A method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions

Abstract

The present invention discloses a method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions, which belongs to the technical field of optimized deployment of roadside perception units. The method includes establishing traffic scenarios, obtaining true probability occupancy grids of various traffic scenarios through multi-time-step continuous simulation; performing parameterized perception modeling of roadside perception units and intelligent connected vehicles, establishing a ray tracing algorithm that considers occlusion effects, obtaining observable probability occupancy grids under vehicle-road cooperative perception conditions through the ray tracing algorithm that considers occlusion effects, and proposing to use the cross entropy of the true probability occupancy grid and the observable probability occupancy grid as an alternative evaluation indicator of vehicle-road cooperative perception for traffic flow monitoring, so as to optimize the deployment of roadside perception units; introducing POG into the field of RPU optimized layout provides a specific and complete technical framework, which solves the problem that the existing technology does not adequately consider the impact of ICV when carrying out RPU optimized layout modeling.

Description

Technical Field

[0001] The present invention relates to the technical field of optimized deployment of roadside perception units, and specifically to a method for optimized deployment of roadside perception units under partial vehicle-road cooperative conditions. Background Art

[0002] Intelligent connected vehicles (ICVs) are an important part of future road traffic flows. Compared with traditional human-driven cars, ICVs can actively perceive traffic scenes and share information with other ICVs or intelligent roadside facilities through wireless communication technology. At present, related research on traffic risk identification, vehicle-level control, etc. in the field of intelligent networking or vehicle-road collaboration is based on the fact that the fine-grained motion data of traffic individuals in the entire road domain can be accurately obtained through vehicle-road collaboration technology. However, it will take a long time to achieve the acquisition of full samples of traffic flow trajectories, during which time the roadside perception system will play an important role. Specifically, the existing road infrastructure needs to deploy roadside perception units (RPUs) to enhance roadside perception capabilities, thereby filling the blind spots of ICV vehicle-side perception and effectively improving the accuracy of traffic scene perception.

[0003] Thanks to wireless communication technology, existing research has found that ICVs and RPUs can improve the accuracy of traffic data collection by complementing each other's field of view (FOV). However, in some stages of vehicle-road collaboration development where ICV penetration is low, the number of RPUs and ICVs in traffic scenarios is limited. In this context, the dynamic occlusion problem of moving vehicles will still significantly affect the perception efficiency of vehicle-road collaboration. Specifically, moving vehicles as obstacles will create blind spots in the perception of ICVs or RPUs, causing the perception system to miss traffic individuals, thereby affecting the normal operation of the vehicle-road collaboration system (such as risk monitoring).

[0004] The existing technologies for optimizing the deployment of RPUs at the microscopic spatial scale can be divided into two categories: those oriented towards static scenarios and those based on direct simulation. Technical solutions for static scenarios account for the majority of existing technologies. Their main problem is that they only consider the perception blind spots caused by static obstacles (such as building facades) on the RPU, without considering any individual traffic entities. However, mixed traffic scenarios are mostly on straight sections of roads, where the impact of roadside obstacles is relatively small. The dynamic occlusion of moving vehicles is the main reason for the decline in vehicle-road collaborative perception performance. Therefore, static scenario methods cannot effectively cope with the optimized deployment of road sensing units under some vehicle-road collaborative conditions.

[0005] Based on the direct simulation technology solution, a traffic scene is built on the existing micro-simulation platform, and the sensor model built into the software is called to quantitatively evaluate the perception efficiency. On this basis, the RPU is optimized. In the study of the optimal layout method of intelligent highway roadside perception equipment, Qiu Xingyou et al. established a multi-lane highway traffic scene through SUMO and called the traffic detector of OMNET++ to simulate the perception of traffic events. However, this study ignored the influence of obstacles and set that traffic individuals can be perceived if they fall within the specific radius of the RPU. In order to model the dynamic occlusion between vehicles, Jin et al. (JIN S, GAO Y, HUI F, et al. A Novel Information Theory-Based Metric for Evaluating Roadside LiDAR Placement [J]. IEEE Sensors Journal, 2022, 22 (21): 21009-21023) used Unity3D to create a three-dimensional traffic scene and simulate the RPU perception function. Considering the influence of dynamic occlusion, a roadside perception efficiency evaluation index based on information entropy was constructed through continuous simulation and its effectiveness was verified. However, this study only considered a single RPU and did not address RPU optimization deployment for improved perception performance. Ma et al. (MA Y, ZHENG Y, WANG S, et al. A Virtual Method for Optimizing Deployment of Roadside Monitoring Lidars at As-Built Intersections [J]. IEEE Transactions on Intelligent Transportation Systems, 2023, 24(11): 11835–11849.) combined data from the same traffic scene acquired by multiple RPUs and quantitatively evaluated perception performance by comparing the differences between individual traffic models under collaborative perception conditions and their original models. However, this study relied on actual collected traffic trajectory data and could not flexibly respond to diverse traffic scenarios.

[0006] However, high-precision sensor simulation models involve a lot of graphics calculations, so when using Carla, Unity3D or other software for perception simulation, the calculation efficiency is low. When optimizing the RPU layout, it may involve multiple times (such as 10 3 To ensure robustness, a single optimization iteration needs to evaluate the average perception performance in multiple time-step scenarios. Therefore, the optimization deployment of RPU based on direct simulation method is very time-consuming.

[0007] In addition to the above problems, although ICV can make up for the road-side perception blind spots, existing research has not adequately considered the impact of ICV when conducting RPU optimization layout modeling, resulting in the inability of existing technical solutions to cope with the optimized layout method of RPU under some vehicle-road cooperative conditions. Summary of the Invention

[0008] In order to make up for the shortcomings of the existing technical problems, the purpose of the present invention is to provide a method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions, introduce POG into the field of RPU optimization layout, provide a specific and complete technical framework, use POG to effectively measure the uncertainty of the scene, and solve the problem that the existing technology does not consider the impact of ICV when carrying out RPU optimization layout modeling; in addition, by evaluating the vehicle-road cooperative perception performance based on POG cross entropy, continuous simulation can be avoided, effectively reducing the time cost of RPU optimization layout.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions, comprising:

[0011] (1) Establish traffic scenarios based on the Vissim-Carla joint simulation method, and obtain the true value probability occupancy grid of various traffic scenarios through multi-time step continuous simulation;

[0012] (2) Conduct parameterized modeling of the perception of roadside perception units and intelligent connected vehicles. To address the adverse effects of mobile occlusion on vehicle-road collaborative perception, establish a ray tracing algorithm that takes occlusion effects into account to simulate the collaborative perception process of roadside perception units and intelligent connected vehicles.

[0013] (3) A ray tracing algorithm that takes occlusion effects into account is used to obtain an observable probabilistic occupancy grid under vehicle-infrastructure cooperative perception conditions. The cross entropy between the true probabilistic occupancy grid and the observable probabilistic occupancy grid is proposed as an alternative evaluation metric for vehicle-infrastructure cooperative perception for traffic flow monitoring.

[0014] (4) The layout parameters of the roadside perception unit and the cross entropy based on the probability occupancy grid are used as decision variables and objective function values ​​respectively, and the Bayesian optimization method is used to complete the optimal deployment of the roadside perception unit.

[0015] In the present invention, the step of obtaining the true value probability occupancy grid in step (1) includes:

[0016] 1) Establish a global xyz coordinate system and determine the region of interest for mixed-use scene perception by constraining the xyz coordinate range;

[0017] 2) At any simulation time step, create the envelope of the traffic individual, voxelize each envelope, and obtain the true value probability occupancy grid through continuous updating of the simulation step.

[0018] In the present invention, voxelization adopts a step-by-step process from local space to global coordinate frame, as follows:

[0019] Step 1: Grid point coordinate space conversion

[0020] Set x v -y v -z v is the local coordinate system of each traffic entity, where y v Corresponding to the real-time moving direction of the individual, x v Perpendicular to the individual's forward direction, horizontally to the right; in the local coordinate frame of the traffic individual, the envelope is outlined by the length, width and height of the vehicle; within the envelope, δ vx Generate regular grid point data for the interval; set t i The azimuth of traffic individual j at time Then, the coordinate frame x is transformed into v -y v -z v The grid point data in is converted to the global coordinate system xyz;

[0021]

[0022] in:

[0023]

[0024] in: It is t i The rigid coordinate transformation matrix of traffic individual j at the moment, and are the local coordinates and global coordinates of individual j grid point, It is t i The global coordinate value of the bottom center of traffic individual j at the moment;

[0025] Step 2: Voxelization

[0026] The grid points after coordinate transformation are voxelized, and the voxel size is consistent with the spacing of the grid points; let V s V is the road traffic scene after voxel processing within the ROI. s The data format is a three-dimensional matrix, each matrix element corresponds to a voxel unit and stores the corresponding occupancy probability value; let M v t i The voxel representing the traffic individual in the scene at the moment Vs The position coordinate set within V s In M v The voxel value is set to 1, which is mathematically expressed as V s |t i ,M v =1; The above is only for the case of a single time step. In order to obtain the POG of a certain traffic scene, continuous simulation and calculation are required; with V s Medium voxel unit v i For example, the corresponding occupancy probability According to equations (3) to (4), we can obtain

[0027]

[0028] in:

[0029]

[0030] Where: T is the total number of simulation steps, v i ∈M v and Respectively refer to whether the voxel vi is included in the coordinate set of the traffic grid point in Vs at a certain time step, is the occupancy probability of the voxel based on the sample evaluation;

[0031] Assuming that λ is the total number of voxel units that are occupied at least once during the T-step simulation process, then 1≤i≤λ; to simplify the calculation process, it is set under the condition of sufficient simulation steps Each voxel unit is regarded as a 0-1 binary variable, and there is Considering that the occupancy state of a single voxel cannot infer the state of other voxels, it is believed that in V s The voxels within the ROI are mutually independently distributed random variables; under this condition, the joint distribution of multivariate variables within the ROI range is calculated according to formula (5);

[0032]

[0033] in

[0034] In the present invention, the perception parameterized modeling of the roadside perception unit and the intelligent connected vehicle includes:

[0035] Let (x R ,y R ,z R ) T is the sensor center of the roadside perception unit, with (x R ,y R ,z R ) TAs the origin, establish the sensor local coordinate system x of the roadside perception unit s -y s -z s ,make With φ s are the vertical and horizontal field of view angles of the roadside perception unit, d s is the effective sensing range of the roadside sensing unit;

[0036] Taking the center of the sensor on the intelligent connected car as the origin, y vs Corresponding to the real-time moving direction of the individual, x vs Perpendicular to the individual's forward direction and horizontally to the right, establish the local coordinate system x of the onboard sensor of the intelligent connected vehicle vs -y vs -z vs , set the horizontal field of view of the ICV sensor to 360 degrees; let ξ vs is the vertical field of view angle of the ICV sensor, d vs is the effective sensing distance of ICV.

[0037] In the present invention, the observable probability occupancy grid under the vehicle-road cooperative perception condition is obtained as follows:

[0038] Under the cooperative perception conditions of the roadside perception unit and the intelligent connected vehicle, the voxel v i Occupancy probability of the condition observed by all sensing units:

[0039]

[0040] in: and are the occupancy probabilities of the non-ICV part of voxel vi under the observation conditions of the mth RPU and the nth ICV sensor, is the occupancy probability of the ICV portion corresponding to voxel vi;

[0041] Assuming that the probability distribution of each voxel unit is still independent under the vehicle-road cooperative observation condition, the joint distribution of conditional occupancy probability is

[0042]

[0043] in: μ is the number of voxels with non-zero conditional occupancy probability within the ROI under the vehicle-road cooperative perception condition. In the present invention, based on the independence assumption, the cross entropy can be calculated voxel by voxel, as shown in formulas (15)-(16); The smaller it is, the closer the occupancy probability under the vehicle-road cooperative perception condition is to the actual situation;

[0044]

[0045] Where: n c It is the number of voxels occupied by non-motor vehicles among the non-zero voxels. If a voxel has been occupied by both motor vehicles and non-motor vehicles, it is considered to be a non-motor vehicle occupied voxel.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. The present invention establishes a systematic method of "simulation-based POG construction - collaborative perception modeling - perception effectiveness substitution evaluation - RPU optimization layout", which provides a specific and complete technical framework for introducing POG into the field of RPU optimization layout. It uses POG to effectively measure the uncertainty of the scenario and solves the problem that the existing technology does not fully consider the impact of ICV when conducting RPU optimization layout modeling.

[0048] 2. The present invention is aimed at traffic scenarios represented by POGs. The occupancy probability of voxel units in POGs is equivalent to the degree of opacity. Combined with the Bayesian reasoning method, a ray tracing technology that considers the occupancy effect is established. This technology can obtain observable POGs for vehicle-road collaborative perception under different intelligent connected vehicle penetration rates and different RPU layout conditions. While considering the uncertainty of traffic scenarios, it can quickly and quantitatively evaluate the collaborative perception effectiveness of intelligent vehicles and roadside perception units for traffic flow scenarios.

[0049] 3. This invention incorporates ICVs into RPU optimization and takes into account the occlusion effect of moving vehicles on RPU or ICV perception. Evaluating vehicle-road cooperative perception performance through POG-based cross-entropy avoids continuous simulations, effectively reducing the time cost of RPU optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Flow chart of the method of the present invention.

[0051] Figure 2 Schematic diagram of the basic traffic scene of the present invention.

[0052] Figure 3 Obtaining the true value probability occupancy grid

[0053] Figure 4 Schematic diagram of the perception parameterization modeling of the roadside perception unit and the intelligent connected vehicle.

[0054] Figure 5 This is a flow chart of the ray tracing algorithm that takes occlusion effects into account in the present invention.

[0055] Figure 6 This is a schematic diagram of the virtual light creation and voxelization processing principles in the present invention.

[0056] Figure 7Schematic diagram of the change of voxel occupancy probability along the virtual ray.

[0057] Figure 8 Schematic diagram of voxel overlap in the present invention.

[0058] Figure 9 Schematic diagram of the collaborative perception process based on POG in the present invention.

[0059] Figure 10 Schematic diagram of the change of voxel occupancy probability along the virtual ray. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] Combine Figure 1-10 , the method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions of the present invention is specifically described.

[0062] 1. Technical Overview

[0063] The technical framework for optimizing the deployment of RPU under partial vehicle-road cooperative conditions proposed in this invention mainly includes the following basic steps: Figure 1 As shown. First, a traffic scene is established based on the Vissim-Carla joint simulation method, and then the probabilistic occupancy grid (POG, a spatial representation method for dynamic environments, that is, dividing the scene into countless voxel units, and each voxel unit stores the occupancy probability of the corresponding grid space area) of various traffic scenes is obtained through multi-time step continuous simulation. Then the perception modeling of RPU and ICV is carried out, and in view of the adverse effects of mobile occlusion effect on vehicle-road collaborative perception, a ray tracing algorithm (occlusion-considered ray-tracing, OCRT) that considers the occlusion effect is established to simulate the collaborative perception process of roadside perception unit (RPU) and intelligent connected vehicle (ICV). The POG observed under the vehicle-road collaborative perception conditions is obtained by the OCRT algorithm, and the cross entropy of the true value POG and the observable POG is proposed as an alternative evaluation indicator of vehicle-road collaborative perception for traffic flow monitoring. Finally, the layout parameters of RPU and the cross entropy based on POG are used as decision variables and objective function values ​​respectively, and the Bayesian optimization method is used to complete the optimal deployment of the road sensing unit. The following will describe in detail each step in the technical solution of the present invention.

[0064] 2. Traffic scene construction

[0065] The present invention focuses on the common two-way multi-lane straight road section in urban road scenes, whose length is L r The number of motor vehicle lanes in each direction is n v , and the outermost lane is a non-motorized vehicle lane. Considering that Vissim can effectively simulate the interaction between motor vehicles and non-motorized vehicles, the present invention uses Vissim as a microscopic traffic flow simulation platform and sets that non-motorized vehicles can occupy the motor vehicle lane closest to the non-motorized vehicle lane, so that the interaction scenario of motor vehicles and non-motorized vehicles mixed is closer to the real situation. Aiming at the effectiveness verification of the subsequent alternative evaluation indicators of vehicle-road collaborative perception effectiveness, the joint simulation of Vissim and Carla is used to achieve the quantitative output of perception indicators of motor vehicles and non-motorized vehicles mixed under vehicle-road collaborative conditions.

[0066] When constructing traffic scenarios, considering the possible impact on the vehicle-road cooperative perception efficiency, the traffic parameters mainly considered in this invention include traffic volume Q t , Traffic composition P t , ICV penetration rate ζ, Figure 2 This demonstrates the basic traffic scenarios used in this invention. For any traffic scenario, based on the sensor simulation capabilities of Carla, the ICVs and designated roadside sensing units within the scene can collaboratively perceive the scene. Considering the advantages of LiDAR in three-dimensional perception and collaborative operation, this invention uses LiDAR as the core sensing device for vehicle-road collaborative perception.

[0067] In order to facilitate the quantitative description of road scenes, Figure 2 The leftmost center of the middle section is the origin O, and a rectangular coordinate system is established with the horizontal rightward direction as the x-axis, the upward direction perpendicular to the x-axis as the y-axis, and the upward direction perpendicular to the xOy plane as the z-axis. In the context of the problem concerned by the present invention, xyz is a global coordinate system.

[0068] 3POG Build

[0069] POG is a spatial representation method for dynamic environments, which divides the scene into countless voxel units, and each voxel unit stores the occupancy probability of the corresponding grid space area. The present invention adopts a two-step method to construct the traffic scene POG, such as Figure 3 As shown in (a). First, by constraining the xyz coordinate range, the region of interest (ROI) required for the perception of mixed traffic scenes is determined. At any simulation time step, since the size of the traffic individual is known, its envelope can be created. Each envelope is voxelized and the POG is established through continuous updating of the simulation step. The specific voxelization process is as follows: Figure 3 (b) shown.

[0070] The voxelization of the present invention adopts a step-by-step process from local space to global coordinate frame. The specific steps are as follows:

[0071] Step 1: Grid point coordinate space conversion.

[0072] Set x v -y v -z v is the local coordinate system of each traffic entity, where y v Corresponding to the real-time moving direction of the individual, x v Perpendicular to the individual's forward direction, horizontally to the right. In the local coordinate frame of the traffic individual, the envelope is drawn by the length, width and height of the vehicle. vx Generate regular grid point data for the interval. Set at t i The azimuth of traffic individual j at time Then, the coordinate frame x is transformed into v -y v -z v The grid point data in is converted to the global coordinate system xyz.

[0073]

[0074] in:

[0075]

[0076] in: It is t i The rigid coordinate transformation matrix of traffic individual j at the moment, the roll angle and pitch angle are not considered in this invention, and are the local coordinates and global coordinates of individual j grid point, It is t i The global coordinate value of the bottom center of traffic individual j at time instant.

[0077] Step 2: Voxelization.

[0078] The grid points after coordinate transformation are voxelized, and the voxel size is consistent with the spacing of the grid points. Let V s V is the road traffic scene after voxel processing within the ROI. s The data format is a three-dimensional matrix, each matrix element corresponds to a voxel unit and stores the corresponding occupancy probability value. Let M v t i The voxel representing the traffic individual in the scene at the moment V s The position coordinate set within V s In M v The voxel value is set to 1, which is mathematically expressed as V s |t i ,M v= 1. The above is only for the case of a single time step. In order to obtain the POG of a certain traffic scene, continuous simulation and calculation are required.

[0079] V s Medium voxel unit v i For example, the corresponding occupancy probability It is obtained by evaluating according to formulas (3) to (4).

[0080]

[0081] in:

[0082]

[0083] Where: T is the total number of simulation steps, v i ∈M v and Respectively refer to whether the voxel vi is included in the coordinate set of the traffic grid point in Vs at a certain time step, is the occupancy probability of a voxel based on sample evaluation.

[0084] Assuming that λ is the total number of voxel units that are occupied at least once during the T-step simulation process, then 1≤i≤λ. To simplify the calculation process, we set Each voxel unit is regarded as a 0-1 binary variable, and there is Considering that the occupancy state of a single voxel cannot infer the state of other voxels, it can be considered that in V s The voxels within the ROI are mutually independent random variables. Under this condition, the joint distribution of multivariate variables within the ROI can be calculated according to formula (5). In the present invention, Ω r Characterization by Figure 3 The POG obtained by the workflow.

[0085]

[0086] in

[0087] 4. Perception Modeling

[0088] Due to the adverse effects of obstacles and the limited sensor field of view (FOV), the observed point of view (POG) under vehicle-infrastructure cooperative perception conditions differs from the actual POG. To quantitatively evaluate the POG under RPU and ICV observation conditions, it is necessary to model the perception process. Figure 4 (a) and 4(b) show the relevant parameters of RPU and ICV sensor respectively.

[0089] Let (x R ,y R ,zR ) T is the sensor center of RPU, with (x R ,y R ,z R ) T As the origin, establish the RPU sensor local coordinate system x s -y s -z s ( Figure 4 (a)). Let With φ s are the vertical and horizontal field of view angles of RPU, d s is the effective sensing range of RPU. Different from RPU, the sensor center of ICV changes with the movement of vehicle position. Taking the sensor center on ICV as the origin, the sensor center is similar to Figure 3 The vehicle local coordinate system x v -y v -z v The local coordinate system x of the ICV vehicle-mounted sensor is established by vs -y vs -z vs ( Figure 4 (b)). In line with the current mainstream configuration of ICV, to ensure that ICV can effectively perceive the surrounding environment, the horizontal field of view of the ICV sensor is set to 360 degrees. Let ξ vs is the vertical field of view angle of the ICV sensor, d vs The effective sensing distance of an ICV is denoted by . Considering the diversity of ICVs in future traffic scenarios, this invention considers both small passenger cars and large trucks as ICVs, and the vehicle type is randomly assigned. This invention defines a vehicle as being large if its vehicle length exceeds 8 meters.

[0090] Ray tracing is an important algorithm for image rendering in the field of computational geometry. When the voxelized road traffic scene (i.e., POG) is used as the observation object, ray tracing technology can be applied to the calculation of the three-dimensional view distance of the road. However, in this type of research, the occlusion of the driver by static obstacles is simulated mainly by judging whether the created ray intersects with the voxel unit. In order to effectively consider the impact of uncertain occlusion, the present invention proposes an OCRT algorithm for voxel units. The algorithm process is as follows: Figure 5 shown in FIG, and includes the following steps.

[0091] Step 1: Virtual light creation.

[0092] The present invention creates a virtual beam through the idea of ​​coordinate transformation. Specifically, χ, ψ and ρ are set to the original information collected by the RPU or ICV visual sensor, namely the horizontal angle, pitch angle and distance parameters, which actually correspond to the three coordinate components of the spherical coordinate system. Figure 6 As shown in the figure, a new coordinate system is established with χ, ψ and ρ as the three components of the Cartesian coordinate system. Through this transformation, the global coordinate system xyz with the sensor center C s The virtual light starting from is mapped to the vertical vector in the χ-ψ-ρ rectangular coordinate system.

[0093] The projection coordinates in the χ-ψ-ρ coordinate frame are (χ k ,ψ k ) T As an example, χ k With ψ k In xyz, it is the horizontal angle and pitch angle of the light. In χ-ψ-ρ, it is possible to sample along the vertical vector to obtain a set of discrete points on the virtual light (χ k ,ψ k ,ρ) T , where 0≤ρ≤d s The discrete points in the χ-ψ-ρ coordinate frame are mapped to the xyz coordinate space through equation (7).

[0094]

[0095] Where: Sphere2Cartesian[.] is the matrix that converts spherical coordinates to Cartesian coordinates, is from the local coordinate system x s -y s -z s The rotation transformation matrix to xyz coordinates, (α, β , γ) are the pitch, roll, and azimuth angles of the sensor in the xyz coordinate system. The meanings of other symbols remain unchanged.

[0096] Step 2: Occupancy probability retrieval.

[0097] The discrete points representing the light in the xyz coordinates are voxelized, that is, (x, y, z) T Convert to voxel position in Vs. s Each voxel unit in stores the occupancy probability value obtained by the previous T-step simulation. Based on the voxel position corresponding to the discrete point of the virtual ray, the original occupancy probability curve of the voxel unit along the ray can be obtained (see Figure 7 ).like Figure 6As shown in Figure 1, all rays within the sensor's field of view are calculated to obtain an occupancy probability matrix. The row number of the matrix represents the index of the virtual ray, and the column number represents the order of the voxel unit along the ray emission direction.

[0098] Step 3: Conditional occupancy probability calculation.

[0099] Since visual sensors cannot penetrate opaque objects, the observability of downstream voxel units along the emission direction of the virtual light will be affected by the upstream voxel units, that is, as the distance along the light direction increases, the observability of the object will decrease. There are two extreme scenarios for observing the state of voxel units: 1) There is a voxel unit with an occupancy probability of 1 upstream, that is, a fixed obstacle, then all downstream voxel units cannot be observed; 2) The occupancy probability of all upstream voxel units is 0, that is, there is no occlusion, then the state of the downstream voxel units can be directly observed without being affected. Based on this

[0100] On this basis, in order to consider the occlusion effect, the present invention obtains the conditional occupancy probability of the voxel unit under the sensor observation condition It is calculated by formula (7).

[0101]

[0102] in: is the observation condition v under the k-th perception unit i The probability of occupancy.

[0103] During the calculation process of step 3, the operation of converting the light into voxel units after discretization will cause the voxel units corresponding to different light rays to overlap, such as Figure 8 Under this condition, the same voxel is calculated on different virtual rays. To solve this problem, when a voxel is on the virtual rays of multiple perception units, the present invention takes The maximum value is taken as the representative value.

[0104] 5. Evaluation of Vehicle-Infrastructure Collaborative Perception Effectiveness

[0105] The steps for evaluating the vehicle-road cooperative perception effectiveness of the present invention are as follows:

[0106] Step 1: Construct the conditional occupancy probability under multiple sensory units.

[0107] Both RPU and ICV can perceive traffic scenes. Under this condition, the state of the same voxel can be observed by multiple perception units. Since the observations of the same voxel by different perception nodes are independent of each other, according to the Bayesian formula and the total probability formula, we have:

[0108]

[0109] Where: i∩S k Represents voxel v i Observed by the kth sensor unit (v i When it is not within the field of view of the kth perception unit, ), i∩S represents v i Conditions observed by all sensory units, n s is the total number of sensory units.

[0110] In formulas (8) and (9), represents the probability of the perception unit appearing. Since the roadside perception unit always exists after being deployed, In contrast, ICV as a sensing node has uncertainty, i.e. When building POG in Section 1.3, Figure 9 As shown in (a), the trajectory point P of the ICV visual sensor vs Voxelization is also performed to obtain the POG corresponding to the ICV sensor, which is recorded as Ω s Ω s Each voxel unit in stores the spatial occupancy probability of the ICV sensor. Therefore, in the present invention, Ω s The spatial position corresponding to the center of the voxel unit is used as the possible perception node of ICV, and Ω s The probability value stored in the voxel unit is used as the ICV perception unit On this basis, formula (9) can be transformed into formula (10).

[0111]

[0112] in: is the voxel v under the observation condition of the mth RPU i Occupancy probability, 1≤m≤n RPU , where n RPU is the number of RPUs, is the voxel v under the observation condition of the nth ICV sensor i Occupancy probability, 1≤n≤n icv , where n icv is Ω s The number of non-zero voxel units in is the probability of the nth ICV sensor appearing, based on Ω s get.

[0113] At the same time, as one of the components of the mixed traffic flow, ICV can actively share information such as the location and size of the vehicle, that is, the part of the POG occupied by the ICV does not need to be sensed. r Can be further divided into Ωicv and Ω nicv , and there is formula (11), under the condition of vehicle-road cooperative perception, the probability of ICV and RPU jointly observing the grid Ω r|S It can be calculated according to formula (12). Specifically, only the non-ICV part of the traffic flow needs to calculate the conditional occupancy probability based on the OCRT algorithm in Section 1.4.

[0114]

[0115]

[0116] in: is the voxel-by-voxel addition, Ω nicv|S It is the POG corresponding to non-ICV traffic individuals under the vehicle-road cooperative perception condition.

[0117] For Ω r|S The conditional probability occupancy value of any voxel unit in is , considering that a single voxel may be occupied by both ICV and non-ICV traffic individuals during traffic operation. Therefore, the occupancy probability of a single voxel unit can also be divided into two parts: ICV part and non-ICV part. Combined with formula (10), the Ω under the vehicle-road cooperative perception condition can be obtained. r|S The conditional occupancy probability of the voxel unit in is:

[0118]

[0119] in: and are the occupancy probabilities of the non-ICV part of voxel vi under the observation conditions of the mth RPU and the nth ICV sensor, is the occupancy probability of the ICV portion corresponding to voxel vi, and the meanings of other symbols remain unchanged.

[0120] In formula (13), and All are calculated by OCRT algorithm (see Figure 5 ). Specifically, the perception field of view is determined with the position of the RPU in the voxelized space or the voxel position corresponding to the ICV sensor as the center. Then, a virtual ray is created within the field of view, and the occupancy probability is retrieved based on Ωicv and Ωnicv. Finally, the conditional occupancy probability of any voxel unit within the perception range is calculated. Since m and n are generally greater than 1, it is necessary to calculate the conditional probability at each possible sensor position in turn, and use formula (14) to complete the addition of the probabilities. It should be noted that in is the occupancy probability of voxel vi obtained through continuous simulation in Section 3. Then order That is, the most ideal state under the conditions of vehicle-road cooperative perception is that the original occupancy probability of the voxel unit is fully observed.

[0121] In order to simplify the calculation process, it is assumed that the probability distribution of each voxel unit is still independent of each other under the vehicle-road cooperative observation condition. On this basis, the joint distribution of conditional occupancy probability can be calculated in a similar way to formula (5).

[0122]

[0123] in: μ is the number of voxels with non-zero conditional occupancy probability within the ROI under vehicle-road cooperative perception conditions. Since some voxels cannot be observed, μ≤λ.

[0124] Step 2: Quantitative evaluation of vehicle-road collaborative perception effectiveness.

[0125] In order to measure the effectiveness of vehicle-road cooperative perception, an intuitive way is to evaluate and The smaller the difference, the higher the perception efficiency. The cross entropy is used to measure the true distribution of voxel occupancy probability. and Considering the small size of non-motor vehicles, the number of voxels they occupy in the voxelized space is relatively small. In order to increase the perception efficiency of the vehicle-road cooperative system for vulnerable road users, the present invention uses weighted cross entropy to measure and Based on the independence assumption, the cross entropy can be calculated voxel by voxel, as shown in Equations (15)-(16). The smaller it is, the closer the occupancy probability under the vehicle-road cooperative perception conditions is to the actual situation.

[0126]

[0127] Where: n c It is the number of voxels occupied by non-motor vehicles among the non-zero voxels. If a voxel has been occupied by both motor vehicles and non-motor vehicles, it is considered to be a non-motor vehicle occupied voxel.

[0128] It should be noted that the present invention mainly focuses on the voxel unit as a binary variable of whether it is occupied, and does not consider the category of traffic individuals. Therefore, formula (15) is not a traditional weighted cross entropy calculation formula. The core idea of ​​formula (15) is to give a greater weight to the voxels occupied by non-motor vehicles. Since μ≤λ, there are conditional occupancy probabilities of some voxel units. is zero (i.e. not perceived by any perception unit), which makes the calculation of formula (15) invalid. Therefore, when calculating the cross entropy, the present invention uses T -1 / 10 to replace the zero value to ensure the numerical stability of the calculation.

[0129] 6. RPU Optimization Deployment

[0130] The complete process of the RPU optimization deployment method in the present invention is as follows: Figure 10 As shown. First, traffic scenarios are created by joint simulation using Vissim and Carla, and different traffic scenarios are simulated by adjusting traffic flow parameters (traffic volume, traffic ratio) and ICV penetration parameters. Then, Ω is obtained through the voxelization process. r ,Ω icv and Ω s The RPU placement parameters are set as decision variables, and the OCRT algorithm is combined with Ω icv Calculate the occupancy probability under the vehicle-road cooperative perception condition (Equation (14)). Compare the conditional occupancy probability Ω r|S Compared with Ω obtained through continuous simulation r and calculate To reduce To optimize the target, the RPU placement parameters are adjusted using a Bayesian optimizer. During the optimization process, constraints can be added as needed, such as requiring the distance between different RPUs to be greater than a certain threshold. The program terminates when the optimization process is complete.

[0131] The above is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for optimizing the deployment of roadside perception units under partial vehicle-road cooperative conditions, characterized in that: include: (1) Establish traffic scenarios based on the Vissim-Carla joint simulation method, and obtain the true value probability occupancy grid of various traffic scenarios through multi-time step continuous simulation; (2) Conduct parameterized modeling of the perception of roadside perception units and intelligent connected vehicles. To address the adverse effects of mobile occlusion on vehicle-road collaborative perception, establish a ray tracing algorithm that takes occlusion effects into account to simulate the collaborative perception process of roadside perception units and intelligent connected vehicles. (3) A ray tracing algorithm that takes occlusion effects into account is used to obtain an observable probabilistic occupancy grid under vehicle-infrastructure cooperative perception conditions. The cross entropy between the true probabilistic occupancy grid and the observable probabilistic occupancy grid is proposed as an alternative evaluation metric for vehicle-infrastructure cooperative perception for traffic flow monitoring. (4) The layout parameters of the roadside perception unit and the cross entropy based on the probability occupancy grid are used as decision variables and objective function values ​​respectively, and the Bayesian optimization method is used to complete the optimal deployment of the roadside perception unit.

2. The method for optimizing deployment of roadside perception units under partial vehicle-road collaboration conditions according to claim 1 is characterized in that: The steps for obtaining the true value probability occupancy grid in step (1) include: 1) Establish a global xyz coordinate system and determine the region of interest for mixed-use scene perception by constraining the xyz coordinate range; 2) At any simulation time step, create the envelope of the traffic individual, voxelize each envelope, and obtain the true value probability occupancy grid through continuous updating of the simulation step.

3. The method for optimizing deployment of roadside perception units under partial vehicle-road cooperation conditions according to claim 2 is characterized in that: Voxelization uses a step-by-step process from local space to a global coordinate frame as follows: Step 1: Grid point coordinate space conversion Set x v -y v -z v is the local coordinate system of each traffic entity, where y v Corresponding to the real-time moving direction of the individual, x v Perpendicular to the individual's forward direction, horizontally to the right; in the local coordinate frame of the traffic individual, the envelope is outlined by the length, width and height of the vehicle; within the envelope, δ vx Generate regular grid point data for the interval; set t i The azimuth of traffic individual j at time Then, the coordinate frame x is transformed into v -y v -z v The grid point data in is converted to the global coordinate system xyz; in: in: It is t i The rigid coordinate transformation matrix of traffic individual j at the moment, and are the local coordinates and global coordinates of individual j grid point, It is t i The global coordinate value of the bottom center of traffic individual j at the moment; Step 2: Voxelization The grid points after coordinate transformation are voxelized, and the voxel size is consistent with the spacing of the grid points; let V s V is the road traffic scene after voxel processing in the region of interest. s The data format is a three-dimensional matrix, each matrix element corresponds to a voxel unit and stores the corresponding occupancy probability value; let M v t i The voxel representing the traffic individual in the scene at the moment V s The position coordinate set within V s In M v The voxel value is set to 1, which is mathematically expressed as V s |t i ,M v =1; The above is only for the case of a single time step. In order to obtain the POG of a certain traffic scene, continuous simulation and calculation are required; with V s Medium voxel unit v i For example, the corresponding occupancy probability According to equations (3) to (4), we can obtain in: Where: T is the total number of simulation steps, v i ∈M v and Respectively refers to the voxel v at a certain time step i Is or is not included in the traffic grid point in V s In the coordinate set of , is the occupancy probability of the voxel based on the sample evaluation; Assuming that λ is the total number of voxel units that are occupied at least once during the T-step simulation process, then 1≤i≤λ; to simplify the calculation process, it is set under the condition of sufficient simulation steps Each voxel unit is regarded as a 0-1 binary variable, and there is Considering that the occupancy state of a single voxel cannot infer the state of other voxels, it is believed that in V s The voxels within the region are independently distributed random variables; under this condition, the joint distribution of the multivariate variables within the region of interest is calculated according to formula (5); in 4. The method for optimizing deployment of roadside perception units under partial vehicle-road collaboration conditions according to claim 1 is characterized in that: The parameterized modeling of perception of roadside perception units and intelligent connected vehicles includes: Let (x R ,y R ,z R ) T is the sensor center of the roadside perception unit, with (x R ,y R ,z R ) T As the origin, establish the sensor local coordinate system x of the roadside perception unit s -y s -z s ,make With φ s are the vertical and horizontal field of view angles of the roadside perception unit, d s is the effective sensing range of the roadside sensing unit; Taking the center of the sensor on the intelligent connected car as the origin, y vs Corresponding to the real-time moving direction of the individual, x vs Perpendicular to the individual's forward direction and horizontally to the right, establish the local coordinate system x of the onboard sensor of the intelligent connected vehicle vs -y vs -z vs , set the horizontal field of view of the ICV sensor to 360 degrees; let ξ vs is the vertical field of view angle of the ICV sensor, d vs is the effective sensing distance of ICV.

5. The method for optimizing deployment of roadside perception units under partial vehicle-road cooperation conditions according to claim 2 is characterized in that: Obtain the observable probability occupancy grid under the vehicle-road cooperative perception condition as follows: Under the cooperative perception conditions of the roadside perception unit and the intelligent connected vehicle, the voxel v i Occupancy probability of the condition observed by all sensing units: in: and are the occupancy probabilities of the non-ICV part of voxel vi under the observation conditions of the mth RPU and the nth ICV sensor, is the voxel v i The occupancy probability of the corresponding ICV part; Assuming that the probability distribution of each voxel unit is still independent under the vehicle-road cooperative observation condition, the joint distribution of conditional occupancy probability is in: μ is the number of voxels with non-zero conditional occupancy probability within the region of interest under vehicle-road cooperative perception conditions.

6. The method for optimizing deployment of roadside perception units under partial vehicle-road collaboration conditions according to claim 1 is characterized in that: Based on the independence assumption, the cross entropy is calculated voxel by voxel, as shown in Equations (15)-(16); The smaller it is, the closer the occupancy probability under the vehicle-road cooperative perception condition is to the actual situation; Where: n c It is the number of voxels occupied by non-motor vehicles among the non-zero voxels. If a voxel has been occupied by both motor vehicles and non-motor vehicles, it is considered to be a non-motor vehicle occupied voxel.

Images