A rule-based target lane prediction method under intersection scenarios

Abstract

The application discloses a rule-based target lane prediction method in an intersection scene, which comprises intention prediction, target exit prediction and target lane prediction. Specifically, the semantic behavior of a vehicle that has entered the intersection is predicted, and different behavior labels are given based on different semantic behaviors. Through the obtained semantic behavior, the target exit that the vehicle may go to is predicted. Through the semantic behavior and the predicted target exit, a reasonable target lane is calculated. The method fully utilizes map elements, selects a reasonable target lane based on geometric calculation, and can analyze the perception result according to human and environmental constraints to give a relatively stable and accurate intersection target lane, thereby refining the prediction problem in the intersection and improving the interpretability of the prediction problem. Through target exit prediction and target lane prediction, the solution space of the prediction problem is reduced, the lateral jump of the predicted trajectory is reduced, and the stability of vehicle driving is improved.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a rule-based method for predicting target lanes in intersection scenarios. Background Technology

[0002] The software algorithms of autonomous driving systems generally consist of several major modules: localization and mapping, perception, and planning and control. The perception module is responsible for understanding the surrounding environment, while the planning and control module is responsible for planning a safe and reliable driving route based on environmental information, with the control module then executing the necessary actions. This "perception-planning and control" model ensures the normal operation of autonomous vehicles, but it tends to struggle in complex road conditions, leading to a sharp increase in takeover and emergency braking. In recent years, with the industry's increasing demands for autonomous driving technology and user experience, more accurate and scene-understanding perception capabilities, along with more flexible, smooth, and timely risk-avoidance planning and control technologies, have become increasingly important goals for R&D personnel. To better analyze the output of the perception module and provide the planning and control module with timely interaction information with dynamic objects, the prediction module has gained increasing importance, becoming a crucial link in the autonomous driving algorithm system.

[0003] The main function and role of the autonomous driving prediction module is to analyze the current and historical information of traffic participants (vehicles, pedestrians, and non-motorized vehicles) output by upstream perception fusion, combine it with real-time high-precision maps and environmental information, analyze and predict the future behavior or future movement trajectory of all traffic participants near the autonomous vehicle, and combine it with the downstream planning module to provide it with future behavior patterns of traffic participants that include a variety of uncertainties, so that it can better interact with the environment.

[0004] Current predictions can be mainly divided into three types:

[0005] The first method is a prediction method based on simple kinematic reasoning. This method only uses the orientation angle, velocity and acceleration values ​​output by the perception fusion module. It uses a simple kinematic model, uniform velocity model or acceleration model to predict the motion trajectory that extends forward for 5 or 8 seconds along the current orientation angle. This method is relatively simple and can reflect the current motion state, but it lacks understanding and analysis of the environment.

[0006] The second approach combines the environmental features of high-precision maps with the semantic features of the perceived environment to analyze vehicle behavior in advance and provide a motion trajectory that conforms to physical characteristics under environmental constraints. This method has rich functionality and stable performance, but it requires specific development for specific scenarios.

[0007] The third approach combines deep learning models for data-driven prediction. This method has good generalization ability, but the initial data collection, model training and analysis, and subsequent deployment require significant time and manpower.

[0008] In addition, the second method can address problems in a variety of complex scenarios with strong interpretability and provide stable operating results. For example, vehicle lane changing and trajectory generation can provide specific results under human-defined constraints.

[0009] However, for scenarios within intersections, the prediction module's performance in predicting other vehicles is often not good enough. For example, whether a vehicle is turning or going straight, and which lane it will enter after turning, there are still some issues with the predicted trajectory changing or the target lane being predicted unreasonably. Summary of the Invention

[0010] To address the aforementioned problems, the main objective of this invention is to design a rule-based target lane prediction method for intersection scenarios, thereby resolving various issues affecting vehicle stability, such as abrupt changes in vehicle predicted trajectories and unreasonable target lane predictions.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A rule-based target lane prediction method in an intersection scenario includes: intent prediction, target exit prediction, and target lane prediction; wherein the target exit prediction depends on the result of intent prediction, and the target lane prediction depends on the result of target exit prediction.

[0013] The specific prediction steps are as follows:

[0014] Step 1: Intent prediction. Predict the semantic behavior of vehicles that have entered the intersection and assign different behavior labels to the vehicles based on different semantic behaviors.

[0015] Step 2: Target exit prediction. Based on the semantic behavior learned in Step 1, predict the target exit that the vehicle may go to.

[0016] Step 3: Target lane prediction. Based on the semantic behavior obtained in Step 1 and the target exit predicted in Step 2, a reasonable target lane is calculated.

[0017] The semantic behavior mentioned above refers to the vehicle's intention, which includes one of the following: going straight, turning left, turning right, or making a U-turn.

[0018] As a further description of the present invention, the intention prediction is implemented as follows: the vehicle enters the intersection through a global map, and the vehicle's intention and the virtual lane corresponding to the vehicle's intention are obtained. The vehicle's behavior label, i.e., the intention label, is filtered out through the rationality calculation of the intention.

[0019] As a further description of the present invention, the implementation of intent prediction includes the following steps:

[0020] Step 1.1: Obtain the function interface of the global map and confirm whether the vehicle has entered the intersection;

[0021] Step 1.2: After confirming that the vehicle has entered the intersection, query the multiple virtual lanes of the vehicle within the intersection according to the global map interface, obtain all possible and non-repeating intention sets within the intersection from the turning attributes of the virtual lanes, denoted as intention_set, and obtain a virtual lane corresponding to each intention, denoted as lane_set; the turning attributes of the virtual lanes are the turning intentions of the vehicle.

[0022] Step 1.3: Check the validity of each intent in the intent set;

[0023] Step 1.3.1: If the lane the vehicle entered at the intersection is known from the historical information, then the intentions that do not belong to the intention_set are filtered out according to the turning attribute of the lane to obtain the vehicle's intention and determine the rationality of the intention.

[0024] As a further description of the present invention, in step 1.3.1, if the lane the vehicle entered at the intersection cannot be determined from the historical information, the following steps are directly performed to determine the reasonableness of the intent:

[0025] Step 1.3.2: Obtain the vehicle's current information and all historical information. Select one frame from the historical information as a reference. Calculate the vehicle's position deviation angle using the position coordinates in the global map coordinate system under the selected historical information and the current information, and record it as move_heading. If the vehicle has no historical information, record the vehicle's current orientation angle as move_heading.

[0026] Step 1.3.3: Obtain the distance between the vehicle and the center line of each virtual lane in lane_set and the orientation of the nearest point, and calculate the orientation difference between move_heading and the orientation of the nearest point of the virtual lane, and sort the distance and orientation difference respectively;

[0027] Step 1.3.4: If there is only one intent corresponding to intention_set, then assign that intent to the vehicle;

[0028] If there are more than one intention in the intention_set, and the intention corresponding to the smallest orientation difference after sorting is the same as the intention corresponding to the shortest distance, then the vehicle is assigned that intention; otherwise, the orientation difference and distance corresponding to the same intention are normalized and the score is calculated. The scores are then sorted, and the intention corresponding to the highest score is taken as the vehicle's intention at the current moment.

[0029] As a further description of the present invention, the target exit prediction is implemented by: obtaining the target exit through the virtual lane lines within the intersection corresponding to the intent and based on the functional interface of the global map;

[0030] The specific steps are as follows:

[0031] Step 2.1: Based on the vehicle intent predicted by intent and the lane information when the vehicle enters the intersection, query and filter the preceding connecting lanes that meet the vehicle's entry requirements through the global map interface, and record the subsequent connecting exits of the corresponding preceding connecting lanes; each preceding connecting lane corresponds to several subsequent connecting exits.

[0032] The preceding connecting lane is the lane for vehicles to enter the intersection, and the subsequent connecting exit is the lane for vehicles to exit the intersection.

[0033] As a further description of the present invention, if the lane information of the vehicle entering the intersection cannot be obtained during the process of obtaining the target exit, the target exit is predicted through the following steps:

[0034] Step 2.2: Based on the vehicle's intent, locate the target exit within the intersection;

[0035] Step 2.2.1: Set tags according to the vehicle's intent. The tags include clockwise and counterclockwise queries.

[0036] Step 2.2.2: Locate all exit roads within the intersection where the vehicle is located;

[0037] According to the labels set in step 2.2.1, when turning left or making a U-turn, the counterclockwise query selects the starting point of the left side line of the exit; when going straight or turning right, the clockwise query selects the starting point of the right side line of the exit. Using the selected starting point as the initial point, calculate the angle and distance d between the current position of the vehicle and the initial point, and calculate the angle difference r between the vehicle's orientation and the angle of the connecting line.

[0038] Step 2.2.3: For all queried exits, record d and r respectively, sort all d and r respectively, and set the angle difference threshold range and storage quantity limit;

[0039] Step 2.2.4: Map the exit according to the intent;

[0040] Step 2.2.5: Intended to turn left or right:

[0041] Iterate through all sorted angle differences r. If the angle difference r is within the angle difference threshold range, then retain the exit corresponding to that angle difference r. The number of retained exits shall not exceed the storage limit.

[0042] Step 2.2.6: Intended to proceed straight:

[0043] Iterate through all sorted angle differences r. If the angle difference r is within the minimum value of the angle difference threshold range, select the exit corresponding to the minimum angle difference r.

[0044] Step 2.2.7: Intended to turn around:

[0045] Select the exit corresponding to the smallest distance d after sorting.

[0046] As a further description of the present invention, the target lane prediction is implemented as follows: based on the driving intention obtained in step 1 and the target exit selected in step 2, a reasonable target lane is selected by calculation.

[0047] For scenarios with high certainty, choose one lane; for scenarios with high uncertainty, choose two lanes.

[0048] As a further description of the present invention, the target lane prediction is implemented by the following steps:

[0049] Step 3.1: Intended action: left turn, right turn, U-turn.

[0050] For each lane in the target exit obtained in step 2, calculate the distance between the vehicle's current position and the starting point of the lane centerline;

[0051] Step 3.1.1: Select the lane closest to the vehicle. If the number of closest lanes is less than three, exit directly and designate the closest lane as the target lane.

[0052] Step 3.2: Intended to proceed straight:

[0053] For the starting point of each lane in the target exit obtained in step 2, transform to a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis, and calculate the abscissa of the starting point of the lane centerline in coordinate system T. The abscissa represents the lateral distance from the starting point of the lane to the vehicle.

[0054] Step 3.2.1: Select the lane with the closest lateral distance to the vehicle. The number of lanes with the closest lateral distance shall not exceed three. If there is only one lane, exit directly and identify the lane with the closest lateral distance as the target lane.

[0055] As a further description of the present invention, in step 3.1, if there are three or more nearest lanes, the target lane is predicted through the following steps if the intended turn is left, right, or U-turn:

[0056] Step 3.1.2: Arrange the selected nearest lanes from left to right according to the actual map, transform them to a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis, and calculate the vector v of the lane starting point in coordinate system T. Transform the unit vector of the vehicle orientation to coordinate system T and denote it as h. Calculate the cross product of each v and h.

[0057] Step 3.1.3: Select the target lane based on the comparison between v and h:

[0058] The selection method for vehicles to turn left or make a U-turn into the target lane is as follows:

[0059] If the vector h and the vector of the starting point of the leftmost lane have the same sign in coordinate system T, then the two lanes on the left are selected.

[0060] If vector h points between the leftmost and rightmost lanes, then select the two lanes closest to the vehicle.

[0061] If the abscissas of vector h and the starting points of the two left lanes in coordinate system T are opposite in direction, and the cross product of vector h and the starting point of the rightmost lane in coordinate system T is positive, then the two right lanes are selected.

[0062] The method for selecting the target lane when a vehicle turns right is as follows:

[0063] If the vector h and the vector of the starting point of the rightmost lane have the same sign in coordinate system T, then the two rightmost lanes are selected.

[0064] If vector h points between the leftmost and rightmost lanes, then select the two lanes closest to the vehicle.

[0065] If the abscissas of vector h and the starting points of the two right lanes in coordinate system T are in opposite directions, and the cross product of vector h and the starting point of the leftmost lane in coordinate system T is positive, then the two left lanes are selected.

[0066] As a further description of the present invention, in step 3.2.1, if there is more than one lane with the closest lateral distance, the intended straight-ahead target lane is predicted through the following steps:

[0067] Step 3.2.2: Arrange the selected lanes with the closest lateral distance from left to right according to the actual map, and recalculate the x-coordinates of the starting points of the left and right edges of each lane in coordinate system T. Keep the lanes with positive x-coordinates of the starting point of the left edge and negative x-coordinates of the starting point of the right edge as target lanes.

[0068] Step 3.2.3: Determine whether to retain one target lane:

[0069] Obtain the orientation of the lane to be retained in step 3.2.2, calculate the difference between the current orientation of the vehicle and the orientation of the lane to be retained, and if the orientation difference is less than the set value, and the vehicle is between the starting point of the lane edge line and the vehicle maintains the corresponding distance from the edge line laterally, then exit and retain only that lane as the target lane.

[0070] Step 3.2.4: Determine whether to acquire the second target lane:

[0071] If the difference between the vehicle's current orientation and the orientation of the reserved lane exceeds a set value, or if the horizontal coordinate of the starting point of the current reserved lane edge line is within 0.8m, or if the vehicle's orientation points to the outside of the starting point of the reserved lane edge line, then the lane whose orientation points to the outside of the reserved lane and is adjacent to the reserved lane will be designated as the second target lane.

[0072] Compared with the prior art, the technical advantages of the present invention are as follows:

[0073] This invention provides a rule-based target lane prediction method in intersection scenarios. By fully utilizing map elements through the output of high-precision maps and perception fusion, it selects reasonable target lanes based on geometric calculations. It can analyze perception results based on human and environmental constraints and provide relatively stable and accurate target lanes for intersections.

[0074] This prediction method refines the prediction problem within intersections, breaking down the general intention prediction-trajectory prediction into intention prediction-target exit prediction-target lane prediction-trajectory prediction, thereby improving the interpretability of the prediction problem.

[0075] By predicting the target exit and the target lane, the solution space of the prediction problem is narrowed, transforming the prediction problem in a wider space into a prediction problem in a limited space, reducing lateral jumps in the predicted trajectory, and improving the stability of vehicle driving. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the overall prediction method of the present invention;

[0077] Figure 2 This is a schematic diagram illustrating the intention prediction and target export prediction of the present invention;

[0078] Figure 3 This is a schematic diagram of the left-turn target lane prediction of the present invention;

[0079] Figure 4 This is a schematic diagram of the straight-ahead target lane prediction of the present invention. Implementation

[0080] The present invention will now be described in detail with reference to the accompanying drawings:

[0081] In the embodiments disclosed in this invention, a rule-based target lane prediction method in an intersection scenario is disclosed, with reference to... Figure 1-4 As shown, the prediction method includes: intent prediction, target exit prediction, and target lane prediction; the target exit prediction depends on the result of intent prediction, and the target lane prediction depends on the result of target exit prediction.

[0082] The specific prediction steps in this embodiment are as follows:

[0083] Step 1: Intent Prediction. Predict the semantic behavior of vehicles that have entered the intersection and assign different behavior labels to the vehicles based on different semantic behaviors; the semantic behavior is the vehicle's intent, including one of the following: go straight, turn left, turn right, or make a U-turn.

[0084] Step 2: Target exit prediction. Based on the semantic behavior learned in Step 1, predict the target exit that the vehicle may go to.

[0085] Step 3: Target lane prediction. Based on the semantic behavior obtained in Step 1 and the target exit predicted in Step 2, a reasonable target lane is calculated.

[0086] This embodiment analyzes the prediction of intent, target exit, and target lane based on the vehicle's intention to turn left or go straight. The analysis is as follows:

[0087] I. Intent Prediction

[0088] 1. Purpose of Intent Prediction

[0089] In order to predict the semantic behavior of vehicles that have entered the intersection, a certain behavior label is assigned to the predicted vehicle. This makes it easier for subsequent target exit prediction and target lane prediction to enter different prediction logics based on different label information. The semantic behavior can be located as four situations: going straight, turning left, turning right, and making a U-turn. In this embodiment, the semantic behavior is going straight and turning left.

[0090] 2. Implementation methods of intent prediction

[0091] The system confirms the vehicle's entry into the intersection using a global map, obtains the vehicle's intent and the corresponding virtual lane, and filters out the vehicle's behavior label, i.e., intent label, by calculating the reasonableness of the intent.

[0092] 3. Steps for implementing intent prediction

[0093] Step 1.1: Obtain the function interface of the global map and confirm whether the vehicle has entered the intersection;

[0094] If the vehicle has not entered the intersection, the logic of this embodiment is exited. If the vehicle has already entered the intersection, the following method is used to predict the vehicle's intention.

[0095] Step 1.2: After confirming that the vehicle has entered the intersection, query the 10 virtual lanes of the vehicle in the intersection according to the global map interface. The virtual lanes should be within eight meters of the vehicle, and the deviation between the orientation of the lane center line and the orientation of the nearest point of the vehicle and the vehicle's orientation should be within 30°.

[0096] In the turning attributes of the 10 virtual lanes, obtain the set of intentions within all possible and non-repeating intersections, denoted as intention_set, and obtain a virtual lane corresponding to each intention in the intention set, denoted as lane_set; the turning attributes of the virtual lanes are the turning intentions of the vehicles.

[0097] Step 1.3: Check the validity of each intent in the intent set;

[0098] Step 1.3.1: If the vehicle's historical information shows that the vehicle entered the intersection from a certain lane, then filter out intentions that do not belong to the intention_set based on the turning attribute of that lane, obtain the vehicle's intention, and determine the rationality of the intention.

[0099] In step 1.3.1, if the lane the vehicle entered at the intersection cannot be determined from historical information, the following steps are directly performed to determine the reasonableness of the intent:

[0100] Step 1.3.2: Obtain the vehicle's current information and all historical information. If the number of historical information records is more than 5 frames, select the historical information of the 5th frame in the past as a reference. If the number of historical information records is less than 5 frames, select the historical information of the last frame in the past as a reference. Use the selected historical information and the position coordinates in the global map coordinate system under the current information to calculate the deviation angle of the vehicle's position, which is recorded as move_heading. If the vehicle has no historical information, record the vehicle's current orientation angle as move_heading.

[0101] In this embodiment, taking the vehicle's current location coordinates (xc, yc) and historical location coordinates (xh, yh) as an example, then:

[0102] `move_hading = arctan((yc – yh) / (xc – xh))` calculates the orientation of the vector pointing from the historical position to the current position using the arctangent function.

[0103] Step 1.3.3: Obtain the distance between the vehicle and the center line of each virtual lane in lane_set, and the lane orientation of the nearest point on the center line of that virtual lane. Calculate the orientation difference between move_heading and the orientation of the nearest point in the virtual lane, and sort the distance and orientation difference respectively.

[0104] The above orientation difference can be expressed as heading_diff = absolute value(move_heading - orientation of the nearest point on the center line).

[0105] Step 1.3.4: If the smallest orientation difference after sorting is still greater than 15° and the shortest distance exceeds 1.5 meters, it means that the vehicle's intention to move is significantly different from that of all lane center lines. Therefore, the vehicle may not be driving according to the steering attributes marked by the virtual lane and needs to be re-predicted based on the actual movement of the vehicle.

[0106] If it is necessary to re-enter the prediction, clear the intention_set, select the current information of the vehicle and the information of the previous frame, calculate the position deviation angle, and map the deviation angle to the corresponding intention according to the set threshold. If the vehicle does not have historical information, the vehicle's intention is assumed to be to go straight, skip step 1, and proceed to step 2.

[0107] If re-entering the prediction is not required, the single intent of the vehicle is obtained by post-processing the intent_set and the corresponding lane_set. The post-processing method is as follows: if there is only one intent corresponding to the intent_set, the vehicle is assigned that intent, and step 1 is skipped to step 2; if there are more than one intent corresponding to the intent_set, the distance and orientation difference are sorted separately. If the intent corresponding to the smallest orientation difference after sorting is the same intent as the intent corresponding to the shortest distance, the vehicle is assigned that intent, and step 1 is skipped to step 2; otherwise, the orientation difference and distance corresponding to the same intent are normalized and the score is calculated. The shorter the distance and the smaller the orientation difference, the higher the score. The scores are sorted, and the intent corresponding to the highest score is taken as the vehicle's intent. Step 1 ends.

[0108] This embodiment specifically explains the calculation of scores after normalizing the heading difference and distance. Since the heading difference and distance obtained in step 1.3.3 are of different orders of magnitude, they need to be mapped to the same order of magnitude. If there are three heading differences (heading_diff), it means that there may be three lane center lines that match the vehicle's driving intention (three of them: straight, left turn, right turn, and U-turn), denoted as hd1, hd2, and hd3. We need to sum them up (sum = hd1 + hd2 + hd3) and calculate p1 = hd1 / sum, p2 = hd2 / sum, p3 = hd3 / sum, converting them into weights and completing the normalization. The larger the heading difference, the larger the weight. Then, we use exp(-p) to calculate the scores for the three weights respectively. Here, exp(-x) satisfies the condition that the smaller the weight, the higher the score. The distance score is calculated similarly. The scores are calculated as score = 0.5 * heading difference score + 0.5 * distance score. The lane center line with the highest score represents the vehicle's final intention.

[0109] Based on the aforementioned publicly disclosed intentions, and combined with, for example Figure 2 As shown, the rectangles in the intersection represent predicted vehicles, i.e., vehicles. Vehicles enter the intersection from lane 4 and can either turn left or go straight. The intent set selected around the vehicle includes both straight and left turns. The lane set consists of virtual lane R1 representing the intention to go straight and virtual lane R2 representing the intention to turn left. After calculation in step 1.3 above, the intention to turn left can be selected as the vehicle's intent label.

[0110] II. Target Export Forecast

[0111] 1. Purpose of Target Export Forecasting

[0112] To predict the exits of intersections that vehicles whose intentions are known can potentially reach, their driving areas are limited, environmental constraints are added, the prediction range is narrowed, and the accuracy and stability of predictions are improved.

[0113] 2. Methods for achieving target export forecasting

[0114] The target exit is obtained by using the virtual lane lines within the intersection corresponding to the intent and based on the function interface of the global map.

[0115] 3. Steps for achieving target export forecasting

[0116] Step 2.1: Based on the vehicle intent predicted by intent and the lane information when the vehicle enters the intersection, query and filter the preceding connecting lanes that meet the vehicle's entry requirements through the global map interface, and record the subsequent connecting exits of the corresponding preceding connecting lanes; each preceding connecting lane corresponds to several subsequent connecting exits; the preceding connecting lane is the lane for the vehicle to enter the intersection, and the subsequent connecting exit is the lane for the vehicle to exit the intersection.

[0117] If the lane information of the vehicle entering the intersection is unavailable during the process of obtaining the target exit, the target exit is predicted using the following steps:

[0118] Step 2.2: Based on the vehicle's intent, locate the appropriate target exit within the intersection;

[0119] Step 2.2.1: Set a tag based on whether the vehicle's intention is to turn left or make a U-turn. The tags include clockwise and counterclockwise queries.

[0120] Step 2.2.2: Locate all available road exits within the intersection where the vehicle is located;

[0121] Based on the labels set in step 2.2.1, if the query is counterclockwise (i.e., when turning left or making a U-turn), the starting point of the leftmost edge of the exit is selected; if the query is clockwise (i.e., when going straight or turning right), the starting point of the rightmost edge of the exit is selected. Using the selected starting point as the initial point, calculate the angle and distance d between the vehicle's current position and the initial point, and calculate the angle difference r between the vehicle's current orientation and the angle of the connecting line.

[0122] In this embodiment, the initial point is denoted as begin_point(x,y), and the current position of the vehicle is denoted as (xc,yc). The distance between the current position of the vehicle and the initial point is: d=sqrt((x-xc)^2 + (y-yc)^2), and the angle between the current position of the vehicle and the initial point is: pos_heading =arctan((y-yc) / (x-xc)). The current orientation of the vehicle is denoted as velocity heading, and the angle of the connecting line is pos heading. The angle difference between the current orientation of the vehicle and the angle of the connecting line is r=velocity heading – pos heading.

[0123] Step 2.2.3: For all queried exits, record d and r respectively, and sort all d and r respectively. Set the angle difference threshold range and the storage quantity limit; the angle difference threshold range is set to 10°~90°, and the storage quantity limit is set to 2.

[0124] Step 2.2.4: Map the exit according to the intention. If the intention is to turn left or right, proceed to step 2.2.5; if the intention is to go straight, proceed to step 2.2.6; if the intention is to make a U-turn, proceed to step 2.2.7.

[0125] Step 2.2.5: Intended to turn left or right:

[0126] Iterate through all sorted angle differences r. If an angle difference r is within the angle difference threshold range of 10° to 90°, then retain the exit corresponding to that angle difference r. The number of retained exits shall not exceed the upper limit of the storage quantity. If the storage quantity reaches the upper limit or the iteration ends, exit step 2.

[0127] Step 2.2.6: Intended to proceed straight:

[0128] Iterate through all sorted angle differences r. For angle differences r that are within the angle difference threshold range of 10°, select the exit corresponding to the smallest angle difference r and exit step 2.

[0129] Step 2.2.7: Intended to turn around:

[0130] Select the exit corresponding to the smallest distance d after sorting, and exit step 2.

[0131] Based on the aforementioned publicly available target export forecasts, and combined with, for example Figure 2 As shown, if the vehicle to be predicted intends to turn left and exit from lane 4, a series of virtual lanes with the turning attribute of left turn and connected to lane 4 can be filtered out. Taking R1 as an example, R1 is connected to lane 4 in the preceding order and lane 1 in the subsequent exit order. Therefore, the exit E2 where lane 1 is located can be used as one of the target exits of the vehicle. If it is known that the vehicle to be predicted intends to turn left but it is unknown that it will exit from lane 4, then the target exits of the vehicle can be filtered out through the above step 2.2.

[0132] III. Target Lane Prediction

[0133] 1. Purpose of target lane prediction

[0134] To predict the target lane that a vehicle with a known target exit might go to, and to limit its driving area, for scenarios with high certainty, such as when it is facing a certain lane, one lane is selected; for scenarios with high uncertainty, such as when turning or crossing the line, two lanes are selected, so as to cover risks as much as possible.

[0135] 2. Implementation methods of target lane prediction

[0136] Based on the driving intention obtained in step 1 and the target exit selected in step 2, a reasonable target lane is selected through calculation.

[0137] 3. Steps for implementing target lane prediction

[0138] Step 3.1: Intended action: left turn, right turn, U-turn.

[0139] For each lane in the target exit obtained in step 2, the distance between the vehicle's current position and the starting point of the lane centerline is calculated.

[0140] Step 3.1.1: Select the three lanes closest to the vehicle. If the number of closest lanes is less than three, exit Step 3 directly and identify the closest lane as the target lane.

[0141] In step 3.1, if there are three or more nearest lanes, the target lane is predicted through the following steps. This embodiment uses three selected lanes as an example.

[0142] Step 3.1.2: Arrange the three closest lanes selected from left to right according to the actual map, connect the starting point of the lane centerline with the vehicle position to form a vector, establish a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis, and transform all connected vectors to coordinate system T, and calculate the vector v of the lane starting point in the vehicle coordinate system, transform the unit vector of the vehicle orientation to coordinate system T, denoted as h, and calculate the cross product of each v and h respectively;

[0143] It should be noted that the vector connecting the starting point of the lane and the current position of the vehicle is denoted as a, represented by (ax, ay). The current predicted vehicle orientation is represented by coordinates b (cos(heading), sin(heading)), where heading represents the current vehicle orientation. The lane orientation, i.e., the y-axis orientation in the T coordinate system, is denoted as reference heading. The transformation angle t = reference heading – 90 degrees rotates a to the T coordinate system, then v(vx, vy) where vx = ax * cos(t) + ay * sin(t), vy = -ax * sin(t) + ay * cos(t); similarly, h is the vector b rotated by an angle t to the T coordinate system; the outer product is calculated as h × v = hx * vy - hy * vx, following the right-hand rule, where the vector representing the vehicle orientation h comes first, indicating that h rotates towards v. If the outer product is positive, h rotates counterclockwise towards v, otherwise clockwise.

[0144] Step 3.1.3: Select the target lane based on the comparison between v and h, such as... Figure 3 As shown, the selection method for vehicles to turn left or make a U-turn into the target lane is as follows:

[0145] If the vector h and the vector v1 of the starting point of the leftmost lane in coordinate system T have the same x-coordinate (same sign), then the two lanes on the left are selected.

[0146] If vector h points between the leftmost lane v1 and the rightmost lane v3, then select the two lanes closest to the vehicle.

[0147] If the abscissas of vector h and vectors v1 and v2, the starting points of the two lanes on the left, are opposite in direction in coordinate system T, and the cross product of vector h and vector v3, the starting point of the rightmost lane, is positive in coordinate system T, then the two rightmost lanes are selected.

[0148] Similarly, based on the above method for selecting the target lane for a vehicle turning left or making a U-turn, the method for selecting the target lane for a vehicle turning right is as follows:

[0149] If the vector h and the vector of the starting point of the rightmost lane in coordinate system T have the same sign, then the two rightmost lanes are selected.

[0150] If vector h points between the leftmost and rightmost lanes, then select the two lanes closest to the vehicle.

[0151] If the abscissas of vector h and the starting points of the two right lanes in coordinate system T are in opposite directions, and the cross product of vector h and the starting point of the leftmost lane in coordinate system T is positive, then the two left lanes are selected.

[0152] It should also be noted that in this embodiment, if the vehicle does not meet the above conditions during the process of turning left, turning right, or making a U-turn, the two lanes closest to the vehicle are selected as the target lanes.

[0153] Step 3.2: Intended to proceed straight:

[0154] For each lane in the target exit road obtained in step 2, connect the starting point of the lane edge line with the vehicle position to form a vector. Establish a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis. Transform all connected vectors to coordinate system T, such as... Figure 4 As shown, calculate the abscissa of the starting point of the lane centerline in coordinate system T. The abscissa represents the lateral distance from the starting point of the lane to the vehicle.

[0155] The lateral distance of the above vehicles is calculated as follows: Connect the starting point of the lane with the current position of the vehicle, denoted as vector a, represented by (ax,ay). The lane orientation, i.e. the orientation of the y-axis of the T coordinate system, is denoted as reference heading. The transformation angle t = reference heading – 90 degrees is used to rotate a to the T coordinate system. Then v(vx,vy) is the coordinate of the starting point of the lane centerline in the T coordinate system after the transformation, where vx = ax*cos(t) + ay*sin(t), vy = -ax*sin(t) + ay*cos(t).

[0156] Step 3.2.1: Select the lane with the closest lateral distance to the vehicle. The number of lanes with the closest lateral distance shall not exceed three. If there is only one lane, exit directly and identify the lane with the closest lateral distance as the target lane.

[0157] In step 3.2.1, if there is more than one lane with the closest lateral distance, the intended straight-ahead target lane is predicted through the following steps:

[0158] Step 3.2.2: Arrange the selected lanes with the closest lateral distance from left to right according to the actual map. Calculate the x-coordinates of the starting points of the left and right sides of each lane in coordinate system T. The calculated x-coordinates can be regarded as virtual extension lines of each lane at the road exit. If the x-coordinate of the starting point of the left side of a lane in coordinate system T is negative and the x-coordinate of the starting point of the right side is positive, then retain that lane, assuming that the vehicle is currently facing that lane and considers it one of the possible target lanes.

[0159] Step 3.2.3: Determine whether to retain one target lane:

[0160] Obtain the orientation of the lane pl to be retained in step 3.2.2, calculate the orientation difference between the current orientation of the vehicle and the orientation of the lane to be retained (heading_diff), if the orientation difference (heading_diff) is less than the set value (usually 7°), and the vehicle is pointing between the starting points of the left and right edge lines of the lane, and the vehicle maintains the corresponding distance from the left and right edge lines laterally, then exit step 3 and retain only that lane.

[0161] For ease of calculation, the lane heading of the starting point of the lane line of the reserved lane is selected, and the difference between the predicted current vehicle heading and the velocity heading is taken as heading_diff.

[0162] Step 3.2.4: Determine whether to acquire the second target lane:

[0163] If the difference between the vehicle's current orientation and the orientation of the reserved lane (heading_diff) exceeds the set value of 7° and points to the left, or if the horizontal coordinate of the starting point of the left line of the current reserved lane (pl) is within 0.8m (meaning the vehicle is laterally closer to the left line of the reserved lane (pl), or if the vehicle's orientation points to the left of the starting point of the left line of the reserved lane (pl), then the lane immediately adjacent to the left of the left side of the reserved lane (pl) will be selected as the second target lane. If there is no lane on the left, then step 3 will be exited directly.

[0164] Specifically, the vehicle body is within 0.8 meters of the left lane line. Here, because the starting point (x, y) of the left lane line is transformed to the T coordinate system, the transformed coordinates are (xt, yt). Since the T coordinate system takes the vehicle itself as the origin, the x-coordinate of the transformed coordinates of the starting point of the left lane line can be used directly; here, the absolute value (xt) < 0.8.

[0165] It should also be noted that in this embodiment, the intention is to select the right lane of the vehicle while traveling straight, which is the same as the implementation method described above.

[0166] Based on the publicly available target lane predictions mentioned above, and combined with, for example Figure 2 , 3 As shown, the intention prediction and target exit prediction results are extracted and considered as a vehicle turning left from lane 4. The target exit E2 consists of lanes 1, 2, and 3. A coordinate system T is established with the average lane orientation as the y-axis and the vehicle itself as the origin. h is the unit vector of the vehicle's orientation in coordinate system T. From step 3.1.1, we can obtain the three lanes closest to the vehicle as 1, 2, and 3. From step 3.1.2, we calculate the line vectors v1, v2, and v3 connecting the vehicle to the starting points of the three lanes in coordinate system T, as well as the unit vector h of the vehicle's orientation. In the current scenario, as we know from step 3.1.3, h is in the same direction as v1, so lanes 1 and 2 are selected at the current moment. If h points between v1 and v3, then two lanes are selected based on the shortest distance. If the vehicle has already crossed lane 1 in the x-direction, that is, h and v1 are in opposite directions and h points outside v3, then lanes 2 and 3 are selected.

[0167] In addition, based on the aforementioned publicly available target lane prediction, and combined with, for example... Figure 2 , 4 As shown, the vehicle travels straight towards the target exit, which consists of lanes 1, 2, and 3 on the right. A coordinate system T is established with the lane orientation as the y-axis and the vehicle itself as the origin. h is the unit vector of the vehicle's orientation in coordinate system T. Lane 1 is the lane that the vehicle is currently facing, and it is selected as one of the target lanes. At this time, h points to the outside of the line connecting the vehicle and the right side line of lane 1, so it is assumed that the vehicle has a tendency to veer to the right. Therefore, lane 2, which is adjacent to lane 1 on the right, is selected as the second target lane.

[0168] The prediction of the target lane is based on the predicted intent, which is then converted into a road exit at the actual physical level, providing a priori conditions for intent prediction in this embodiment. That is, if the lane L where the vehicle enters the intersection can be obtained, the global map interface can be queried to retain all virtual lanes with lane L as the preceding connected lane in the intersection. Among the retained lanes, lane driving attributes that are different from the vehicle's intent are filtered out, and all virtual lanes with driving attributes that are the same as the intent are retained. Finally, the exit information of the subsequent connected lanes of the retained lanes is selected.

[0169] The advantage of this invention over the prior art is that, through the output of high-precision maps and perception fusion, it makes full use of map elements, selects reasonable target lanes based on geometric calculations, and can analyze perception results based on human and environmental constraints to provide relatively stable and accurate target lanes at intersections.

[0170] The prediction method disclosed in this invention refines the prediction problem within intersections, breaking down the general intention prediction-trajectory prediction into intention prediction-target exit prediction-target lane prediction-trajectory prediction, thereby improving the interpretability of the prediction problem.

[0171] This invention narrows the solution space of the prediction problem by predicting the target exit and the target lane, transforming the prediction problem in a wider space into a prediction problem in a limited space, reducing lateral jumps in the predicted trajectory, and improving the stability of vehicle driving.

[0172] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A rule-based method for predicting target lanes in an intersection scenario, characterized in that, include: Intent prediction, target exit prediction, target lane prediction; the target exit prediction depends on the result of intent prediction, and the target lane prediction depends on the result of target exit prediction. The specific prediction steps are as follows: Step 1: Intent prediction. Predict the semantic behavior of vehicles that have entered the intersection and assign different behavior labels to the vehicles based on different semantic behaviors. The intention prediction is implemented as follows: the vehicle enters the intersection through the global map, and the vehicle's intention and the virtual lane corresponding to the vehicle's intention are obtained. The vehicle's behavior label, i.e., the intention label, is filtered out through the rationality calculation of the intention. The implementation of intent prediction includes the following steps: Step 1.1: Obtain the function interface of the global map and confirm whether the vehicle has entered the intersection; Step 1.2: After confirming that the vehicle has entered the intersection, query the multiple virtual lanes of the vehicle within the intersection according to the global map interface, obtain all possible and non-repeating intention sets within the intersection from the turning attributes of the virtual lanes, denoted as intention_set, and obtain a virtual lane corresponding to each intention, denoted as lane_set; the turning attributes of the virtual lanes are the turning intentions of the vehicle. Step 1.3: Check the validity of each intent in the intent set; Step 1.3.1: If the lane the vehicle entered at the intersection is known from the historical information, then the intentions that do not belong to the intention_set are filtered out according to the turning attribute of the lane to obtain the vehicle's intention and determine the rationality of the intention. Step 2: Target exit prediction. Based on the semantic behavior learned in Step 1, predict the target exit that the vehicle may go to. Step 3: Target lane prediction. Based on the semantic behavior obtained in Step 1 and the target exit predicted in Step 2, a reasonable target lane is calculated. The semantic behavior mentioned above refers to the vehicle's intention, which includes one of the following: going straight, turning left, turning right, or making a U-turn.

2. The method for predicting target lanes based on rules in an intersection scenario according to claim 1, characterized in that: In step 1.3.1, if the lane the vehicle entered at the intersection cannot be determined from historical information, the following steps are directly performed to determine the reasonableness of the intent: Step 1.3.2: Obtain the vehicle's current information and all historical information. Select one frame from the historical information as a reference. Calculate the vehicle's position deviation angle using the position coordinates in the global map coordinate system under the selected historical information and the current information, and record it as move_heading. If the vehicle has no historical information, record the vehicle's current orientation angle as move_heading. Step 1.3.3: Obtain the distance between the vehicle and the center line of each virtual lane in lane_set and the orientation of the nearest point, and calculate the orientation difference between move_heading and the orientation of the nearest point of the virtual lane, and sort the distance and orientation difference respectively; Step 1.3.4: If there is only one intent corresponding to intention_set, then assign that intent to the vehicle; If there are more than one intention in the intention_set, and the intention corresponding to the smallest orientation difference after sorting is the same as the intention corresponding to the shortest distance, then assign that intention to the vehicle. Otherwise, the orientation difference and distance corresponding to the same intention are normalized and the score is calculated. The scores are then sorted, and the intention corresponding to the highest score is taken as the vehicle's intention at the current moment.

3. The method for predicting target lanes based on rules in an intersection scenario according to claim 2, characterized in that: The target exit prediction is achieved by obtaining the target exit through virtual lane lines within the intersection corresponding to the intent and based on the functional interface of the global map. The specific steps are as follows: Step 2.1: Based on the vehicle intent predicted by intent and the lane information when the vehicle enters the intersection, query and filter the preceding connecting lanes that meet the vehicle's entry requirements through the global map interface, and record the subsequent connecting exits of the corresponding preceding connecting lanes; each preceding connecting lane corresponds to several subsequent connecting exits. The preceding connecting lane is the lane for vehicles to enter the intersection, and the subsequent connecting exit is the lane for vehicles to exit the intersection.

4. The method for predicting target lanes based on rules in an intersection scenario according to claim 3, characterized in that: If the lane information of the vehicle entering the intersection is unavailable during the process of obtaining the target exit, the target exit is predicted using the following steps: Step 2.2: Based on the vehicle's intent, locate the target exit within the intersection; Step 2.2.1: Set tags according to the vehicle's intent. The tags include clockwise and counterclockwise queries. Step 2.2.2: Locate all exit roads within the intersection where the vehicle is located; According to the labels set in step 2.2.1, when turning left or making a U-turn, the counterclockwise query selects the starting point of the left side line of the exit; when going straight or turning right, the clockwise query selects the starting point of the right side line of the exit. Using the selected starting point as the initial point, calculate the angle and distance d between the current position of the vehicle and the initial point, and calculate the angle difference r between the vehicle's orientation and the angle of the connecting line. Step 2.2.3: For all queried exits, record d and r respectively, sort all d and r respectively, and set the angle difference threshold range and storage quantity limit; Step 2.2.4: Map the exit according to the intent; Step 2.2.5: Intended to turn left or right: Iterate through all sorted angle differences r. If the angle difference r is within the angle difference threshold range, then retain the exit corresponding to that angle difference r. The number of retained exits shall not exceed the storage limit. Step 2.2.6: Intended to proceed straight: Iterate through all sorted angle differences r. If the angle difference r is within the minimum value of the angle difference threshold range, select the exit corresponding to the minimum angle difference r. Step 2.2.7: Intended to turn around: Select the exit corresponding to the smallest distance d after sorting.

5. The method for predicting target lanes based on rules in an intersection scenario according to claim 4, characterized in that: The target lane prediction is achieved by: based on the driving intention obtained in step 1 and the target exit selected in step 2, a reasonable target lane is selected by calculation. For scenarios with high certainty, choose one lane; for scenarios with high uncertainty, choose two lanes.

6. The method for predicting target lanes based on rules in an intersection scenario according to claim 5, characterized in that: The target lane prediction is implemented through the following steps: Step 3.1: Intended action: left turn, right turn, U-turn. For each lane in the target exit obtained in step 2, calculate the distance between the vehicle's current position and the starting point of the lane centerline; Step 3.1.1: Select the lane closest to the vehicle. If the number of closest lanes is less than three, exit directly and designate the closest lane as the target lane. Step 3.2: Intended to proceed straight: For the starting point of each lane in the target exit obtained in step 2, transform to a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis, and calculate the abscissa of the starting point of the lane centerline in coordinate system T. The abscissa represents the lateral distance from the starting point of the lane to the vehicle. Step 3.2.1: Select the lane with the closest lateral distance to the vehicle. The number of lanes with the closest lateral distance shall not exceed three. If there is only one lane, exit directly and identify the lane with the closest lateral distance as the target lane.

7. The method for predicting target lanes based on rules in an intersection scenario according to claim 6, characterized in that: In step 3.1, if there are three or more nearest lanes, the target lane is predicted according to the following steps if the intended turn is left, right, or U-turn: Step 3.1.2: Arrange the selected nearest lanes from left to right according to the actual map, transform them to a coordinate system T with the vehicle as the origin and the average lane orientation as the y-axis, and calculate the vector v of the lane starting point in coordinate system T. Transform the unit vector of the vehicle orientation to coordinate system T and denote it as h. Calculate the cross product of each v and h. Step 3.1.3: Select the target lane based on the comparison between v and h: The selection method for vehicles to turn left or make a U-turn into the target lane is as follows: If the vector h and the vector of the starting point of the leftmost lane have the same sign in coordinate system T, then the two lanes on the left are selected. If vector h points between the leftmost and rightmost lanes, then select the two lanes closest to the vehicle. If the abscissas of vector h and the starting points of the two left lanes in coordinate system T are opposite in direction, and the cross product of vector h and the starting point of the rightmost lane in coordinate system T is positive, then the two right lanes are selected. The method for selecting the target lane when a vehicle turns right is as follows: If the vector h and the vector of the starting point of the rightmost lane have the same sign in coordinate system T, then the two rightmost lanes are selected. If vector h points between the leftmost and rightmost lanes, then select the two lanes closest to the vehicle. If the abscissas of vector h and the starting points of the two right lanes in coordinate system T are in opposite directions, and the cross product of vector h and the starting point of the leftmost lane in coordinate system T is positive, then the two left lanes are selected.

8. The method for predicting target lanes based on rules in an intersection scenario according to claim 7, characterized in that: In step 3.2.1, if there is more than one lane with the closest lateral distance, the intended straight-ahead target lane is predicted through the following steps: Step 3.2.2: Arrange the selected lanes with the closest lateral distance from left to right according to the actual map, and recalculate the x-coordinates of the starting points of the left and right edges of each lane in coordinate system T. Keep the lanes with positive x-coordinates of the starting point of the left edge and negative x-coordinates of the starting point of the right edge as target lanes. Step 3.2.3: Determine whether to retain one target lane: Obtain the orientation of the lane to be retained in step 3.2.2, calculate the difference between the current orientation of the vehicle and the orientation of the lane to be retained, and if the orientation difference is less than the set value, and the vehicle is between the starting point of the lane edge line and the vehicle maintains the corresponding distance from the edge line laterally, then exit and retain only that lane as the target lane. Step 3.2.4: Determine whether to acquire the second target lane: If the difference between the vehicle's current orientation and the orientation of the reserved lane exceeds a set value, or if the horizontal coordinate of the starting point of the current reserved lane edge line is within 0.8m, or if the vehicle's orientation points to the outside of the starting point of the reserved lane edge line, then the lane whose orientation points to the outside of the reserved lane and is adjacent to the reserved lane will be designated as the second target lane.

Images