A method for autonomous driving decision-making and planning based on optimizing multiple tasks, electronic equipment and computer storage medium

Abstract

The present invention proposes an automatic driving decision planning method, electronic device and computer storage medium based on optimizing multi-tasks, the method comprising: obtaining upstream input data, performing security checks on the upstream input data, and converting the data structure of the upstream input data; storing the upstream input data and the upstream input data after the data structure conversion locally or in the cloud; extracting candidate lane sets from high-precision map data and perception data, providing traffic semantic information, and self-checking the high-precision map data and perception data through tool methods, and cognitively grading the high-precision map data and perception data based on data type and quality. Advantages of the present invention: better scalability of decision scenarios, better scalability and decision performance of traffic rule decisions, efficient lane change decisions and strong experience, efficient intersection decisions and strong experience, and high efficiency of decision planning algorithms.

Description

Technical Field

[0001] The present invention relates to the field of vehicle autonomous driving technology, and in particular to an autonomous driving decision-making planning technology based on optimizing multiple tasks. Background Art

[0002] Autonomous driving vehicles are comprehensive intelligent systems that integrate multiple functions such as environmental perception, planning and decision-making, behavior control and execution, covering multiple disciplines such as mechanics, control, sensor technology, signal processing, pattern recognition, artificial intelligence and computer technology. The development of intelligent vehicles with autonomous driving capabilities is of great practical significance for the development of active safety assisted driving products with independent intellectual property rights in my country, improving the intelligence level of my country's independent brand cars, improving road traffic safety, and developing intelligent transportation systems. In autonomous driving vehicles, decision planning is a key technology and a hot and difficult issue in current research. It connects perception, prediction and control and is an important part of intelligence. Therefore, the study of decision planning algorithms is of great significance to improving the performance of autonomous driving systems.

[0003] The current decision planning algorithms mainly include:

[0004] First, decision-making planning based on rule state machines and geometric curves. Combining the current state of the vehicle, surrounding traffic participants and traffic safety conditions, the corresponding decision state machine and its state switching rules are designed, and the possible planning trajectory is obtained using geometric curves.

[0005] Second, end-to-end decision-making planning based on deep learning. The input is camera images, lidar point clouds, positioning, high-precision maps and other data. The model is trained through a large amount of data, and the output is vehicle control instructions.

[0006] Third, decision-making planning based on deep reinforcement learning. In a simulated environment, the reward function is designed, and the policy network and value network are trained through continuous trial and error, so that the intelligent vehicle can obtain better cumulative rewards after making sequential decisions, presenting better intelligent decision-making results.

[0007] The shortcomings of decision-making planning methods based on rule state machines and geometric curves: In the physical world, the scenarios are complex and it may be difficult to enumerate all the scenarios, which requires a good abstract ability for the scenarios; for geometric planning, certain rules need to be designed, and it may not be easy to integrate complex factors to obtain the optimal or approximately optimal solution.

[0008] The shortcomings of the end-to-end decision planning method based on deep learning: the deep learning model is a black box, which is not easy to find and correct when there are problems with the decision results; the deep learning model is a good fitting method, but may lack certain reasoning capabilities; the deep learning model requires a large amount of data, has certain requirements for data collection in complex interactive environments, and has high requirements for data quality; the computing platform operation overhead has certain challenges.

[0009] The shortcomings of the end-to-end decision-making and planning method based on deep reinforcement learning: there are certain requirements for the simulation quality of the simulation environment; there is a gap in the migration of intelligent agents from the simulation environment to the real scene; the end-to-end model has a certain degree of unexplainability and requires certain safety measures; the computing platform operation overhead has certain challenges.

[0010] The Chinese patent document with the publication number CN113879339A discloses a technology named "A decision-making planning method, electronic device and computer storage medium for autonomous driving". The technology involves a method including: obtaining driving perception information of the object to be decided in a continuous behavior space, wherein the driving perception information includes: geometric information related to the object to be decided, historical driving trajectory information and map information; according to the driving perception information and driving target information, obtaining multiple planning strategies that conform to the Gaussian distribution and the strategy evaluation corresponding to each planning strategy; according to the multiple planning strategies and the strategy evaluation corresponding to each planning strategy, making decision planning for the object to be decided. This technology can effectively make decision planning for strong interactive scenarios in autonomous driving and improve the decision effect. However, this technology does not involve the detailed application of reference lines based on the Frenet coordinate system, that is, it does not involve similar means, and the processing of acquisition is also different. The technology is more complicated and costly to process data. In addition, this technology does not provide a specific path and speed planning method, nor does it provide a method for calculating the optimal trajectory.

[0011] The Chinese patent document with the publication number CN113830108A discloses a technology named "a decision-making planning method and device for an autonomous driving vehicle", which involves the following methods: obtaining parameter information of traffic participants, obtaining the potential field of traffic participants according to the parameter information of traffic participants; constructing a decision sequence tree, which includes multiple decision sequences; constructing a corresponding fusion potential field for each possible domain in the decision sequence according to the potential field of traffic participants; calculating the planning trajectory corresponding to all decision sequences in the time domain according to the fusion potential field and the preset system state equation; evaluating the planning trajectory corresponding to all decision sequences to obtain multiple evaluation results; sorting the multiple evaluation results to obtain the sorted evaluation results; performing collision detection on the sorted evaluation results to obtain the detection results; and making long-term decision planning, which is easy to expand and has high stability. This technology only obtains the parameter information of traffic participants, and does not involve the specific application of the reference line based on the Frenet coordinate system, and does not involve the process of speed decision, scene decision, path, and speed planning.

[0012] A Chinese patent document with publication number CN110103987A discloses a technology named "A decision-making planning method and device for autonomous driving vehicles". The method involved in this technology includes: determining trigger information, wherein the trigger information includes at least one of a target decision time and a target vehicle speed, and the target decision time and the target vehicle speed are respectively determined according to the duration and vehicle speed of the autonomous driving vehicle performing decision-making planning operations within a recent preset time period; determining a trigger frequency based on the trigger information; and performing decision-making planning operations based on the trigger frequency. The golden fox improves the adaptability of the trigger frequency to different vehicle speed scenarios and different processing devices. This technology does not involve operations on acquiring data, does not involve the specific use of reference lines based on the Frenet coordinate system, and does not involve the process of speed decision, scene decision, path, and speed planning, and does not provide a specific decision-making planning method.

[0013] The Chinese patent document with the publication number CN114580302A discloses a technology named "A method for decision-making and planning of autonomous driving vehicles based on maximum entropy reinforcement learning". The method involved in the technology includes: constructing a decision-making and planning model based on maximum entropy reinforcement learning, which includes: state space, action space, reward function, strategy function and evaluation model; constructing an interactive high-speed highway simulation training scene: using a two-degree-of-freedom vehicle kinematic model to describe the movement of vehicles in the training scene, and using a rule-based intelligent decision-making and planning model to control environmental vehicles, so that the environment has interactive characteristics; training a decision-making and planning model based on maximum entropy reinforcement learning. This technology uses maximum entropy reinforcement learning to improve the optimality and stability of the high-speed highway decision-making and planning strategy of autonomous driving vehicles. The technical means of this technology do not involve the specific application of reference lines based on the Frenet coordinate system, and do not involve the process of speed decision, scene decision, path planning, and speed planning, and do not provide a specific decision-making and planning method. Summary of the invention

[0014] One of the purposes of the present invention is to provide an automatic driving decision-making and planning method based on optimizing multiple tasks, and the technical problems solved are: the decision-making and planning algorithms in the prior art have high requirements on computing power and low operating efficiency, and at the same time, the acquired data is not processed to facilitate the processing of the algorithm model; the candidate lane set is not verified, resulting in failure of the subsequent algorithm when executing the task, resulting in a safety accident of the vehicle; there is no cognitive classification of high-precision map data and perception data, which makes the algorithm's operating efficiency low; the traditional decision-making and planning algorithm does not involve the specific application of reference lines based on the Frenet coordinate system, does not involve the process of path decision planning and speed decision planning, and does not provide a method for calculating the optimal trajectory; the decision-making scenario scalability of the traditional decision-making and planning method is poor, the scalability of traffic rules decisions is poor, and the decision-making performance is low; the lane change decision is inefficient and the experience is poor; the intersection decision is inefficient and the experience is poor; the second purpose is to provide an electronic device; the third purpose is to provide a computer storage medium; the fourth purpose is to provide a computer program product.

[0015] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is as follows: an automatic driving decision planning method based on optimizing multiple tasks, the method comprising:

[0016] Acquire upstream input data, perform security checks on the upstream input data, and perform data structure conversion on the upstream input data;

[0017] storing the upstream input data and the upstream input data after data structure conversion;

[0018] Extracting a candidate lane set from the upstream input data, performing self-checking on the upstream input data, and cognitively grading the upstream input data based on data type and quality;

[0019] The lane centerline in the candidate lane set is combined with the lane boundary line and the vehicle body width as constraints to obtain a smooth reference line based on the Frenet coordinate system;

[0020] Traffic rules and scene decisions are made based on the candidate Frenet coordinate system and boundary, and the decision planning algorithm loops through the parameters related to traffic rules and scenes in the main memory, and the parameters related to the traffic rules and scenes are extended through the configuration file;

[0021] Make path and speed decisions based on the values ​​of parameters related to the scenario;

[0022] Based on the path decision and speed decision, path planning and speed planning are carried out based on the Frenet coordinate system;

[0023] Based on the path trajectories planned by multiple candidate lanes, multi-trajectory feature extraction is performed, and the optimal trajectory is calculated according to the multi-trajectory evaluation strategy;

[0024] The lateral and longitudinal movement of the vehicle is controlled based on traffic regulations decision, scene decision, speed decision planning, and path decision planning.

[0025] Preferably,

[0026] The smooth reference line is obtained as follows: The optimization variable is the coordinates of the discrete points (x i ,y i ), where x i ,y i All refer to the distance in the coordinate system; the optimization targets are smoothness, length and offset relative to the original point;

[0027] The optimization formula is:

[0028] Among them, P k is an unknown point;

[0029] The constraints include position constraints and curvature constraints.

[0030] The formula is:

[0031] where x k_ref represents the reference point, x u represents the upper boundary, y u Represents the upper boundary of the y-axis, stack represents the slack variable, and cur represents the curvature.

[0032] Preferably,

[0033] The path planning optimization is performed as follows: Path planning is constructed based on Frenet Frame; the optimization variable is the discrete lateral offset l i , l i ',l i "; The objective function is the lateral displacement l i , lateral displacement l i For the first, second, and third order derivatives of s; the constraints are obstacle avoidance, kinematic constraints, and path continuity.

[0034] Preferably,

[0035] Speed ​​planning optimization is performed in the following ways: speed planning is constructed based on the ST graph; heuristic speed planning based on dynamic programming provides rough analysis cost factors that meet the planning goals; speed smoothing based on quadratic programming performs fine calculations on the rough analysis results.

[0036] Preferably,

[0037] The rough analysis results are finely calculated in the following way: the optimized variables are the position, velocity and acceleration of the discrete points; the objective function is the heuristic planning position penalty acceleration and jerk corresponding to the position variable; the constraints are the continuity of the velocity value and the range of acceleration and jerk.

[0038] Preferably,

[0039] The calculation method of the optimal trajectory includes: obtaining candidate trajectories, matching the lateral and longitudinal data based on the main decision sequence number, matching the lateral data, longitudinal data and the main decision data; performing multi-trajectory feature extraction; judging whether it is a triggered lane change, and if it is a triggered lane change, performing a triggered lane change multi-trajectory evaluation to obtain the optimal trajectory; if it is not a triggered lane change, performing an expert-based trajectory evaluation to judge whether the same lane change trajectory is selected for multiple consecutive frames, and if so, performing a safety check to finally obtain the optimal trajectory; when the same lane change trajectory is not selected for multiple consecutive frames, a decision reference line priority multi-trajectory evaluation is performed, and then a safety check is performed to obtain the optimal trajectory.

[0040] Preferably,

[0041] Expert-based trajectory evaluation is conducted from the following dimensions: reference path length benefit, planned path length benefit, relative speed benefit, curvature benefit, lane change type benefit, and front space benefit;

[0042] The reference path length benefit is:

[0043] R benefit =R L -R S

[0044] Among them, R S , R L are the reference path lengths of the ego vehicle and the left lane, respectively;

[0045] The planning path length benefit is:

[0046] P benefit =P L -P S

[0047] Among them, P S , P L are the planned path lengths of the vehicle and the left lane respectively;

[0048] The relative speed benefit is:

[0049] V benefit =min(v set -v P ,v LP -v P )

[0050] Among them, v set is the set speed of the vehicle, v LP is the speed of the target car on the left, v P is the speed of the target vehicle in front, v P ' is the current real-time speed of the vehicle;

[0051] The curvature efficiency is:

[0052] K benefit =K S -K L

[0053] Among them, K S , K L are the maximum curvatures of the vehicle and the left lane, respectively;

[0054] The lane change type benefits are:

[0055] L benefit =L L -L S

[0056] Among them, L S , L L are the lane change type weights of the vehicle and the left lane respectively;

[0057] The front space benefit is:

[0058] G benefit =G LP -G P

[0059] Among them, G LP is the distance between the left front vehicle and the vehicle itself, G P is the distance between the vehicle in front and the vehicle itself.

[0060] Preferably,

[0061] After extracting the candidate lane set, traffic semantic information is provided, where the traffic semantic information includes traffic lights, pedestrian crossings, stop lines, and on / off ramps.

[0062] Preferably,

[0063] The upstream input data includes perception data, prediction data, high-precision map data, positioning data and chassis data.

[0064] Preferably,

[0065] The self-check content of the high-precision map data includes whether there are dynamic and static high-precision maps, whether there is navigation information and whether it meets the integrity, the current LinkIndex, the current RoutingIndex, the lane index and the lane line index; the self-check content of the perception data includes whether there are perceived lane lines and whether the perceived lane line curvature meets the non-integrity constraint; the cognitive classification is divided into high-precision map confidence level, static map lame level and perceived lane line slow level based on data type and quality;

[0066] The values ​​of the parameters related to the traffic regulations include traffic lights, pedestrian crossings, stop signs, yield signs, no-parking zones and lane speed limits;

[0067] The values ​​of the parameters involved in the scenarios include lane following scenario, protected intersection scenario, unprotected intersection scenario, pull-over scenario, perpendicular parking scenario, and roundabout scenario;

[0068] The values ​​of the path parameters include the path boundary within the lane, the path boundary for lane avoidance, the path boundary for lane change, and the path boundary for on-ramp and off-ramp.

[0069] The values ​​of the speed parameters include road speed limit, vehicle speed limit, pedestrian speed limit, bicycle speed limit, crosswalk speed limit, and speed bump speed limit;

[0070] The multi-trajectory features include lane change type, exploration distance, planned trajectory length in the same time interval, path curvature, maximum acceleration and deceleration, distance from the intersection and S-direction distance of the target in front and behind the target lane, TTC, speed, acceleration, and driving intention.

[0071] The present invention also provides an electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus; the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the above-mentioned autonomous driving decision-making planning method based on optimized multi-tasks.

[0072] The present invention also provides a computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned automatic driving decision-making planning method based on optimizing multiple tasks.

[0073] The present invention also provides a computer program product, comprising computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the above-mentioned autonomous driving decision-making planning method based on optimizing multiple tasks.

[0074] Beneficial effects of the present invention:

[0075] (1) The present invention first obtains prediction data, high-precision map data, positioning data, chassis data and perception data, then performs data preprocessing on the obtained data, verifies and transfers them, and then establishes a decision-making environment model. The model is based on high-precision maps and perception data, extracts a candidate lane set, evaluates the candidate lanes, obtains a smooth reference line based on the suboptimal candidate lane set, and then makes a traffic rule decision based on the candidate Frenet coordinate system and boundary, makes an obstacle decision based on the traffic rule stop_s, and makes a horizontal and vertical decision mark based on the obstacle. Then, a scene decision is made, and task planning is performed based on the scene, stage and task of the vehicle. Then, trajectory planning is performed based on the path boundary and speed boundary constraints in the Frenet coordinate system, and then the optimal trajectory is obtained through multi-trajectory evaluation. Compared with traditional decision-making and planning algorithms, fewer functional modules are involved, and the coordination between the functional modules is not a tree structure. The entire algorithm is not prone to bugs when running. From the perspective of the overall architecture, the operating efficiency is improved.

[0076] (2) In the present invention, the upstream input data obtained by the decision planning algorithm is subjected to a data security check and a data structure conversion. The data security check provides a good foundation for the healthy operation of subsequent modules. The data is converted into the required structure and stored, which can facilitate the use of subsequent algorithms and facilitate subsequent strategy optimization and model training.

[0077] (3) The present invention extracts and refines the upstream high-precision map and perception data into a data structure that is convenient for use by the decision-making planning algorithm. At the same time, it serves as a flexible isolation between the decision-making planning algorithm and the upstream high-precision map data, facilitating the expansion of the upstream high-precision map and perception data.

[0078] (4) The present invention also performs self-inspection on the high-precision map and perception data and cognitively grades the high-precision map and perception data by data type and quality. The self-inspection ensures the accuracy of the data, and cognitive grading of the data facilitates the algorithm to efficiently utilize the data.

[0079] (5) In the present invention, the lane center line is extracted based on the high-precision map, and the lane boundary line and the vehicle body width are combined as constraints to obtain a smooth reference line. The present invention provides a specific implementation method, which makes it possible to obtain a smooth reference line based on the Frenet coordinate system.

[0080] (6) Traffic regulation decisions are made based on the candidate Frenet coordinate system and boundaries. The decision-making planning algorithm loops through the parameters related to traffic regulations in the main memory, and the parameters related to traffic regulations are expanded through the configuration file, which improves the scalability of traffic regulations. At the same time, because the data obtained by the decision-making planning algorithm is more comprehensive than before, the decision-making performance of the algorithm involving traffic regulations is improved.

[0081] (7) When the present invention performs scene decision making based on the Frenet coordinate system and boundaries, the decision planning algorithm loops through the parameters involved in the scene in the main memory, and the parameters involved in the scene are expanded through the configuration file. Because the values ​​of the scene parameters become diverse through the configuration file, the scalability of the decision scene is improved.

[0082] (8) The present invention can make lane change decisions efficiently while taking into account physical perception under the premise of ensuring safety, and can also make intersection decisions efficiently while taking into account physical perception under the premise of ensuring safety.

[0083] (9) The present invention can efficiently decide the path and speed, reasonably plan the path and speed, and provide a new multi-trajectory evaluation strategy to obtain an accurate optimal trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 One of the principle diagrams of the present invention;

[0085] Figure 2 This is the second schematic diagram of the present invention;

[0086] Figure 3 A flowchart of extracting a candidate lane set according to the present invention;

[0087] Figure 4 A schematic diagram of the effect of obtaining a smooth reference line in the present invention;

[0088] Figure 5 It is a schematic diagram of traffic regulation decision;

[0089] Figure 6 It is a schematic diagram of scenario decision;

[0090] Figure 7 This is a schematic diagram of the path decision;

[0091] Figure 8 Flowchart for calculating the optimal trajectory. DETAILED DESCRIPTION

[0092] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0093] The purpose of this embodiment is to: improve the efficiency of lane change decisions while taking into account physical perception under the premise of ensuring safety; improve the efficiency of intersection decisions while taking into account physical perception under the premise of ensuring safety; improve the operating efficiency of the decision-making planning algorithm; improve the scalability of decision-making scenarios and improve the scalability and decision-making performance of traffic rules decisions.

[0094] like Figure 1 As shown, the general principle of this embodiment is: obtaining prediction data, high-precision map data, positioning data and chassis data; then preprocessing, verifying and transferring the data, and then establishing a decision-making environment model, a high-precision map-based and perception-based model, a crowdsourcing map and perception combined model, and then evaluating the candidate lanes based on the candidate lane set, obtaining a smooth reference line based on the suboptimal candidate lane set, making traffic rule decisions based on the candidate Frenet coordinate system and boundaries, making obstacle decisions based on the traffic rule stop_s, and making scene decisions based on the horizontal and vertical decision marks of the obstacles. The scene decisions include lane following, intersections, garage formation, garage parking and U-turns, and task planning based on the scene, stage and task of the vehicle, trajectory planning based on the path boundary and speed boundary constraints in the Frenet coordinate system, and then memorizing multi-trajectory feature extraction, and the multi-trajectory feature extraction also includes multi-trajectory expert evaluation and multi-trajectory evaluation (data-driven), and then performing safety checks and functional data filling to finally obtain the optimal trajectory.

[0095] like Figure 1 and Figure 2 As shown, this embodiment proposes an automatic driving decision planning method based on optimizing multiple tasks, and the method includes:

[0096] The upstream input data is obtained, and a security check is performed on the upstream input data, and a data structure conversion is performed on the upstream input data, wherein the upstream input data includes perception data, prediction data, high-precision map data, positioning data, and chassis data; the modules involved in this embodiment include a decision-making environment model, reference line generation, construction constraints, trajectory planning, and trajectory selection; the method simply includes upstream input, main decision, path planning, speed planning, post-decision, and downstream output. The upstream input data includes perception data, prediction data, high-precision map data, positioning data, and chassis data. Then the data is checked for security and the data structure conversion is performed.

[0097] The upstream input data and the upstream input data after data structure conversion are stored locally or in the cloud; the upstream input data and intermediate variable data are stored locally or in the cloud to facilitate subsequent strategy optimization and model training.

[0098] like Figure 3 As shown, a candidate lane set is extracted from the high-precision map data and perception data, traffic semantic information is provided, and self-inspection is performed on the high-precision map data and perception data through tool methods. At the same time, cognitive classification is performed on the high-precision map data and perception data based on data type and quality, wherein the candidate lane set includes lane center lines, lane boundary lines, and lane lengths; the decision-making environment model involved is to extract and refine the upstream high-precision map and perception data into a data structure that is convenient for the decision-making planning algorithm to use, and at the same time, it serves as a flexible isolation between the algorithm and the upstream high-precision map data, facilitating the expansion of the upstream high-precision map and perception data. The function of the decision-making environment model is to extract a set of candidate lanes from the high-precision map and environmental perception, and provide traffic semantic information. The traffic semantic information includes traffic lights, crosswalks, stop lines, on and off ramps, etc. Input: high-precision map road navigation and lane information, perception lane lines, positioning; the output is a candidate lane set, which includes lane center lines, lane boundary lines, lane lengths, and tool methods include query, self-inspection, and repair. Specifically, as shown in the following table,

[0099]

[0100]

[0101] Input: high-precision map, perceived lane lines, positioning; Output: candidate tool set and tool methods; Data conversion: vehicle posture, destination position, link_id where the main vehicle is located, number of lanes, lane number, distance between the main vehicle and the end position of the link, map data format conversion; Data self-check: current LinkIndex, current RoutingIndex, lane index, lane line index; Data correction: lane number where the main vehicle is located, distance between the left and right boundaries of the lane where the main vehicle is located, positioning correction based on perceived lane lines, lane boundary correction based on perceived lane lines; Generate lane candidate lane set: safety check, establish mapping table, fill in initial feasible lanes, and traverse all lane sequences within the exploration distance; Tool methods: obtain link_id and lane_num of the inlet and outlet points, traffic light information ahead, pedestrian crossing information ahead, stop line information ahead, traffic sign information ahead, rationality check and repair, and static map limping only. Data self-inspection and cognitive grading: Self-inspection of high-precision maps and perception data. The content of high-precision map self-inspection includes whether there are dynamic and static high-precision maps, whether there is navigation information and whether it satisfies integrity, current LinIndex, current RoutingIndex, lane index and lane line index; the content of perceived lane line self-inspection includes whether there are perceived lane lines and whether the curvature of perceived lane lines satisfies non-integrity constraints; cognitive grading is performed based on data type and quality, which is divided into high-precision map confidence level, static map lame level and perceived lane line slow-moving level.

[0102] like Figure 4 As shown in the figure, the lane centerline is combined with the lane boundary line and the vehicle width as constraints to obtain a smooth reference line based on the Frenet coordinate system; the lane centerline is extracted based on the high-precision map, and the lane boundary line and the vehicle width are combined as constraints to obtain a smooth reference line. The optimization variable is the discrete point coordinate (x i ,y i ), the optimization objectives are smoothness, length, and offset relative to the original point. The optimization formula is:

[0103]

[0104] Constraints: position constraints and curvature constraints, the formula is:

[0105]

[0106] where x k_ref represents the reference point, x u represents the upper boundary, y u Represents the upper boundary of the y-axis, stack represents the slack variable, and cur represents the curvature.

[0107] like Figure 5As shown, traffic regulation decisions are made based on the candidate Frenet coordinate system and boundaries, and the decision planning algorithm loops through the parameters related to traffic regulations in the main memory, and the parameters related to the traffic regulations are expanded through the configuration file; the values ​​of the variable "traffic regulations" include traffic lights, crosswalks, stop signs, yield signs, no-parking areas, and lane speed limits.

[0108] To make a scene decision, the decision planning algorithm loops through the parameters related to the scene in the main memory, and the parameters related to the scene are extended through the configuration file; Figure 6 As shown, the values ​​of the scenario parameters include lane following scenario, protected intersection scenario, unprotected intersection scenario, pull-over scenario, perpendicular parking scenario and roundabout scenario. The U-turn scenario includes approach and U-turn, the lane following scenario includes lane following, the intersection scenario includes approach and intersection cruising, the garage driving scenario includes approaching the garage entrance, passing the garage entrance, and then lane following, and the garage driving scenario also includes approaching the garage exit, passing the garage entrance (including the lifting rod), and then lane following, and the garage parking scenario includes approaching and parking. The scenario is extensible, and the stage tasks are extensible.

[0109] Path decision and speed decision are made based on the values ​​of parameters related to the scene, wherein the path decision and speed decision are based on the values ​​of the path parameters and the values ​​of the speed parameters respectively; Figure 7 As shown in the figure, the path parameter values ​​include the path boundary within the lane, the path boundary for using lanes to avoid lanes, the path for changing lanes, and the path for on-ramp and off-ramp. The speed parameter values ​​include the road speed limit, the speed limit for passing vehicles, pedestrians, and bicycles, the speed limit for pedestrian crossings, the speed limit for speed bumps, and the speed limit for more congested roads.

[0110] Based on the path decision and speed decision, the path planning and speed planning based on the Frenet coordinate system are carried out; in the path planning, the path planning optimization problem is constructed based on the Frenet Frame; the objective function is the lateral displacement l i , lateral displacement l iFor the first, second, and third order derivatives of s; the constraints are obstacle avoidance, kinematic constraints, and path continuity. In speed planning, speed planning is constructed based on the ST graph; heuristic speed planning based on dynamic programming provides a rough analysis cost factor that meets the planning goal; speed smoothing based on quadratic programming performs fine calculations on the rough analysis results. Speed ​​planning optimization problem is constructed based on the ST graph; heuristic speed planning based on dynamic programming provides a rough analysis cost factor that meets the planning goal: obstacle cost—as far away from obstacles as possible, spatial potential energy cost—as close to the farthest s as possible, the cost of the parent node is used for accumulation, and the edge cost between the parent node—speed (as close to the target speed as possible), acceleration (as small as possible), and jerk (as small as possible); speed smoothing based on quadratic programming performs fine calculations on the rough analysis results: the optimization variables are the position, speed, and acceleration of discrete points, the objective function is the position variable close to the corresponding heuristic planning position, penalizing acceleration and jerk, and the constraints are the continuity of the speed value and the range of acceleration and jerk.

[0111] Based on the path trajectories planned by multiple candidate lanes, multi-trajectory feature extraction is performed, and the optimal trajectory is calculated according to the multi-trajectory evaluation strategy; the trajectory features include lane change type (none, left, right), exploration distance (exploration reference line length), path curvature of the planned trajectory length in the same time interval, maximum acceleration and deceleration (comfort), S-direction distance of the target front and rear in the target lane, TCC, speed, acceleration, driving intention (aggressive, regular, conservative), and distance from the intersection.

[0112] The calculation method of the optimal trajectory includes: obtaining candidate trajectories, matching the lateral and longitudinal data based on the main decision sequence number, matching the lateral data, longitudinal data and the main decision data; performing multi-trajectory feature extraction; judging whether it is a triggered lane change, and if it is a triggered lane change, performing a triggered lane change multi-trajectory evaluation to obtain the optimal trajectory; if it is not a triggered lane change, performing an expert-based trajectory evaluation to judge whether the same lane change trajectory is selected for multiple consecutive frames, and if so, performing a safety check to finally obtain the optimal trajectory; when the same lane change trajectory is not selected for multiple consecutive frames, a decision reference line priority multi-trajectory evaluation is performed, and then a safety check is performed to obtain the optimal trajectory.

[0113] Expert-based trajectory evaluation is conducted from the following dimensions: reference path length benefit, planned path length benefit, relative speed benefit, curvature benefit, lane change type benefit, and front space benefit;

[0114] The longer the reference line is, the farther the vehicle goes and the greater the intention to reach the corresponding lane. The reference path length benefit is:

[0115] R benefit =R L -R S

[0116] Among them, R S , R L are the reference path lengths of the ego vehicle and the left lane, respectively;

[0117] The reference line road does not consider obstacles. It is possible that the planned reference path is 800m (the target is tentatively set at 800m). In this case, the reference line path length benefit evaluation factor will be invalid. Therefore, consider introducing the planning path length evaluation factor. The longer the longitudinal planning path length, the greater the intention to reach the corresponding lane. The planning time is 8s, and the planning path length benefit is:

[0118] P benefit =P L -P S

[0119] Among them, P S , P L are the planned path lengths of the vehicle and the left lane respectively;

[0120] For an autonomous vehicle, the future speed can be converted to the speed of the vehicle ahead, so the relative speed benefit is:

[0121] V benefit =min(v set -v P ,v LP -v P )

[0122] Among them, v set is the set speed of the vehicle, v LP is the speed of the target car on the left, v P is the speed of the target vehicle in front, v P ' is the current real-time speed of the vehicle;

[0123] The left lane has priority;

[0124] Considering comfort, it is necessary to consider introducing the path curvature evaluation factor. Among the three planned trajectories, try to choose the one with little curvature change. The flatter the curvature, the more obvious the intention to reach the corresponding lane. The curvature efficiency is:

[0125] K benefit =K S -K L

[0126] Among them, K S , K L They are the maximum curvatures of the vehicle and the left lane respectively; Note: this part can be added for safety. If the maximum curvature is greater than the threshold, it is considered that comfort and safety are insufficient.

[0127] It is believed that lane changing for an autonomous vehicle is a relatively unsafe operation compared to driving straight. Try to change lanes as little as possible unless the benefits of lane changing are great. The weight of the lane change type evaluation factor should be relatively large. The benefits of lane change type are:

[0128] L benefit =L L -L S

[0129] Among them, L S , L L are the lane change type weights of the vehicle and the left lane, assuming that the straight lane is 0.6 and the left and right lane changes are 0.2 respectively. The larger the front space, the greater the benefit, and the greater the intention to enter the corresponding lane. The front space benefit is:

[0130] G benefit =G LP -G P

[0131] Among them, G LP is the distance between the left front vehicle and the vehicle itself, G P is the distance between the vehicle in front and the vehicle itself.

[0132] The lateral and longitudinal movement of the vehicle is controlled based on traffic regulations decision, scene decision, speed decision planning, and path decision planning.

[0133] This embodiment also provides an electronic device, including: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus; the memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the autonomous driving decision planning method based on optimized multi-task.

[0134] This embodiment also provides a computer storage medium on which a computer program is stored. When the program is executed by a processor, the automatic driving decision planning method based on optimizing multiple tasks as described above is implemented.

[0135] This embodiment also provides a computer program product, including computer instructions, which instruct a computing device to perform operations corresponding to the above-mentioned autonomous driving decision-making planning method based on optimizing multiple tasks.

[0136] The advantages of the present invention are: better scalability of decision scenarios, better scalability and decision performance of traffic rules decisions; it can make lane change decisions efficiently and take into account physical perception under the premise of ensuring safety; it can make intersection decisions efficiently and take into account physical perception under the premise of ensuring safety; the decision-making planning algorithm has high operating efficiency. The present invention can better complete the L4 level urban autonomous driving task. The present invention relates to the characteristics and evaluation strategies of multiple trajectories.

[0137] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.

Claims

1. An autonomous driving decision planning method based on optimizing multiple tasks, characterized in that: The method comprises: Acquire upstream input data, perform security checks on the upstream input data, and perform data structure conversion on the upstream input data; storing the upstream input data and the upstream input data after data structure conversion; Extracting a candidate lane set from the upstream input data, performing self-checking on the upstream input data, and cognitively grading the upstream input data based on data type and quality; The lane centerline in the candidate lane set is combined with the lane boundary line and the vehicle width as constraints to obtain the Frenet Smooth reference lines of the coordinate system; Traffic rules and scene decisions are made based on the candidate Frenet coordinate system and boundary, and the decision planning algorithm loops through the parameters related to traffic rules and scenes in the main memory, and the parameters related to the traffic rules and scenes are extended through the configuration file; Make path and speed decisions based on the values ​​of parameters related to the scenario; Based on the path decision and speed decision, path planning and speed planning are carried out based on the Frenet coordinate system; Based on the path trajectories planned by multiple candidate lanes, multi-trajectory feature extraction is performed, and the optimal trajectory is calculated according to the multi-trajectory evaluation strategy; Control the lateral and longitudinal movement of the vehicle based on traffic rules decision, scene decision, speed decision planning, and path decision planning; The smooth reference line is obtained as follows: The optimization variable is the coordinates of the discrete points (x i ,y i ), where x i ,y i All refer to the distance in the coordinate system; the optimization targets are smoothness, length and offset relative to the original point; The optimization formula is: Among them, P k is an unknown point; The constraints include position constraints and curvature constraints. The formula is: where x k_ref represents the reference point, x u represents the upper boundary, y u Represents the upper boundary of the y-axis, stack represents the slack variable, and cur represents the curvature.

2. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 1, characterized in that: The path planning optimization is performed as follows: Path planning is constructed based on Frenet Frame; the optimization variable is the discrete lateral offset l i , l i ',l i "; The objective function is the lateral displacement l i , lateral displacement l i For the first, second, and third order derivatives of s; the constraints are obstacle avoidance, kinematic constraints, and path continuity.

3. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 1, characterized in that: Speed ​​planning optimization is performed in the following ways: speed planning is constructed based on the ST graph; heuristic speed planning based on dynamic programming provides rough analysis cost factors that meet the planning goals; speed smoothing based on quadratic programming performs fine calculations on the rough analysis results.

4. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 3 is characterized in that: The rough analysis results are finely calculated in the following way: the optimized variables are the position, velocity and acceleration of the discrete points; the objective function is the heuristic planning position penalty acceleration and jerk corresponding to the position variable; The constraints are the continuity of velocity values ​​and the range of acceleration and jerk.

5. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 1, characterized in that: The calculation method of the optimal trajectory includes: obtaining candidate trajectories, matching the lateral and longitudinal data based on the main decision sequence number, matching the lateral data, longitudinal data and the main decision data; performing multi-trajectory feature extraction; judging whether it is a triggered lane change, and if it is a triggered lane change, performing a triggered lane change multi-trajectory evaluation to obtain the optimal trajectory; if it is not a triggered lane change, performing an expert-based trajectory evaluation to judge whether the same lane change trajectory is selected for multiple consecutive frames, and if so, performing a safety check to finally obtain the optimal trajectory; when the same lane change trajectory is not selected for multiple consecutive frames, a decision reference line priority multi-trajectory evaluation is performed, and then a safety check is performed to obtain the optimal trajectory.

6. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 5, characterized in that: Expert-based trajectory evaluation is conducted from the following dimensions: reference path length benefit, planned path length benefit, relative speed benefit, curvature benefit, lane change type benefit, and front space benefit; The reference path length benefit is: R benefit =R L -R S Among them, R S , R L are the reference path lengths of the ego vehicle and the left lane, respectively; The planning path length benefit is: P benefit =P L -P S Among them, P S , P L are the planned path lengths of the vehicle and the left lane respectively; The relative speed benefit is: V benefit =min(v set -v P ,v LP -v P ) Among them, v set is the set speed of the vehicle, v LP is the speed of the target car on the left, v P is the speed of the target vehicle in front, v P ' is the current real-time speed of the vehicle; The curvature efficiency is: K benefit =K S -K L Among them, K S , K L are the maximum curvatures of the vehicle and the left lane, respectively; The lane change type benefits are: L benefit =L L -L S Among them, L S , L L are the lane change type weights of the vehicle and the left lane respectively; The front space benefit is: G benefit =G LP -G P Among them, G LP is the distance between the left front vehicle and the vehicle itself, G P is the distance between the vehicle in front and the vehicle itself.

7. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 1, characterized in that: After extracting the candidate lane set, traffic semantic information is provided, where the traffic semantic information includes traffic lights, pedestrian crossings, stop lines, and on / off ramps.

8. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 1, characterized in that: The upstream input data includes perception data, prediction data, high-precision map data, positioning data and chassis data.

9. The autonomous driving decision-making planning method based on optimizing multiple tasks according to claim 8, characterized in that: The self-check content of the high-precision map data includes whether there are dynamic and static high-precision maps, whether there is navigation information and whether it meets the integrity, the current LinkIndex, the current RoutingIndex, the lane index and the lane line index; the self-check content of the perception data includes whether there are perceived lane lines and whether the perceived lane line curvature meets the non-integrity constraint; the cognitive classification is divided into high-precision map confidence level, static map lame level and perceived lane line slow level based on data type and quality; The values ​​of the parameters related to the traffic regulations include traffic lights, pedestrian crossings, stop signs, yield signs, no-parking zones and lane speed limits; The values ​​of the parameters involved in the scenarios include lane following scenario, protected intersection scenario, unprotected intersection scenario, pull-over scenario, perpendicular parking scenario, and roundabout scenario; The values ​​of the parameters related to the path include the path boundary within the lane, the path boundary for avoiding the lane, the path boundary for changing lanes, and the path boundary for on-ramp and off-ramp; The values ​​of the parameters related to the speed include road speed limit, vehicle speed limit, pedestrian speed limit, bicycle speed limit, crosswalk speed limit, and speed bump speed limit; The multi-trajectory features include lane change type, exploration distance, planned trajectory length in the same time interval, path curvature, maximum acceleration and deceleration, distance from the intersection and S-direction distance of the target in front and behind the target lane, TTC, speed, acceleration, and driving intention.

10. An electronic device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the automatic driving decision planning method based on optimized multi-task as described in any one of claims 1 to 9.

11. A computer storage medium having a computer program stored thereon, which, when executed by a processor, implements the autonomous driving decision-making planning method based on optimized multi-task as described in any one of claims 1 to 9.

12. A computer program product, comprising computer instructions, wherein the computer instructions instruct a computing device to perform operations corresponding to the autonomous driving decision planning method based on optimized multi-task as described in any one of claims 1 to 9.

Images