics, the deployment plan has spatiotemporal coupling optimization characteristics and pays more attention to the areas where dynamic occupancy events occur in the scene. At the same time, due to the introduction of macro-micro layered information indicators, information quantification not only focuses on the total information changes caused by micro grids, but also pays more attention to the information changes in specific macro areas, making the generated sensor deployment plan more adaptable to real traffic scenarios, thereby solving at least one technical problem involved in the background technology.

[0006] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0007] A method for optimizing the deployment of autonomous driving perception sensors, comprising the following steps:

[0008] Step S1: construct a three-dimensional perception space with the vehicle as the origin. At the microscopic level, the perception space is divided into multiple continuous cubic grids; at the macroscopic level, the perception space is divided into four perception areas according to the front, back, left, and right directions of the vehicle.

[0009] Step S2: Using a continuous hemispherical geometric model as a proxy model for the vehicle, and using rays emitted from the sensor center point to represent the sensor's perception process of the perception space, a sensor observation model is established;

[0010] Step S3: Load traffic flow data, count the number of times each target in the traffic flow occupies a cube grid at the micro level and the number of times it occupies a perception area at the macro level over a period of time, calculate the sensor's observation information gain for the micro cube grid and the macro perception area, and use the weighted sum of the observation information gains at the micro and macro levels as the perception entropy indicator;

[0011] In step S4, the position and posture of each sensor are used as variables, and the optimization goal is to maximize the overall observation information gain, construct an optimization function, and solve the optimization function to obtain the optimal value of the position and posture of each sensor.

[0012] As a preferred improvement, the process of constructing the vehicle agent model includes the following steps:

[0013] Step S211: A plane parallel to the ground with the center of the vehicle as the origin is For a plane, a Cartesian coordinate system is established with the vertical upward direction as the z-axis direction;

[0014] Step S212: Using a continuous hemispherical geometric model as a proxy model for the vehicle, aligning the center of the hemispherical geometric model with the center of the vehicle, and determining the minimum hemispherical radius that completely envelops the vehicle. r .

[0015] As a preferred improvement, the construction of the sensor observation model specifically includes the following steps:

[0016] Step S21: Express the coordinates of each point on the hemispherical geometric model in the form of polar coordinates, and construct a connection between the hemispherical points. and the axial point on the z axis Rays ;

[0017] Step S22: Mesh the vehicle body and use the Moller-Trumbore algorithm to detect rays. Intersection point with the vehicle body grid, select a valid intersection point , the intersection is the actual installation position of the sensor, and the sensor orientation angle is determined by the ray direction is determined.

[0018] As a preferred improvement, the sensor includes a camera and a laser radar. For the camera, the number of rays is the product of the horizontal and vertical resolutions of the camera, which is calculated by the number of horizontal and vertical pixels. u、vCalculate the horizontal and vertical angles of each ray. The starting point of the ray is the optical center, which is assumed to be the installation coordinates of the camera. For the lidar, the starting point of the ray is the intersection of each ray, which is assumed to be the installation coordinates of the lidar. The number of rays is the number of channels, combined with the horizontal and vertical field of view angles , horizontal and vertical angle resolution 、 Calculate the horizontal and vertical heading angles of each ray.

[0019] As a preferred improvement, the horizontal and vertical angles of any laser radar ray Defined as:

[0020]

[0021]

[0022] in, 、 are the horizontal and vertical orientation angles of the laser radar respectively; 、 Respectively represent the horizontal and vertical field of view angles of the laser radar; 、 Respectively represent the horizontal and vertical angular resolutions of the lidar;

[0023] The horizontal and vertical angles of any ray of the camera Defined as:

[0024]

[0025]

[0026] in, 、 are the horizontal and vertical angles of the camera respectively, , are the horizontal and vertical field of view of the camera respectively; u、v Respectively represent the number of horizontal and vertical pixels of the camera;

[0027] The cube grid of the perception space through which the ray passes is calculated by the Bresenham line algorithm. The end point of the ray is determined by the boundary surface of the perception space. When observing the perception space, the grid that is traversed by the ray is the observed grid, and the set of observed grids is Expressed as:

[0028]

[0029] Where, Indicates Bresenham's line algorithm calculation.

[0030] As a preferred improvement, the perceptual entropy index is expressed as:

[0031]

[0032] Where, 、 Represent the observation information gain at the micro and macro levels respectively; Represents the weight coefficient, which is used to balance the importance of macro and micro information, where:

[0033] ;

[0034] ;

[0035] Where, 、 represents the conditional entropy under sensor s observation; 、 They represent the amount of information quantified by sensor s observing the occupancy status of a single grid and area.

[0036] As a preferred improvement, the optimization function is expressed as:

[0037] ;

[0038] Where, represents the optimization objective; Indicates the position and attitude parameters that each sensor needs to optimize, Design variables of the sensor composition, , 、 、 Respectively represent the horizontal orientation angle, vertical orientation angle and height of the sensor; represents the overall observation information gain, which is expressed as:

[0039] ;

[0040] Where, and Respectively represent Grid and regional information gain brought by sensors; Indicates the number of sensors.

[0041] As a preferred improvement, the optimization function satisfies the following constraints:

[0042] When multiple sensors of the same type need to consider their collaborative perception, the field of view overlap cannot be too large. The constraint can be expressed as:

[0043]

[0044] Where, Indicates sensor Horizontal orientation angle; Indicates sensor The horizontal orientation angle, Indicates the limit value of the horizontal orientation angle of the homogeneous sensor;

[0045] When different types of sensors need to consider their collaborative perception, the field of view overlaps to ensure that heterogeneous information about the same target object is complementary. The constraint is expressed as:

[0046]

[0047] Where, Indicates the limit value of the horizontal orientation angle of the heterogeneous sensor.

[0048] The beneficial effects of the present invention are:

[0049] (1) Improved multi-sensor collaborative perception performance. By using overlapping or mutually exclusive field of view constraints of multiple heterogeneous sensors (multiple cameras + multiple lidars), complementary perception ranges and complete target information can be achieved.

[0050] (2) Improved adaptability of sensing deployment in dynamic scenarios and specific areas. The macro-micro hierarchical sensing information quantification model based on traffic flow statistics enables the deployment plan to have spatiotemporal coupling optimization characteristics, paying more attention to the dynamic occupation event area in the scene. At the same time, due to the introduction of macro-micro hierarchical information indicators, information quantification not only focuses on the total information changes caused by micro grids, but also pays more attention to the information changes in specific macro areas. The sensor deployment plan is more adaptable to real traffic scenarios.

[0051] (3) Improved design efficiency and vehicle model adaptability. Based on the hemispherical proxy model, the sensor parameter design variables are continuous, which reduces the optimization time. Since the hemispherical proxy model only considers the size of the vehicle body bounding box, autonomous driving vehicles with different shapes are also applicable to this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0053] Figure 1 Schematic diagram showing the perceptual space model at the micro level;

[0054] Figure 2 Schematic diagram of the perceptual space model at the macro level;

[0055] Figure 3 Represents a schematic diagram of a sensor model;

[0056] Figure 4 A diagram showing the occupancy of the target bounding box on the grid in the corresponding frame at a certain moment;

[0057] Figure 5 A schematic diagram showing the grid observed by the sensor rays;

[0058] Figure 6 Schematic diagram showing the body spherical proxy model and the body discrete point model. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] Please refer to Figures 1-6 This embodiment provides a method for optimizing the deployment of autonomous driving perception sensors, including the following steps:

[0061] In step S1, a three-dimensional perception space is constructed with the vehicle as the origin. At the microscopic level, the perception space is divided into multiple continuous cubic grids. At the macroscopic level, the perception space is divided into four perception areas according to the front, back, left, and right directions of the vehicle.

[0062] like Figure 1 As shown in Figure 2, at the microscopic level, the length, width, and height of the perceived space are L, W, and H respectively. The size of the cube grid is × × , then the number of cube grids N is calculated as follows:

[0063] ;

[0064] A collection of volume meshes Expressed as:

[0065]

[0066] Where, Represents an arbitrary cube grid, and the spatial coordinates of its center point are expressed as .

[0067] like Figure 2 As shown, at the macro level, the set of four perception areas is represented as .

[0068] In step S2, a continuous hemispherical geometric model is used as a proxy model of the vehicle, and rays emitted from the center point of the sensor are used to represent the sensor's perception process of the perception space, thereby establishing a sensor observation model.

[0069] The process of building a vehicle agent model includes the following steps:

[0070] Step S211: A plane with the center of the vehicle as the origin and parallel to the ground is For a plane, a Cartesian coordinate system is established with the vertical upward direction as the z-axis direction;

[0071] Step S212: Using a continuous hemispherical geometric model as a proxy model for the vehicle, aligning the center of the hemispherical geometric model with the center of the vehicle, and determining the minimum hemispherical radius that completely envelops the vehicle. r .

[0072] The hemispherical geometric model is used as a proxy model for the vehicle, which can avoid the massive discrete variables that need to be processed by the traditional vehicle body discrete position model, thereby reducing the computational complexity.

[0073] The construction of the sensor observation model specifically includes the following steps:

[0074] Step S21: Express the coordinates of each point on the hemispherical geometric model in the form of polar coordinates, and construct a connection between the hemispherical points. and the axial point on the z axis Rays .

[0075] Step S22: Mesh the vehicle body and use the Moller-Trumbore algorithm to detect rays. Intersection point with the vehicle body grid, select a valid intersection point , the intersection is the actual installation position of the sensor, and the sensor orientation angle is determined by the ray direction is determined.

[0076] The vehicle body is typically a complex spatial surface, difficult to express using a unified mathematical expression. Therefore, a set of discrete points can be sampled on the vehicle body to represent the vehicle body. These discrete points can become potential sensor installation locations. However, this discrete installation location set can result in the loss of the optimal installation location. For example, the optimal location may be between points A and B, but this point does not exist in the set. Therefore, this application meshes the vehicle and obtains discrete points on the vehicle body through dense sampling. These discrete points are then connected in a triangular form to form a continuous triangular mesh approximation of the vehicle body, which can avoid the loss of the optimal installation location.

[0077] The gridded vehicle body is used as the design domain for the sensor installation location, and the sensor installation coordinates The upper and lower bounds can be obtained from the mesh body or the discrete point set The maximum and minimum values ​​of the components are determined, and the actual design domain of the installation position coordinates can be obtained, where any The combination will be searched by the optimization algorithm to find the best position, but there is a problem here: any The combination may indicate an invalid installation position coordinate, because the complex body surface does not necessarily have such coordinates. Therefore, it is necessary to determine whether the body has such coordinates. However, due to the lack of analytical expression for the body, the existence of such coordinates is difficult to determine. Based on this, the application determines whether there are feasible installation position coordinates by finding the intersection point between the ray and the body grid. Therefore, it is necessary to construct a ray with the starting point at axis Within the range, this starting point is also the axial point ,in , then at a given ray angle 、 , , , we can determine the analytical expression of the ray. Since the car body is composed of a triangular mesh, the Moller-Trumbore algorithm can quickly determine and solve the intersection of the ray and the triangular mesh, thereby obtaining the actual installation position of the sensor. At the same time, when the intersection exists, the direction angle of the ray 、 Therefore, the installation position of the sensor is determined by the orientation angle and axial point of the ray. Then these three variables will replace the installation position of the sensor and become the new design variables, that is, the orientation angle and the starting point of the ray become a proxy model, in which the orientation angle is set. 、 It satisfies the spherical constraint on a hemisphere (i.e. a point represented by polar coordinates on the sphere), and the orientation angle satisfies the spherical constraint. In addition, the design variable is changed from the original three-dimensional rectangular coordinate to the two-dimensional angle of the proxy sphere, which reduces the difficulty of the optimization algorithm.

[0078] The sensor model constructed is as follows Figure 3 As shown, the sensor Emitted rays The horizontal and vertical directions are 、 .

[0079] Sensors include cameras and lidars. For cameras, the number of rays is the product of the horizontal and vertical resolutions of the camera, which is calculated by the number of horizontal and vertical pixels. u、v Calculate the horizontal and vertical angles of each ray. The starting point of the ray is the optical center, which is assumed to be the installation coordinates of the camera. For the lidar, the starting point of the ray is the intersection of each ray, which is assumed to be the installation coordinates of the lidar. The number of rays is the number of channels, combined with the horizontal and vertical field of view angles , horizontal and vertical angle resolution 、 Calculate the horizontal and vertical heading angles of each ray.

[0080] The horizontal and vertical angles of any laser radar ray Defined as:

[0081]

[0082]

[0083] in, 、 are the horizontal and vertical orientation angles of the laser radar respectively; 、 Respectively represent the horizontal and vertical field of view angles of the laser radar; 、 Respectively represent the horizontal and vertical angular resolutions of the lidar;

[0084] The horizontal and vertical angles of any ray of the camera Defined as:

[0085]

[0086]

[0087] in, 、 are the horizontal and vertical angles of the camera respectively, , are the horizontal and vertical field of view of the camera respectively; u、v Respectively represent the number of horizontal and vertical pixels of the camera.

[0088] The cube grid of the perception space through which the ray passes is calculated by the Bresenham line algorithm. The end point of the ray is determined by the boundary surface of the perception space. When observing the perception space, the grid that is traversed by the ray is the observed grid, and the set of observed grids is Expressed as:

[0089] ;

[0090] Where, Indicates Bresenham's line algorithm calculation.

[0091] Step S3: load traffic flow data, count the number of times each target in the traffic flow occupies the cube grid at the micro level and the number of times it occupies the perception area at the macro level over a period of time, calculate the sensor's observation information gain for the micro cube grid and the macro perception area, and use the weighted sum of the observation information gains at the micro and macro levels as the perception entropy indicator.

[0092] T-frame traffic flow data is collected in simulation or real scenes, and each frame of data contains a set of 3D bounding boxes of all dynamic targets in the perception space.

[0093] At the micro level, for each cube mesh, we traverse all frames and count the number of times it is covered by the target bounding box. Figure 4 The shaded area represents the grid occupied by the target bounding box. Specifically, the overlapping volume of the target bounding box and the grid is calculated as As the occupancy weight, if the occupancy weight is greater than the threshold , then the grid index position is marked as , indicating an occupied state, if it is less than the threshold, it is marked as 0, indicating that it is not occupied. After the frame, these marks can be accumulated to get the occupancy value . Statistics all Cube grid in frame The total number of times occupied, that is, the occupancy value of each cube grid is obtained , then for the cube grid , its occupation probability can be calculated Expressed as:

[0094] ;

[0095] Probability Reflects the target object in the cube grid The likelihood of occurrence.

[0096] Similarly, at the macro level, statistics of each area in a frame The number of grid observations , then the occupied status of each area of ​​the frame is , is the total number of grids in the area. Then after accumulating T frames, the occupancy value can also be calculated , we can get the occupancy probability of the area :

[0097] .

[0098] To simplify the calculation, the present invention assumes that the occupancy events of each cube and each area are independent of each other. Therefore, under the independence assumption, the joint probability density of all cube grids in the perception space being occupied is It can be decomposed into the product of the occupancy probabilities of all cube grids:

[0099] .

[0100] Similarly, the independence assumption is applied to the perception area , the joint probability density of all four regions in the perception space being occupied It can be decomposed into the product of the occupancy probabilities of all cube grids:

[0101] .

[0102] For the observed grid set, under the independence assumption, the product of the probability of the occupied cubic grid observed by sensor s can also be calculated :

[0103] ;

[0104] Where, It is expressed as the probability that the grid is occupied under the observation of sensor s.

[0105] The probability of four regions being observed simultaneously Expressed as:

[0106] ;

[0107] Where, Indicates the The probability that a region is observed.

[0108] Therefore, the above modeling steps can be summarized as counting the number of times the grid area inside the perception space is occupied by the target and the position index within a period of time, counting the position index of the grids passed by the sensor ray, and the intersection of the target occupied grid and the sensor passed grid (obtained by the position index) is the sensor's observation result of the occupied grid and area. This observation result can be converted into the amount of information observed by the sensor on the grid and area, namely, information entropy, as shown in Figure 5As shown, sensor S1 emits ray L1, sensor S2 emits ray L2, and the observed grid is shown as g1. According to information theory, the grid information entropy in the perception space is The probability density The regional information entropy of the above four perception areas can be obtained by Find them and express them as:

[0109] ;

[0110] ;

[0111] Among them, the information amount of a single grid / area occupancy state (including occupied and unoccupied) and It can be calculated by the following formula:

[0112] ;

[0113] .

[0114] The conditional entropy under sensor s observation is 、 The probability density and Solve, then the grid and regional observation information gain brought by sensor s 、 The expressions are:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] Where, 、 They represent the amount of information quantified by sensor s observing the occupancy status of a single grid and area.

[0120] Observation Information Gain 、 It is expressed as the basic quantitative index of sensor observation information. Therefore, for sensor observation, it is necessary to consider the information changes of the detail area in the perception space, that is, the micro-grid, and also to pay attention to the information changes of the macro-region in the perception space. For this purpose, the hierarchical perception entropy index is obtained:

[0121] ;

[0122] in, is the weight coefficient, which is used to balance the importance of macro and micro information.

[0123] In step S4, the position and posture of each sensor are used as variables, and the optimization goal is to maximize the overall observation information gain, construct an optimization function, and solve the optimization function to obtain the optimal value of the position and posture of each sensor.

[0124] According to the information quantification model, sensors observe occupancy events by traversing the grid with rays. The relevant grid intersection parameters are quantified as information gain, which can be set as the basic objective function in the deployment optimization model. For multiple lidars or cameras, the gain effect of the overall autonomous vehicle sensor solution can be obtained by accumulating the information gain of each sensor, that is, the overall observation information gain. , which can be expressed as:

[0125] ;

[0126] Where, and Respectively represent Grid and regional information gain brought by sensors; Indicates the number of sensors.

[0127] Therefore, the overall observation information gain is set as the objective function of the deployment optimization model, that is, it is necessary to solve the sensor position and posture brought by the maximum observation information gain.

[0128] Polar coordinates are used to define the sensor position and orientation that need to be optimized. The horizontal orientation angle range is , the vertical orientation angle range is , height parameter .

[0129] Therefore, for the Design variables of the sensor , which is expressed as follows:

[0130] ;

[0131] When considering collaborative perception between multiple sensors of the same type, the overlap of their fields of view cannot be too large. Therefore, a collaborative constraint 1 can be established. For a solution involving two sensors of the same type, the constraint can be expressed as:

[0132]

[0133] Where, Indicates sensor Horizontal orientation angle; Indicates sensor The horizontal orientation angle, Indicates the limit value of the horizontal orientation angle of the homogeneous sensor.

[0134] When different types of sensors need to be considered for collaborative perception, overlapping fields of view are necessary to ensure complementary heterogeneous information about the same target object. For example, the information observed by a lidar and a camera on the same target is completely heterogeneous. Therefore, collaborative constraint 2 can be established. For a solution involving two different types of sensors, the constraint can be expressed as:

[0135] ;

[0136] Where, Indicates the limit value of the horizontal orientation angle of the heterogeneous sensor.

[0137] Thus, the main components of the optimization deployment model can be obtained, and the optimization problem of the model can be expressed as:

[0138] ;

[0139] Where, represents the optimization objective; Indicates the position and attitude parameters that need to be optimized for each sensor, composition.

[0140] The optimization function can be solved using intelligent swarm optimization tools such as genetic algorithms to obtain the best deployment solution.

[0141] The above describes the embodiments of the present invention, but the present invention is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A method for optimizing the deployment of autonomous driving perception sensors, characterized in that: The steps include: Step S1: construct a three-dimensional perception space with the vehicle as the origin. At the microscopic level, the perception space is divided into multiple continuous cubic grids; at the macroscopic level, the perception space is divided into four perception areas according to the front, back, left, and right directions of the vehicle. Step S2: Using a continuous hemispherical geometric model as a proxy model for the vehicle, and using rays emitted from the sensor center point to represent the sensor's perception process of the perception space, a sensor observation model is established; Step S3: Load traffic flow data, count the number of times each target in the traffic flow occupies a cube grid at the micro level and the number of times it occupies a perception area at the macro level over a period of time, calculate the sensor's observation information gain for the micro cube grid and the macro perception area, and use the weighted sum of the observation information gains at the micro and macro levels as the perception entropy indicator; Step S4, taking the position and posture of each sensor as a variable and maximizing the overall observation information gain as the optimization goal, constructing an optimization function, and solving the optimization function to obtain the optimal value of the position and posture of each sensor; The process of building a vehicle agent model includes the following steps: Step S211: A plane with the center of the vehicle as the origin and parallel to the ground is For a plane, a Cartesian coordinate system is established with the vertical upward direction as the z-axis direction; Step S212: Using a continuous hemispherical geometric model as a proxy model for the vehicle, aligning the center of the hemispherical geometric model with the center of the vehicle, and determining the minimum hemispherical radius that completely envelops the vehicle. r ; The construction of the sensor observation model specifically includes the following steps: Step S21: Express the coordinates of each point on the hemispherical geometric model in the form of polar coordinates, and construct a connection between the hemispherical points. and the axial point on the z axis Rays ; Step S22: Mesh the vehicle body and use the Moller-Trumbore algorithm to detect rays. Intersection point with the vehicle body grid, select a valid intersection point , the intersection is the actual installation position of the sensor, and the sensor orientation angle is determined by the ray direction is determined.

2. The method for optimizing the deployment of autonomous driving perception sensors according to claim 1, characterized in that: Sensors include cameras and lidars. For cameras, the number of rays is the product of the horizontal and vertical resolutions of the camera, which is calculated by the number of horizontal and vertical pixels. u、 v Calculate the horizontal and vertical angles of each ray. The starting point of the ray is the optical center, which is assumed to be the installation coordinates of the camera. For the lidar, the starting point of the ray is the intersection of each ray, which is assumed to be the installation coordinates of the lidar. The number of rays is the number of channels, combined with the horizontal and vertical field of view angles , horizontal and vertical angle resolution 、 Calculate the horizontal and vertical heading angles of each ray.

3. The method for optimizing the deployment of autonomous driving perception sensors according to claim 2, characterized in that: The horizontal and vertical angles of any laser radar ray Defined as: in, 、 are the horizontal and vertical orientation angles of the laser radar respectively; 、 Respectively represent the horizontal and vertical field of view angles of the laser radar; 、 Respectively represent the horizontal and vertical angular resolutions of the lidar; The horizontal and vertical angles of any ray of the camera Defined as: in, 、 are the horizontal and vertical angles of the camera respectively, , are the horizontal and vertical field of view of the camera respectively; u、v Respectively represent the number of horizontal and vertical pixels of the camera; The cube grid of the perception space through which the ray passes is calculated by the Bresenham line algorithm. The end point of the ray is determined by the boundary surface of the perception space. When observing the perception space, the grid that is traversed by the ray is the observed grid, and the set of observed grids is Expressed as: Where, Indicates Bresenham's line algorithm calculation.

4. The method for optimizing the deployment of autonomous driving perception sensors according to claim 1, wherein: The perceptual entropy index is expressed as: Where, 、 Represent the observation information gain at the micro and macro levels respectively; Represents the weight coefficient, which is used to balance the importance of macro and micro information, where: ； ； Where, 、 represents the conditional entropy under sensor s observation; 、 They represent the amount of information quantified by sensor s observing the occupancy status of a single grid and area.

5. The method for optimizing the deployment of autonomous driving perception sensors according to claim 4, characterized in that: The optimization function is expressed as: ； Where, represents the optimization objective; Indicates the position and attitude parameters that each sensor needs to optimize, Design variables of the sensor composition, , 、 、 Respectively represent the horizontal orientation angle, vertical orientation angle and height of the sensor; represents the overall observation information gain, which is expressed as: ； Where, and Respectively represent Grid and regional information gain brought by sensors; Indicates the number of sensors.

6. The method for optimizing the deployment of autonomous driving perception sensors according to claim 5, characterized in that: The optimization function satisfies the following constraints: When multiple sensors of the same type need to consider collaborative perception, the field of view overlap cannot be too large. The constraint can be expressed as: Where, Indicates sensor Horizontal orientation angle; Indicates sensor The horizontal orientation angle, Indicates the limit value of the horizontal orientation angle of the homogeneous sensor; When different types of sensors need to consider their collaborative perception, the field of view overlaps to ensure that heterogeneous information about the same target object is complementary. The constraint is expressed as: Where, Indicates the limit value of the horizontal orientation angle of the heterogeneous sensor.

Images