Method, computer system, readable medium for target selection in proximity of a vehicle

Abstract

Methods, computer systems, readable media for target selection in proximity to a vehicle are provided. The method includes obtaining vehicle state information including dynamic information about the vehicle; predicting, by a processor, a first trajectory of the vehicle for a first prediction time layer based on the vehicle state information; detecting, by a sensor of the vehicle, a road user in proximity to the vehicle; determining state information from the detected road user including dynamic information about the road user; predicting, by the processor, a second trajectory of the vehicle for the first prediction time layer based on the vehicle state information and the road user state information; and performing, by the processor, a first similarity comparison of the predicted first trajectory and the predicted second trajectory of the vehicle to determine, for the first prediction time layer, whether the detected road user is a potential target of the vehicle.

Description

Technical Field

[0001] This disclosure relates to a method for target selection in the vicinity of a vehicle, particularly a vehicle with advanced driver assistance systems and / or an autonomous vehicle, a computer system and a non-transitory computer-readable medium. Background Technology

[0002] Driving a vehicle involves continuous interaction with other road users. The advanced driver assistance systems (ADAS) or autonomous driving features of modern vehicles rely on the perception of these road users to adjust their control parameters (e.g., adaptive cruise control (ACC), automatic emergency braking (AEB)), issue warnings (e.g., lateral collision warning (LCW)), and / or plan their own autonomous driving (AD) routes.

[0003] Target selection is traditionally a rule-based system, and different rules are required for different applications. For example, an ACC system can guide a vehicle to adjust its speed based on targets selected from perceived road users. Autonomous driving systems may need to consider a variety of potential driving behaviors of road users and therefore select appropriate targets to ensure safe interaction with other users. However, selecting appropriate targets is computationally intensive.

[0004] Therefore, an improved method for selecting objectives for road users is needed. Summary of the Invention

[0005] This disclosure provides a method implemented by a computer, a computer system, and a non-transitory computer-readable medium. Embodiments are shown in the specification and accompanying drawings.

[0006] In one aspect, the present invention relates to a computer-implemented method for target selection near a vehicle. The method includes: obtaining vehicle state information in a first step, wherein the vehicle state information includes dynamic information about the vehicle. The method includes: in a further step, predicting a first trajectory of the vehicle for a first prediction time horizon based on the vehicle state information. The method includes: in a further step, detecting road users near the vehicle. The method includes: in a further step, determining state information from the detected road users, wherein the state information includes dynamic information about the road users. The method includes: in a further step, predicting a second trajectory of the vehicle for the first prediction time horizon based on the road user state information and the vehicle state information. In a further step, the method includes performing a first similarity comparison between the first and second predicted trajectories to determine whether, for the first prediction time horizon, the detected road user is a potential target of the vehicle.

[0007] The vehicle can be a vehicle with autonomous driving (AD) features or a vehicle with advanced driver assistance systems (ADAS). Vehicle status information includes dynamic information about the vehicle. In this context, dynamic means information about changes and / or information related to dynamics (i.e., moving surrounding objects). Specifically, vehicle status information may include information about position, steering angle, throttle and / or brake input, vehicle acceleration and / or speed, turn signal status, AD and / or ADAS feature status, and / or the current navigation route. This information can be obtained from appropriate sensors, such as range sensors integrated into the vehicle. Vehicle status information may specifically include vehicle status information from the past.

[0008] In this method, vehicle state information is processed to predict a first trajectory of the vehicle for a first prediction time layer. The prediction time layer is a future given time point in time from which the prediction is performed. The first trajectory predicted based on the vehicle state information is specifically the future trajectory of the vehicle, i.e., the path the vehicle may take in the future, considering only the vehicle itself. In particular, the first trajectory is predicted based solely on the vehicle state information.

[0009] A further step involves detecting road users near the vehicle. This can be done using sensors, such as one or more image sensors, and particularly one or more radar, lidar, and / or camera sensors integrated into the vehicle.

[0010] Similarly, road user state information includes dynamic information about road users. Specifically, state information from road users may include information about the location, acceleration, and / or speed of other road users different from the vehicle. A road user is, for example, another vehicle, motorcycle, or truck using the same road and / or located near the vehicle. The total number of road users may include at least one or more other road users. Road user state information may specifically include road user state information from the past, specifically from earlier and / or previous prediction timeframes.

[0011] Then, the road user state information is processed together with the vehicle state information. Thus, a second trajectory of the vehicle can be predicted for the same first prediction time level. Specifically, the second trajectory predicted based on the road user state information and the vehicle state information depends on the future trajectory of the vehicle and other road users; that is, it considers the overall situation of other road users and the vehicle itself to determine the possible paths the vehicle might take in the future.

[0012] Then, a first similarity comparison is performed by comparing the vehicle's first predicted trajectory based on vehicle state information with the vehicle's second predicted trajectory based on both road user state information and vehicle state information. This determines whether a road user is a potential target for the vehicle at the first prediction time level. A target is a road user who may have a significant impact on the vehicle's future driving trajectory. Specifically, if the similarity comparison yields a similarity below a predetermined threshold, the road user is considered a potential target. Conversely, if the similarity comparison yields a similarity above the predetermined threshold, the road user is not considered a potential target and can be discarded.

[0013] In other words, a first predicted trajectory that only considers the vehicle is compared with a second predicted trajectory that considers the vehicle and at least one other road user. If the two trajectories are similar to each other to some extent, i.e., their similarity is above a certain threshold, then there is no significant difference between the two trajectories, and thus the other road user is not considered a potential target. If the two trajectories are different from each other to some extent, i.e., their similarity is below a certain threshold, then there is a significant difference between the two trajectories, and thus another road user is considered a potential target.

[0014] According to the implementation, the vehicle status information also includes static information about the vehicle. Alternatively or additionally, the status information from road users includes static information about the road users.

[0015] In this context, "static" refers to unchanging information and / or information related to static, i.e., non-moving surroundings. Specifically, in this context, "static" means surroundings that restrict the vehicle itself and / or the manipulation of a particular road user. As an example, vehicle status information may include information about the location of buildings, roads, or other non-moving structures (such as lane structures, shoulders, and / or guardrails) near the vehicle. Additionally or alternatively, vehicle status information may include information about traffic rules, speed limits, lane directions, and / or traffic lights associated with the vehicle.

[0016] Similarly, status information from road users may include information about the location of a particular road user or buildings, roads, or other non-moving structures (such as lane structures, curbs, and / or guardrails) near the road user. Additionally or alternatively, vehicle status information may include information about traffic rules, speed limits, lane directions, and / or traffic lights relevant to a particular road user.

[0017] Static information can be derived from map data and / or perception sensors, such as camera-based traffic light detection, traffic sign detection, and / or lane marking detection.

[0018] According to the implementation method, the step of performing a first similarity comparison between the first predicted trajectory of the vehicle based on vehicle state information and the second predicted trajectory of the vehicle based on road user state information and vehicle state information is used to determine the relevance threshold of the road user to the first prediction time layer.

[0019] The relevance threshold is a threshold above which other road users are considered targets, and below which other road users are not considered targets of the vehicle. In particular, by considering all road users, a separate threshold can be determined for any given situation.

[0020] Therefore, road users who are irrelevant to the target determination are not considered, and computational power can be reduced.

[0021] According to an implementation, the road user includes a first road user. The step of determining state information from the detected road user includes determining state information from the first road user. The method further includes: predicting a third trajectory of the vehicle for a first prediction time layer based on the first road user state information and the vehicle state information; performing a second similarity comparison between a second predicted trajectory of the vehicle based on the road user state information and the vehicle state information and a third predicted trajectory of the vehicle based on the first road user state information and the vehicle state information; and determining, based on the correlation threshold, whether the first road user is a potential target of the vehicle for the first prediction time layer.

[0022] Status information from the first road user may include dynamic and / or static information about a single first road user among all road users, as described above.

[0023] According to an implementation, the road user includes a second road user. The step of determining state information from detected road users includes determining state information from the second road user. The method further includes: predicting a fourth trajectory of the vehicle for the first prediction time layer based on the second road user state information and the vehicle state information; performing a third similarity comparison between the fourth predicted trajectory of the vehicle based on the second road user state information and the vehicle state information and a second predicted trajectory of the vehicle based on the road user state information and the vehicle state information; and determining, based on the correlation threshold, whether the second road user is a potential target of the vehicle for the first prediction time layer.

[0024] Similarly, the status information from the second road user may include dynamic and / or static information about a single second road user among all road users, as described above.

[0025] The first road user is different from the second road user, and both are different from the vehicle.

[0026] Therefore, road users can be ranked based on their impact on vehicles.

[0027] According to the implementation method, the steps of the method are repeated for a second prediction time layer that is different from the first prediction time layer.

[0028] Specifically, the steps can be repeated for the second prediction time layer after all steps for the first prediction time layer are completed, i.e., sequentially. In another embodiment, the steps are performed in parallel for the first and second prediction time layers.

[0029] The second prediction time layer is another point in the future from which the prediction is performed; specifically, it is another point in the future from which the first prediction time layer is performed.

[0030] This allows for particularly robust target determination.

[0031] According to the implementation method, similarity comparison includes performing a distance metric.

[0032] According to the implementation method, the distance metric includes performing at least one of the Wasserstein algorithm, L1 algorithm, L2 algorithm, and Mahalanobis algorithm.

[0033] According to the implementation method, the prediction is performed by a machine learning algorithm. In particular, the prediction of the first trajectory, the second trajectory, the third trajectory, and / or the fourth trajectory can be performed using a machine learning algorithm.

[0034] Specifically, machine learning algorithms can be scene prediction algorithms based on the perception of surrounding objects. In particular, machine learning algorithms can be multimodal prediction algorithms based on machine learning. Furthermore, machine learning algorithms can use convolutional neural networks and / or recurrent neural networks, which can be used together to jointly learn and predict the movements of one or more road users.

[0035] Specifically, for a given scenario, the prediction of road users involves forecasting the future movement of one or more road users within a region of interest. The number of road users can vary depending on the scenario, necessitating the use of a shared data structure to store this variable number of road users. As input and output data structures, a series of fixed-size 2D grids, also known as images, are used, allowing the algorithm to jointly encode the trajectories of one or more road users within the region of interest and supporting the prediction of road users.

[0036] Prediction can be performed as described in Maximilian Schaefer et al.: "Context-Aware Scene Prediction Network (CASPNet)", https: / / arxiv.org / abs / 2201.06933, the entire contents of which are incorporated herein by reference.

[0037] Similarly, predictions can be made as described in the following literature. K. Zhao, M. Bühren and A. Kummert, "Context-Aware Scene Prediction Network (CASPNet)," 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), Macau, China, 2022, pp. 3970-3977, doi:10.1109 / ITSC55140.2022.9921850, the entire contents of which are incorporated herein by reference.

[0038] In another aspect, this disclosure relates to a computer system configured to perform some or all of the steps of the computer-implemented methods described herein.

[0039] A computer system may include a processing unit, at least one memory unit, and at least one non-transitory data memory. The non-transitory data memory and / or memory unit may include computer programs for instructing the computer to perform some or all of the steps or aspects of the computer-implemented methods described herein.

[0040] In another aspect, the present invention relates to a vehicle comprising the computer system described above.

[0041] In another aspect, this disclosure relates to a non-transitory computer-readable medium comprising instructions for performing several or all of the steps or aspects of the computer-implemented methods described herein. The computer-readable medium may be configured as: an optical medium, such as an optical disc (CD) or digital versatile disc (DVD); a magnetic medium, such as a hard disk drive (HDD); a solid-state drive (SSD); a read-only memory (ROM), such as flash memory; and so on. Furthermore, the computer-readable medium may be configured as data storage accessible via a data connection such as an Internet connection. The computer-readable medium may, for example, be an online database or cloud storage.

[0042] This disclosure also relates to a computer program for instructing a computer to perform some or all of the steps or aspects of the computer-implemented methods described herein. Attached Figure Description

[0043] This document describes exemplary embodiments and functions of the present disclosure in conjunction with the following schematically illustrated figures:

[0044] Figure 1 A system for target selection near a vehicle is shown.

[0045] Figure 2 A flowchart illustrating a method for target selection near a vehicle is shown, and

[0046] Figure 3 Exemplary results of a method for target selection near a vehicle are shown.

[0047] List of reference numerals

[0048] 10 processors

[0049] 20. Memory

[0050] 30-range sensor

[0051] 40 Image Sensors

[0052] 100 System

[0053] 200 methods

[0054] 210 Method and Steps

[0055] 220 Method and Steps

[0056] 230 Methods and Steps

[0057] 240 Methods and Steps

[0058] 250 Methods and Steps

[0059] 260 Methods and Steps

[0060] 265 Methods and Steps

[0061] 270 Methods and Steps

[0062] 280 Methods and Steps

[0063] 290 Methods and Steps

[0064] 295 Methods and Steps

[0065] 1000 vehicles

[0066] 1001 Intersection

[0067] 1010 First Road User

[0068] 1020 Second Road Users

[0069] 1030 Third Road Users

[0070] 1040 Fourth Road User

[0071] 1100 First Prediction Layer

[0072] 1200 Second Prediction Layer

[0073] 1300 Third Prediction Layer

[0074] 1400 Fourth Prediction Layer

[0075] 1500 Fifth Prediction Layer

[0076] 1600 Sixth Prediction Layer Detailed Implementation

[0077] Figure 1 A system 100 for target selection near a vehicle is depicted. The system 100 includes a processor 10, a memory 20, a range sensor 30, and an image sensor 40.

[0078] System 100 is adapted to obtain vehicle status information by means of processor 10. The vehicle status information includes dynamic and static information about the vehicle. Dynamic vehicle status information is obtained by means of range sensor 30, and static vehicle status information is obtained by the processor from data stored in memory 20.

[0079] System 100 is also adapted to execute machine learning algorithms by means of processor 10 to predict the first trajectory of the vehicle for a first prediction time layer based on vehicle state information.

[0080] System 100 is also adapted to detect road users near the vehicle by means of image sensor 40. System 100 is also adapted to determine state information from the detected road users who are different from the vehicle by means of processor 10. This state information includes dynamic and static information about the road users.

[0081] System 100 is also adapted to perform machine learning algorithms on the state information of road users and vehicle state information by means of processor 10 to predict the second trajectory of the vehicle for a first prediction time layer.

[0082] System 100 is also adapted to perform a first similarity comparison, by means of processor 10, of a first predicted trajectory of the vehicle based on vehicle state information and a second predicted trajectory of the vehicle based on road user state information and vehicle state information, to determine whether a road user is a potential target of the vehicle for a first prediction time layer.

[0083] The step of processor 10 performing a first similarity comparison between a first predicted trajectory of a vehicle based on vehicle state information and a second predicted trajectory of a vehicle based on road user state information and vehicle state information is used to determine the correlation threshold of the detected road user for the first prediction time layer.

[0084] The road user includes a first road user, and the step of determining the state information from the road user includes determining the state information from the first road user. The system 100 is further adapted to perform a machine learning algorithm on the state information of the first road user and the vehicle state information using a processor 10 to predict a third trajectory of the vehicle for a first prediction time layer, and to perform, using the processor 10, a second similarity comparison between the predicted trajectory of the vehicle based on the second predicted trajectory of the vehicle based on the road user state information and the vehicle state information, and the third predicted trajectory of the vehicle based on the first road user state information and the vehicle state information, to determine, based on the correlation threshold, whether the first road user is a potential target of the vehicle for the first prediction time layer.

[0085] Furthermore, the road users include second road users, wherein the step of determining state information from road users includes determining state information from second road users. The system 100 is also adapted to perform a machine learning algorithm on the state information of the second road users and vehicle state information using the processor 10 to predict a fourth trajectory of the vehicle for a first prediction time layer, and to perform, using the processor 10, a third similarity comparison between the fourth predicted trajectory of the vehicle based on the state information of the second road users and vehicle state information and the second predicted trajectory of the vehicle based on the state information of the road users and vehicle state information, to determine, based on a correlation threshold, whether the second road user is a potential target of the vehicle for the first prediction time layer.

[0086] System 100 is also adapted to determine whether the first road user or the second road user has a higher priority based on a second similarity comparison and a third similarity comparison.

[0087] System 100 is also adapted to repeat the aforementioned steps for a second prediction time layer that is different from the first prediction time layer by means of processor 10.

[0088] System 100 is also adapted to perform distance metrics, which may include at least one of the Wasserstein algorithm, L1 algorithm, L2 algorithm, and Mahalanobis algorithm.

[0089] In particular, the system 100 will now be described by way of example with respect to various features that can serve as the basis of the system 100:

[0090] A human driver's driving trajectory is influenced by their own dynamics and their surrounding environment, which consists of other road users (dynamic environment) and static environment or surrounding objects. Given the vehicle's past dynamic state x and the past dynamic states x1, x2, ..., xn of all N other road users... n Given static information c of the scene, the future trajectory y of the vehicle at the predicted time t is... t It can be defined as a conditional distribution:

[0091] P(y t |x,x1,x2,…,x n ,c)(1)

[0092] In particular, sensors cannot observe a driver's intention to go somewhere, such as a person's thoughts. For example, at an intersection, a driver may turn left, right, or go straight, and their trajectory may be influenced by their surroundings, but the most important factor is where the driver wants to go.

[0093] This factor is unobservable. Therefore, the predicted distribution y t It must be multimodal to cover multiple possible real-world future trajectories. Depending on where the driver might want to go, the destination can also be different.

[0094] Among them, based on a single road user x i The impact, the future trajectory of the vehicle, can be expressed as a conditional distribution P. i :

[0095] P i (y t |x,x i ,c),(2)

[0096] Similarly, the distribution P0 describes the distribution when y t The prediction does not take into account dynamic surroundings (no other road users):

[0097] P0(y t |x,c)(3)

[0098] Suppose there exists a distance function L that measures the distance d between two distributions. For example, between distribution (1) and distribution (2):

[0099] d i =L(P, P) i (4)

[0100] Furthermore, the distance to the predicted distribution can be calculated in both cases with and without dynamic surrounding elements:

[0101] d0=L(P,P0) (5)

[0102] Fundamentally, small distances imply similar predicted distributions, while large distances indicate different distributions.

[0103] Given the above formulas and definitions, the system 100 for target selection can be described as the following process:

[0104] 1. For each prediction time level,

[0105] a. Predicting distributions P and P' using machine learning algorithms. i i∈[1,2,…,N] and P0.

[0106] b. By calculating the distance d0 = L(P, P0(y) t |x,c)) is used to check whether dynamic surrounding things have an effect on the prediction.

[0107] c. Loop through all road users, for each road user x i Calculate its distance from distribution P:

[0108] d i =L(P,P i )

[0109] d. Maintain road users whose distance from distribution P is less than d0.

[0110] e. Road users at the smallest distance have the most significant impact on the driving of vehicles.

[0111] 2. Repeat the above steps for all forecast time periods to select the most relevant target (if any) for each forecast time period.

[0112] Therefore, a machine learning-based trajectory prediction system based on surrounding object perception was used to test which road user has the most significant impact on the vehicle's future trajectory by comparing trajectory predictions based on complete surrounding object perception with trajectory predictions based on only one road user as a dynamic surrounding object. Furthermore, the influence of road users (dynamic surrounding objects) was also tested.

[0113] Now refer to Figure 2 This will be described in more detail. Figure 2A flowchart of a method 200 for target selection near a vehicle is shown.

[0114] In step 210, method 200 obtains vehicle status information.

[0115] Then in step 220, a machine learning algorithm is executed to predict the first trajectory of the vehicle for the first prediction time layer based on the vehicle state information.

[0116] Prediction is made in step 230.

[0117] In step 240, status information from road users different from those of the vehicle is obtained.

[0118] In step 250, a machine learning algorithm is performed on the road user state information and vehicle state information to predict the second trajectory of the vehicle for the first prediction time layer.

[0119] Prediction is made in step 260.

[0120] Then, in step 265, a first similarity comparison is performed based on the first predicted trajectory of the vehicle output in step 230 and the second predicted trajectory of the vehicle output in step 260 to determine whether the road user is a potential target of the vehicle for the first prediction time layer.

[0121] In step 270, state information from only one road user (i.e., the first road user) is used as input along with vehicle state information. In step 280, a machine learning algorithm is executed on this information to predict the third trajectory of the vehicle for the first prediction time layer. The third trajectory is output in step 290.

[0122] Then, in step 295, a second similarity comparison is performed based on the second predicted trajectory of the vehicle output in step 260 and the third predicted trajectory of the vehicle output in step 290 to determine whether the road user is a potential target of the vehicle for the first prediction time layer.

[0123] This is done based on the relevance threshold determined earlier based on the first similarity comparison in step 265.

[0124] Then, based on the status information of another road user among the road users, namely the status information of the second, third and / or fourth road users, along with the vehicle status information, these last four steps 270, 280, 290 and 295 can be repeated to perform the third and / or fourth similarity comparison.

[0125] Similarly, for the second prediction time layer, the third time layer, the fourth time layer, etc., method 200 can be repeated sequentially or in parallel.

[0126] The aforementioned system 100 and method 200 provide a general target selection framework that works regardless of highway / city conditions, road structure, driving scenario, or the complexity of the scenario.

[0127] Specifically, the above implementation provides a data-driven approach that learns from real-world user driving, where no specific rules need to be explicitly defined, thus mimicking the user's real-world decision-making.

[0128] Goal determination involves a wide range of possible user driving behaviors, taking into account multimodalities. Therefore, goal determination is application-independent and relies on different prediction time layers to provide flexibility for different goal determination scenarios.

[0129] Figure 3 It shows how to combine Figure 2 Exemplary results of the method for target selection near vehicle 1000.

[0130] The example shown is a multi-lane intersection 1001 with a first prediction layer 1100 in the next 0.5s, a second prediction layer 1200 in the next 1.0s, a third prediction layer 1300 in the next 1.5s, a fourth prediction layer 1400 in the next 2.0s, a fifth prediction layer 1500 in the next 2.5s, and a sixth prediction layer 1600 in the next 3.0s.

[0131] In this model, the vehicle 1000, which performs the target selection method, is displayed at the center of each prediction layer. In the first prediction layer 1100, none of the other road users are potential targets for the vehicle. This is due to the shorter future predicted in the first prediction layer 1100.

[0132] In the second prediction layer 1200, among multiple road users, three road users have been identified as potential targets of vehicle 1000: first road user 1010, second road user 1020, and third road user 1030. Other road users not shown in the accompanying drawings have been identified as irrelevant or below a previously determined relevance threshold.

[0133] The smaller road user ID indicates higher relevance. As can be seen in the second prediction layer 1200, the method has identified the first road user 1010 as the most relevant, the second road user 1020 as less relevant, and the third road user 1030 as the least relevant.

[0134] However, as can be seen from the third prediction layer 1300, the first road user 1010 is no longer identified as relevant, while the second road user 1020 is now considered the most relevant road user, followed by the third road user 1030. This remains the same in the fourth prediction layer 1040.

[0135] As can be seen from the fifth prediction layer 1050, the second road user 1020 is now considered the only relevant road user. However, in the sixth prediction layer 1060, the third road user becomes relevant again, but with a much lower correlation.

Claims

1. A computer-implemented method for target selection near a vehicle, the method comprising the following steps: The processor obtains vehicle status information, which includes dynamic information about the vehicle. The processor predicts the first trajectory of the vehicle for the first prediction time layer based on the vehicle state information; The vehicle's sensors detect road users near the vehicle; Determine the status information from the detected road users, which includes dynamic information about the road users; The processor predicts the second trajectory of the vehicle for the first prediction time layer based on the vehicle state information and the road user state information. as well as The processor performs a first similarity comparison between the predicted first trajectory and the predicted second trajectory of the vehicle to determine whether the detected road user is a potential target of the vehicle for the first prediction time layer.

2. The method according to claim 1, in, The dynamic information includes information on at least one of the following: the vehicle's position, the vehicle's steering angle, the vehicle's throttle input, the vehicle's braking input, the vehicle's acceleration, the vehicle's speed, the vehicle's turning signal status, autonomous driving characteristic status, and / or the vehicle's navigation route.

3. The method according to claim 1 or 2, in, The vehicle status information also includes static information about the vehicle.

4. The method according to claim 1 or 2, in, The step of performing the first similarity comparison by the processor is used to determine the relevance threshold of the detected road users for the first prediction time layer.

5. The method according to claim 4, in, The status information from the road user includes status information from the first road user, and the method further includes: The processor predicts the third trajectory of the vehicle for the first prediction time layer based on the state information of the first road user and the vehicle state information; and The processor performs a second similarity comparison between the predicted second trajectory and the predicted third trajectory to determine, based on the correlation threshold, whether the first road user is a potential target of the vehicle for the first prediction time layer.

6. The method according to claim 5, in, The status information from the road user includes status information from the second road user, and the method further includes: The processor predicts the fourth trajectory of the vehicle for the first prediction time layer based on the state information of the second road user and the vehicle state information; and The processor performs a third similarity comparison between the predicted second trajectory and the predicted fourth trajectory to determine, based on the correlation threshold, whether the second road user is a potential target of the vehicle for the first prediction time layer.

7. The method according to claim 6, further comprising: The first road user or the second road user is given higher priority based on the second similarity comparison and the third similarity comparison.

8. The method according to claim 1 or 2, wherein, For a second prediction time layer that is different from the first prediction time layer, repeat the steps of the method.

9. The method according to claim 1 or 2, in, The similarity comparison includes performing a distance metric.

10. The method according to claim 9, in, The distance metric includes performing at least one of the Wasserstein algorithm, L1 algorithm, L2 algorithm, and Mahalanobis algorithm.

11. The method according to claim 1 or 2, in, The prediction is performed using a machine learning algorithm.

12. A computer system configured to perform a computer-implemented method according to at least one of claims 1 to 11.

13. A non-transitory computer-readable medium comprising instructions for performing a computer-implemented method according to at least one of claims 1 to 11.

14. A vehicle comprising the computer system of claim 12.

Images