A target vehicle selection method and device

Abstract

The application discloses a target vehicle selection method and device, wherein the target vehicle selection method comprises the following steps: acquiring the state information of the vehicle, the road geometry information and the state information of the candidate vehicles within the preset range of the vehicle in real time; calculating the lane information of the vehicle according to the acquired state information of the vehicle and the road geometry information; inputting the obtained lane information of the vehicle and the state information of the candidate vehicles into a pre-trained machine learning model to determine whether each candidate vehicle is a target vehicle; if multiple target vehicles are determined, screening out a unique key target vehicle according to the relative motion relationship between the vehicle and the multiple target vehicles. The application introduces a machine learning method to determine the target vehicle, thereby improving the accuracy of the target vehicle determination; and the unique key target vehicle is accurately selected from the multiple target vehicles for output, so that the vehicle is not controlled or adjusted too frequently due to the influence of the multiple target vehicles, thereby improving the traffic efficiency and the driving comfort.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle technology, specifically relating to a target vehicle selection method and device. Background Technology

[0002] In real-world road environments, complex driving conditions pose a severe challenge to the perception, decision-making, and control capabilities of intelligent driving vehicles. The ability to accurately and efficiently select the most important target vehicle from multiple candidate vehicles in complex environments directly impacts the safety and comfort of intelligent vehicle operation.

[0003] Due to the high real-time nature and complexity of real-world driving scenarios, the selection of the target vehicle is currently primarily based on manually defined thresholds determined by the positions of candidate vehicles and the current vehicle, as well as their relative lateral speeds. To avoid selection errors, the thresholds are set quite strictly, often resulting in a delay before the target vehicle can be locked after vehicles from adjacent lanes have entered the lane, leading to unnecessary sudden deceleration or even dangerous situations. Conversely, if the thresholds are set too leniently, it can easily cause frequent acceleration, deceleration, and attitude adjustments by the vehicle, reducing both traffic efficiency and passenger comfort. Summary of the Invention

[0004] The technical problem to be solved by the embodiments of the present invention is to provide a target vehicle selection method and apparatus to reduce the error rate of target vehicle selection by intelligent driving vehicles under complex working conditions and improve driving safety and comfort.

[0005] To solve the above-mentioned technical problems, the present invention provides a target vehicle selection method, comprising:

[0006] Real-time acquisition of vehicle status information, road geometry information, and candidate vehicle status information within a preset range;

[0007] The lane information is calculated based on the acquired vehicle status information and road geometry information;

[0008] The obtained lane information and candidate vehicle status information are input into a pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle.

[0009] If multiple target vehicles are identified, the single key target vehicle is selected based on the relative motion relationship between the vehicle and the multiple target vehicles.

[0010] Furthermore, the step of selecting a unique key target vehicle based on the relative motion relationship between the vehicle and multiple target vehicles specifically includes:

[0011] Calculate the time to collision (TTC) between this vehicle and each target vehicle: TTC = s / v rels is the distance from the front of this vehicle to the rear of the target vehicle, v rel The relative speed between the two vehicles;

[0012] Select the target vehicle with the lowest TTC value from the calculated TTC values ​​as the key target vehicle.

[0013] Furthermore, the real-time acquired vehicle status information includes vehicle speed, acceleration, steering wheel angle, and yaw rate; road geometry information includes lane line information and road edge information; and candidate vehicle status information includes candidate vehicle size, orientation angle, lateral and longitudinal position, lateral and longitudinal velocity, and lateral and longitudinal acceleration.

[0014] Furthermore, the geometry of the lane lines for this lane is calculated based on the following lane line equations:

[0015] y = c0 + c1x + c2x 2 +c3x 3

[0016] Where y is the lateral position in the vehicle coordinate system, x is the longitudinal position, and c0, c1, c2, and c3 are the coefficients of the zeroth to third order terms, respectively.

[0017] Furthermore, the method also includes: based on the acquired lane line information of the current lane, further determining whether the vehicle is in a state of driving along the lane line; if so, directly determining the edge of the current lane based on the lane line information; if not in a state of driving along the lane line or without valid lane line input, predicting the trajectory of the vehicle based on the road edge information and the vehicle's motion state, and widening the predicted trajectory of the vehicle to both sides by a certain distance as the virtual edge of the current lane.

[0018] Furthermore, the trajectory prediction of this vehicle is based on the uniform speed model: X K =A×X K-1 X K This is a general recursive formula, representing the recursive relationship from time k-1 to time k, where A = [Δt, Δt, 1, 1]; time k-1 is time t, and time k is time t + Δt.

[0019] By setting Δt as the sampling point at different times, the state vector X of the vehicle at different times is obtained. K =x(t) = (x,y,vx,vy) T ,x(t+Δt)=(x(t)+Δt×vx,y(t)+Δt×vy,vx,vy) T Where x, y, vx, and vy represent the longitudinal position, lateral position, longitudinal velocity, and lateral velocity of the vehicle in the vehicle coordinate system, respectively; x(t) is the state vector of the vehicle at time t; and x(t+Δt) is the state vector of the vehicle at time t+Δt.

[0020] Furthermore, the pre-trained machine learning model is built based on the determined lane information and candidate vehicle state information, specifically in the following ways:

[0021] Whether a candidate vehicle belongs to the target vehicle is determined by whether the depth of its intrusion into this lane exceeds the intrusion threshold based on the candidate vehicle's historical trajectory.

[0022] If the candidate vehicle is determined to be the target vehicle, then time backtracking and spatial backtracking are performed:

[0023] Time backtracking is defined as the moment when the depth of the intrusion into the lane exceeds the intrusion threshold as the end time of time backtracking. The time backtracking is carried out for a time threshold duration to obtain the start time of time backtracking. From the start time of time backtracking, the candidate vehicle is identified as the target vehicle and used as the target vehicle identification data in the time dimension.

[0024] Spatial backtracking involves tracing back candidate vehicles whose intrusion depth exceeds the intrusion threshold to the position where their intrusion depth exceeds the spatial threshold. From that position, the candidate vehicle is identified as the target vehicle and used as target vehicle identification data in the spatial dimension.

[0025] The present invention also provides a target vehicle selection device, comprising:

[0026] The acquisition module is used to acquire the vehicle's status information, road geometry information, and the status information of candidate vehicles within a preset range in real time.

[0027] The calculation module is used to calculate the lane information based on the acquired vehicle status information and road geometry information;

[0028] The judgment module is used to input the obtained lane information and candidate vehicle status information into the pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle.

[0029] The output module is used to filter out the unique key target vehicle based on the relative motion relationship between the current vehicle and the multiple target vehicles when multiple target vehicles are identified.

[0030] Furthermore, the output module filters out a unique key target vehicle based on the relative motion relationship between the vehicle and multiple target vehicles, specifically including:

[0031] Calculate the time to collision (TTC) between this vehicle and each target vehicle: TTC = s / v rel s is the distance from the front of this vehicle to the rear of the target vehicle, v rel The relative speed between the two vehicles;

[0032] Select the target vehicle with the lowest TTC value from the calculated TTC values ​​as the key target vehicle.

[0033] Furthermore, the calculation module is also used to: determine whether the vehicle is driving along the lane line based on the obtained lane line information of the current lane; if so, determine the edge of the current lane directly based on the lane line information of the current lane; if not driving along the lane line or without valid lane line input, predict the trajectory of the vehicle based on the road edge information and the vehicle's motion state, and widen the predicted trajectory of the vehicle to both sides by a certain distance as the virtual edge of the current lane.

[0034] Implementing this invention has the following beneficial effects: it introduces machine learning methods to determine target vehicles, improving the accuracy of target vehicle determination; by accurately selecting a single key target vehicle from multiple target vehicles for output, it reduces the need for excessively frequent vehicle control or adjustment due to the influence of multiple target vehicles, thereby improving traffic efficiency and driving comfort; at the same time, it fully utilizes the vehicle's motion information and road geometry information to solve the target vehicle selection problem in the case of missing lane lines; furthermore, it uses historical trajectory information to establish a machine learning dataset for target vehicle selection through spatiotemporal backtracking, and as the dataset becomes richer, it has the ability to continuously improve the accuracy of target vehicle determination. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart illustrating a target vehicle selection method according to an embodiment of the present invention.

[0037] Figure 2 This is a schematic diagram illustrating the specific process of a target vehicle selection method according to an embodiment of the present invention. Detailed Implementation

[0038] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0039] Please refer to Figure 1 As shown, Embodiment 1 of the present invention provides a target vehicle selection method, including:

[0040] Real-time acquisition of vehicle status information, road geometry information, and candidate vehicle status information within a preset range;

[0041] The lane information is calculated based on the acquired vehicle status information and road geometry information;

[0042] The obtained lane information and candidate vehicle status information are input into a pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle.

[0043] If multiple target vehicles are identified, the single key target vehicle is selected based on the relative motion relationship between the vehicle and the multiple target vehicles.

[0044] Specifically, the real-time vehicle status information includes vehicle speed, acceleration, steering wheel angle, yaw rate, etc. Road geometry information includes lane line information, road edge information, etc.

[0045] This lane information can be calculated based on road geometry and vehicle status information. A deep learning-based lane recognition module can provide a cubic fit curve for the lane line, deriving the lane line's geometry based on the following lane line equation:

[0046] y = c0 + c1x + c2x 2 +c3x 3

[0047] Where y is the lateral position in the vehicle coordinate system, x is the longitudinal position, and c0, c1, c2, and c3 are the coefficients of the zeroth to third order terms of the equation, respectively.

[0048] If lane line information for this lane is obtained, it is further determined whether the vehicle is traveling along the lane line. If so, the edge of the lane is determined directly based on the lane line information. If the vehicle is not traveling along the lane line or there is no valid lane line input, the vehicle trajectory is predicted based on the road edge information and the vehicle's motion state. If the road geometry information is missing, the trajectory is also predicted based on the vehicle's motion state.

[0049] The trajectory of this vehicle can be predicted using a uniform velocity model, i.e., X. K =A×X K-1 X K This is a general recursive formula, representing the recursive relationship from time k-1 to time k, where A = [Δt, Δt, 1, 1]; time k-1 is time t, and time k is time t + Δt. By setting Δt as the sampling point at different times, the state vector X of the vehicle at different times can be obtained. K =x(t) = (x,y,vx,vy) T ,x(t+Δt)=(x(t)+Δt×vx,y(t)+Δt×vy,vx,vy) T x, y, vx, vy represent the longitudinal position, lateral position, longitudinal velocity, and lateral velocity of the vehicle in the local map (which can be the vehicle coordinate system at this moment), respectively. x(t) represents the state vector of the vehicle at time t, and x(t+Δt) represents the state vector of the vehicle at time t+Δt.

[0050] After predicting the vehicle's trajectory, the lane is widened to both sides by a certain distance (as an example, the widening distance is generally 1.5 meters on each side) to form the virtual lane edge, thereby obtaining lane information containing the virtual lane edge.

[0051] The candidate vehicle status information acquired in real time in this embodiment includes candidate vehicle size, orientation angle, lateral and longitudinal position, lateral and longitudinal velocity, and lateral and longitudinal acceleration. The obtained lane information and candidate vehicle information data are input into a pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle.

[0052] A pre-trained machine learning model may identify multiple target vehicles. In this embodiment, a single key target vehicle is selected based on the relative motion relationship between the vehicle and the multiple target vehicles, and output to the downstream module. This avoids taking excessive and unnecessary control or adjustment actions on the vehicle for multiple target vehicles. Specifically, the time to collision (TTC) between the vehicle and each target vehicle can be calculated separately: TTC = s / v rel s represents the distance from the front of this vehicle to the rear of the target vehicle, v rel The relative speeds of the two vehicles are represented by their respective Total Time Toll (TTC) values. Then, the vehicles with the lowest TTC values ​​are selected as the critical target vehicles. TTC can be calculated directly in the vehicle coordinate system. It is understood that other TTC calculation methods considering acceleration or image-based TTC estimation methods can also be used. Through the aforementioned screening mechanism for multiple target vehicles, this invention enables the accurate and efficient selection of the most critical target vehicles from multiple candidate vehicles ahead in complex environments. This allows for accurate and timely adjustments to the vehicle's position, speed, and attitude, while also improving vehicle safety and comfort.

[0053] In this embodiment, the pre-trained machine learning model is built based on the determined lane information and candidate vehicle status information. During data annotation, a historical information backtracking method is used, that is, whether a candidate vehicle belongs to the target vehicle is determined based on whether the depth of its historical trajectory intrusion into the lane exceeds an intrusion threshold (as an example, the intrusion threshold can generally be set to 0.2 meters). Setting the intrusion threshold can avoid errors caused by accidental driving mistakes entering the lane and being mistakenly identified as a target vehicle. If a candidate vehicle is determined to be a target vehicle, then time backtracking and spatial backtracking are performed. The time backtracking duration is set to a time threshold (as an example, the time threshold can generally be set to 0.6 seconds). That is, the time backtracking ends when the depth of the intrusion into the lane exceeds the intrusion threshold. Going back 0.6 seconds, the start time of the time backtracking is obtained. From this start time, the candidate vehicle is determined to be a target vehicle, and this is used as the target vehicle judgment data in the time dimension. Spatial backtracking is performed until the candidate vehicle begins to intrude into the lane and reaches the spatial threshold (as an example, the spatial threshold can generally be set to 0 meters). That is, for a candidate vehicle whose depth of intrusion into the lane exceeds the intrusion threshold, backtracking is performed to the position where it first showed the tendency to intrude (the spatial threshold is 0 meters). From this position, the candidate vehicle is determined to be a target vehicle, and this is used as the target vehicle judgment data in the spatial dimension.

[0054] Once the annotations are complete, the machine learning model can be trained using the data from each time step and the M frames preceding that time step, along with the annotation results. Each frame is a time slice, for example, every 0.1 seconds is considered one frame; 10 frames can be acquired per second, with 6 frames representing the target vehicle and 4 frames representing non-target vehicles. Supervised learning methods such as decision trees or support vector machines can be used to train the model. The specific learning strategy of the model can be adjusted in various ways to achieve high efficiency and accuracy.

[0055] Corresponding to the target vehicle selection method described in Embodiment 1 of the present invention, Embodiment 2 of the present invention further provides a target vehicle selection device, comprising:

[0056] The acquisition module is used to acquire the vehicle's status information, road geometry information, and the status information of candidate vehicles within a preset range in real time.

[0057] The calculation module is used to calculate the lane information based on the acquired vehicle status information and road geometry information;

[0058] The judgment module is used to input the obtained lane information and candidate vehicle status information into the pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle.

[0059] The output module is used to filter out the unique key target vehicle based on the relative motion relationship between the current vehicle and the multiple target vehicles when multiple target vehicles are identified.

[0060] Furthermore, the output module filters out a unique key target vehicle based on the relative motion relationship between the vehicle and multiple target vehicles, specifically including:

[0061] Calculate the time to collision (TTC) between this vehicle and each target vehicle: TTC = s / v rel s is the distance from the front of this vehicle to the rear of the target vehicle, v rel The relative speed between the two vehicles;

[0062] Select the target vehicle with the lowest TTC value from the calculated TTC values ​​as the key target vehicle.

[0063] Furthermore, the calculation module is also used to: determine whether the vehicle is driving along the lane line based on the obtained lane line information of the current lane; if so, directly determine the edge of the current lane based on the lane line information; if not driving along the lane line or without a valid lane line input, predict the trajectory of the vehicle based on the road edge information and the vehicle's motion state, and widen the predicted trajectory of the vehicle to both sides by a certain distance as the virtual edge of the current lane.

[0064] For the working principle and process of this embodiment, please refer to the description of Embodiment 1 of the present invention, which will not be repeated here.

[0065] As can be seen from the above description, compared with the prior art, the beneficial effects of the present invention are as follows: It introduces machine learning methods to determine target vehicles, improving the accuracy of target vehicle determination; by accurately selecting a single key target vehicle from multiple target vehicles for output, it reduces the need for excessively frequent control or adjustment of the vehicle due to the influence of multiple target vehicles, thereby improving traffic efficiency and driving comfort; simultaneously, it fully utilizes the vehicle's motion information and road geometry information to solve the target vehicle selection problem in the case of missing lane lines; furthermore, it utilizes historical trajectory information to establish a machine learning dataset for target vehicle selection through spatiotemporal backtracking, and as the dataset becomes richer, it has the ability to continuously improve the accuracy of target vehicle determination.

[0066] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for selecting a target vehicle, characterized in that, include: The system acquires real-time vehicle status information, road geometry information, and candidate vehicle status information within a preset range. The real-time vehicle status information includes vehicle speed, acceleration, steering wheel angle, and yaw rate. The road geometry information includes lane line information and road edge information. The candidate vehicle status information includes candidate vehicle size, orientation angle, lateral and longitudinal position, lateral and longitudinal velocity, and lateral and longitudinal acceleration. The lane information is calculated based on the acquired vehicle status information and road geometry information; The obtained lane information and candidate vehicle status information are input into a pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle. If multiple target vehicles are identified, the unique key target vehicle is selected based on the relative motion relationship between this vehicle and the multiple target vehicles. The pre-trained machine learning model is built based on the determined lane information and candidate vehicle state information, specifically in the following ways: Whether a candidate vehicle belongs to the target vehicle is determined by whether the depth of its intrusion into this lane exceeds the intrusion threshold based on the candidate vehicle's historical trajectory. If a candidate vehicle is determined to be a target vehicle, then time backtracking and spatial backtracking are performed: Time backtracking starts at the moment when the depth of the intrusion into the lane exceeds the intrusion threshold, and then backtracks by a time threshold duration to obtain the start time of the time backtracking. From the start time of the time backtracking, the candidate vehicle is determined to be a target vehicle, and this is used as the target vehicle determination data in the time dimension. Spatial backtracking involves tracing back the candidate vehicle whose depth of intrusion into the lane exceeds the intrusion threshold to the position where the depth of intrusion into the lane exceeds the spatial threshold. From the position corresponding to the spatial threshold, the candidate vehicle is determined to be a target vehicle, and this is used as the target vehicle determination data in the spatial dimension.

2. The method according to claim 1, characterized in that, The process of selecting a unique key target vehicle based on the relative motion relationship between this vehicle and multiple target vehicles specifically includes: Calculate the time of collision (TTC) between this vehicle and each target vehicle: TTC = s / v rel , s This is the distance from the front of this vehicle to the rear of the target vehicle. v rel The relative speed between the two vehicles; Select the target vehicle with the lowest TTC value from the calculated TTC values ​​as the key target vehicle.

3. The method according to claim 1, characterized in that, Calculate the geometry of the lane lines for this lane based on the following lane line equations: y=c 0 +c 1 x+c 2 x 2 +c 3 x 3 in, y This represents the lateral position in the vehicle coordinate system. x Vertical position c 0、 c 1. c 2. c 3 represents the coefficients of the zeroth to third-order terms.

4. The method according to claim 3, characterized in that, Also includes: Based on the lane line information of this lane, it is further determined whether the vehicle is traveling along the lane line. If so, the edge of this lane is determined directly based on the lane line information. If the vehicle is not traveling along the lane line or there is no valid lane line input, the trajectory of this vehicle is predicted based on the road edge information and the vehicle's motion state. The predicted trajectory of this vehicle is then widened to both sides by a certain distance as the virtual edge of this lane.

5. The method according to claim 4, characterized in that, The vehicle trajectory is predicted using a uniform speed model: X K =A × X K-1 , X K This is a general recursive formula, representing... k-1 Time's up k The recursive relationship of time, A=[ Δ t, Δ t,1,1] ; k-1 Time for t time, k Time for t+ Δ t time; By Δ t Using sampling points at different times, calculate the state vector of the vehicle at different times. X K = x(t)=(x,y, vx,vy) T , x ( t +Δ t )=( x ( t )+Δ t × vx,y ( t )+Δ t × vy, vx,vy ) T ; in, x, y , vx,vy These represent the vehicle's longitudinal position, lateral position, longitudinal velocity, and lateral velocity in the vehicle coordinate system, respectively. x(t) for t The state vector of this vehicle at any given time. x ( t +Δ t )for t +Δ t The state vector of this vehicle at any given time.

6. A target vehicle selection device, characterized in that, include: The acquisition module is used to acquire the vehicle's status information, road geometry information, and candidate vehicle status information within a preset range in real time. The real-time acquired vehicle status information includes the vehicle's speed, acceleration, steering wheel angle, and yaw rate; the road geometry information includes lane line information and road edge information; and the candidate vehicle status information includes candidate vehicle size, orientation angle, lateral and longitudinal position, lateral and longitudinal velocity, and lateral and longitudinal acceleration. The calculation module is used to calculate the lane information based on the acquired vehicle status information and road geometry information; The judgment module is used to input the obtained lane information and candidate vehicle status information into a pre-trained machine learning model to determine whether each candidate vehicle belongs to the target vehicle. The pre-trained machine learning model is built based on the determined lane information and candidate vehicle status information. Specifically, it includes: determining whether a candidate vehicle belongs to the target vehicle based on whether the depth of its intrusion into the lane exceeds an intrusion threshold according to its historical trajectory; if the candidate vehicle is determined to belong to the target vehicle, then time backtracking and spatial backtracking are performed: time backtracking starts at the moment when the depth of intrusion into the lane exceeds the intrusion threshold, and backtracks by a time threshold duration to obtain the start time of time backtracking. From the start time of time backtracking, the candidate vehicle is determined to be the target vehicle, serving as the target vehicle judgment data in the time dimension; spatial backtracking involves tracing back the candidate vehicle whose depth of intrusion into the lane exceeds the intrusion threshold to the position where the depth of intrusion into the lane exceeds the spatial threshold, and from the position corresponding to the spatial threshold, the candidate vehicle is determined to be the target vehicle, serving as the target vehicle judgment data in the spatial dimension. The output module is used to filter out the unique key target vehicle based on the relative motion relationship between the current vehicle and the multiple target vehicles when multiple target vehicles are identified.

7. The apparatus according to claim 6, characterized in that, The output module selects a unique key target vehicle based on the relative motion relationship between the current vehicle and multiple target vehicles, specifically including: Calculate the time of collision (TTC) between this vehicle and each target vehicle: TTC = s / v rel , s This is the distance from the front of this vehicle to the rear of the target vehicle. v rel The relative speed between the two vehicles; Select the target vehicle with the lowest TTC value from the calculated TTC values ​​as the key target vehicle.

8. The apparatus according to claim 6, characterized in that, The calculation module is also used to: further determine whether the vehicle is driving along the lane line based on the obtained lane line information of the current lane; if so, directly determine the edge of the current lane based on the lane line information of the current lane; if not driving along the lane line or without valid lane line input, predict the trajectory of the vehicle based on the road edge information and the vehicle's motion state, and widen the predicted trajectory of the vehicle to both sides by a certain distance as the virtual edge of the current lane.

Images