Path planning decision method for autonomous driving guided by high-precision maps

Abstract

The high-precision map-guided autonomous driving path planning decision method of this application is based on the large amount of rich and accurate road static information provided by the high-precision map, based on the navigation application mode, combined with the path planning decision method based on the perception system and the simple element map, and proposes a path planning decision system framework, modules and methods suitable for smart cars, and focuses on three aspects of improvement: first, the lane trajectory level path planning method guided by the high-precision map, second, the driving behavior decision method guided by the high-precision map, and third, the local path planning and selection method guided by the high-precision map. The three different indicators of decision cost, road cost and comfort cost that affect the selection of driving curves are integrated through cost calculation, and the trajectory line is selected by weighted summation instead of the minimum cost. Through the above three improvements, the reliability, safety, intelligence and economy of autonomous driving have been greatly improved.

Description

Technical Field

[0001] The present application relates to a vehicle assisted driving path planning decision system, and in particular to a high-precision map-guided automatic driving path planning decision method, which belongs to the field of automotive AI intelligent driving technology. Background Art

[0002] In recent years, autonomous driving technology has flourished. Traditional autonomous driving systems are an extension of the robot architecture and are divided into three modules: perception, planning and decision-making, and control. Perception is the use of various sensors such as cameras and radars to detect and identify the surrounding environment, so as to "see" the surrounding environment, which is analogous to the human eye. Planning and decision-making is based on the fusion of multiple sensor information and the premise of having a clear understanding of itself and the surrounding environment. In the simulated manual driving state, the human brain makes a comprehensive judgment on the importance of each factor and finally makes various decisions and plans different driving paths. Therefore, the planning and decision-making module of the autonomous driving system is also the most complex part. Control is to implement the brain's planning into the actual control system, including brakes, accelerators, and brakes.

[0003] Currently, a large part of traffic accidents are caused by human driving errors. The bad habits of human driving also cause the congestion to increase. Compared with human driving, machines do not need to rest like humans, nor do they feel tired. They always keep detecting the surrounding environment to minimize the risk of collision. Therefore, the application of autonomous driving technology will greatly improve the travel efficiency of the city, thereby reducing the cost of related industries and improving people's living standards. In addition, excellent autonomous driving technology can reduce the occurrence of traffic accidents. Therefore, the early realization of autonomous driving technology also has great social significance.

[0004] However, practice has proven that the existing autonomous driving technology is not reliable. If the system is assisted by maps, the system can retrieve the entire intersection from a few kilometers away, and by saving historical data through maps, it is easy to find intersections that are prone to traffic accidents. Then the vehicle can be warned in advance, slow down, enhance perception, etc. to avoid accidents. The autonomous driving system has repeatedly made mistakes, and the bloody lessons have made everyone realize that high-precision maps are indispensable in the autonomous driving system, but the current application of high-precision maps in autonomous driving is still in its infancy.

[0005] Due to the extremely high safety requirements of automobiles, the most critical issue restricting the commercialization of autonomous driving technology is the reliability of the technology itself. Therefore, the addition of maps can improve the performance of the perception, positioning and planning decision-making modules that are closely related to the safety of the autonomous driving system, which will help to greatly accelerate the popularization and application of this technology and provide better travel services.

[0006] Currently, high-precision maps have been produced in small batches, but the cost of producing high-precision map data is ten times that of ordinary map data, and there are currently no sufficient application areas. The field of autonomous driving is the focus of high-precision map applications, and solving the application problem of high-precision maps in autonomous driving technology is a top priority. However, at present, the combination of high-precision maps and autonomous driving technology is still in its early stages.

[0007] There are many methods currently applied to smart cars, but these methods are all based on simple feature maps. Due to the lack of support from engineered high-precision maps, they cannot adapt to complex real road conditions and therefore cannot be further engineered. How to use the information from high-precision maps to improve the existing smart car path planning decision-making methods is the core of current research.

[0008] The existing intelligent car behavior decision-making methods rely solely on the detection of the perception system or the support of simple element maps. In some complex scenarios that are not easily recognized or cannot be recognized by the perception system, it is easy to make wrong decisions, thus causing collision risks. How to use the detailed data of high-precision maps to avoid wrong decisions and reduce risks is also a current research focus.

[0009] The path planning decision module is the brain of the smart car and one of the most critical technologies in the field of smart cars. The performance indicators of this module also largely determine the height that the entire system can reach. The smart car planning decision system in the existing technology is divided into four levels. The first level is the long-distance road-level planning based on conventional map data, which is to realize the planning from the current location to the destination, that is, task planning; the second level is the lane trajectory-level path planning guided by high-precision maps, which is based on the road-level planning and further obtains the information of the lane trajectory-level planning; the third level is the behavior decision guided by high-precision maps. This step is to abstract and summarize the vehicle driving experience and traffic planning to assist in the selection of the final path. The fourth level is the local path planning, that is, to solve the driving path within one or two hundred meters in front of the current vehicle. This step takes into account the location and behavior prediction results of obstacles and the movement constraints of the vehicle, and finally selects an executable, collision-free optimal local driving trajectory that conforms to vehicle dynamics. Although there are many research contents on path planning methods, the concept of high-precision maps has not been popularized before, so the current smart car path planning decision system has not made good use of high-precision map data.

[0010] In general, the existing technology still has several problems and defects. The key technical difficulties of the autonomous driving path planning decision system are as follows:

[0011] (1) Existing autonomous driving systems often make mistakes and cannot meet the extremely high safety requirements of automobiles. The most critical issue restricting the commercialization of autonomous driving technology is the reliability of the technology itself. The current autonomous driving system has several problems in the perception, positioning and planning decision modules that are closely related to safety, which is not conducive to the popularization and application of this technology. The current high-precision map data production cost is high, and the combination of high-precision maps and autonomous driving technology is not yet mature. The existing technology methods applied to smart cars are all based on simple element maps. Due to the lack of support for engineered high-precision maps, they cannot adapt to complex real road conditions. Therefore, they cannot be further engineered and cannot use the information of high-precision maps to improve the existing smart car path planning decision methods. Existing smart car behavior decision methods rely solely on the detection of the perception system or the support of simple element maps. In some complex scenes that are not easily recognized or cannot be recognized by the perception system, it is easy to make mistakes in decision-making, thereby causing collision risks. It is impossible to use the detailed data of high-precision maps to avoid decision-making mistakes and reduce risks. The existing technology of autonomous driving cannot meet the requirements in terms of reliability, safety, intelligence and economy.

[0012] (2) The existing technology of autonomous driving lacks a lane trajectory level path planning method guided by high-precision maps. The lane trajectory level path planning method is a necessary step to be carried out in the context of high-precision map data accurate to the lane trajectory level. The conventional lane trajectory level planning method in the existing technology has a relatively simple corresponding map and simply performs lane trajectory level topology calculation. It lacks a high-precision map with high precision and complete elements, and is unable to further calculate the similar lanes of each lane trajectory level planning section. It lacks an accurate description of the drivable area of ​​the structured road, and is unable to provide effective support for the vehicle's lane change and the generation of local planning trajectory clusters. This is fatal to the autonomous driving system and is unable to provide safe and economical path planning decisions, ultimately resulting in the loss of value for large-scale promotion and application of autonomous driving.

[0013] (3) Autonomous driving lacks decision-making methods guided by high-precision maps. The decision-making methods of existing technologies are based on perceived information. However, some information (such as the color and type of lane lines and beyond-visual-range information) is not easy to be perceived and identified, but is of great significance to decision-making (double yellow lines and solid white lines cannot change lanes, and accidents are prone to occur at intersections outside the line of sight). Therefore, there is a certain probability of decision-making errors. Regarding the decision-making risks that may be caused by the uncertainty of the perception system, due to the lack of high-precision maps to provide accurate ground static information, it is impossible to avoid decision-making errors caused by the uncertainty of perception, resulting in huge potential risks to safety and reliability. As a result, autonomous driving decisions are not only unsafe, but also lack scientificity and authority.

[0014] (4) Autonomous driving lacks local path planning guided by high-precision maps. The existing autonomous driving team technology is mostly derived from smart car competitions, so local path planning methods are all implemented based on road network files (domestic and foreign teams are based on RNDF files defined by DARPA), but the road network file information is sparse and the data model is too simple, which cannot adapt to more complex road scenes, thus bringing additional difficulties to local planning. In response to the local road expression dilemma and planning boundary fuzziness caused by road network files, the existing technology lacks a detailed road expression model combined with high-precision maps and accurate lane geometry and topology information, and cannot solve the road boundary defuzzification problem of local path planning methods. At the same time, the path selection results cannot fully match the structured road driving rules, lack lane and lane line data combined with high-precision maps, lack the cost calculation of path selection, and cannot solve the path selection to ensure that the vehicle drives in the center of the structured road as much as possible, which makes autonomous driving have serious safety risks and is also easy to affect the overall order of road traffic. Such autonomous driving models have lost their practical value. Summary of the invention

[0015] The current autonomous driving technology cannot achieve sufficient safety. It only relies on the perception system without the assistance of maps, and its stability is insufficient. After combining high-precision maps with the multi-level advanced autonomous driving system, the role of high-precision maps can not only assist the vehicle's perception system, but also the positioning system can assist the vehicle's path planning and decision-making system, which has greatly upgraded the stability of autonomous driving. Specifically, through lane trajectory-level path planning guided by high-precision maps, behavioral decision-making methods guided by high-precision maps, and local path planning and selection guided by high-precision maps, the safety, scientificity and authority of autonomous driving decisions have been greatly improved. Tests have proved that the autonomous driving of this application, as a substitute for manual driving, can not only significantly reduce the incidence of traffic accidents, but also greatly improve the operating efficiency of urban highway traffic. It has huge market value and social value, and effectively solves the most direct safety problem that restricts the commercialization of autonomous driving technology.

[0016] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0017] The HD map-guided autonomous driving path planning decision-making method first constructs the elements and model data of the HD map: based on the optimal lane planning, a comprehensive plan for multiple suboptimal alternative lanes is obtained. The plan includes three parts. The first part is to perform lane trajectory-level topology calculation based on the one-to-many inclusion relationship between the road LinkID and the lane LaneID on the basis of road-level planning to ensure connectivity; the second part is to obtain the lane LaneID and ensure lane connectivity, and then trace back layer by layer to calculate the same type of lanes using the differences in attributes between lanes; the third part is to segment the road segment based on the HD map data itself, and make virtual lane connections for smoothing the step part according to the tangent of the road; the lane trajectory-level path planning is refined in two points:

[0018] (1) Constructing similar lanes. In addition to the LaneID obtained by lane trajectory-level planning, similar lanes are also provided to ensure local path planning and lane change requirements, reflecting the actual road conditions to the vehicle;

[0019] (2) Perform tangent point search on the search results and use the cosine function to perform tangent translation to ensure smooth connection between the center lines of the two lanes;

[0020] Then, the HD map-guided path planning decision system is divided into three layers: the first layer is the lane trajectory level path planning layer, the second layer is the behavior decision layer, and the third layer is the local path planning layer. Each layer method analyzes the HD map data for the path planning decision in the direction of autonomous driving application.

[0021] 1) Lane trajectory level path planning method guided by high-precision map: First, the Dijkstra method is improved to perform lane trajectory level path planning for high-precision maps. Then, the correspondence between roads and lanes in the map is used to reversely calculate the drivable area of ​​the same type of lane expression plan to meet the vehicle's obstacle avoidance and lane change requirements. Then, a virtual lane is constructed through a cosine function to perform tangential smoothing on the planned lane to solve the lane jump problem caused by map division in lane trajectory level planning and meet the requirements of path planning smoothness. Finally, based on the advantage of minimal resource consumption in short-distance map topological relationship calculation, a lane trajectory level path planning trigger method is constructed, including two items: insufficient planning distance and lane change driving.

[0022] High-precision map lane trajectory level topology calculation: Through high-precision positioning, the current lane s_lane and the end lane e_lane are matched. Each long-distance planning is completed by the road-level planning calculation. Each lane trajectory level planning only needs to calculate the data of the first 1 to 2 km.

[0023] Based on the lane trajectory level topology information, the lane change action must be achieved by connecting multiple lane LaneIDs belonging to the same road LinkID;

[0024] The lane change itself is converted into a distance cost like the lane length and is involved in the path calculation. The lane cost is the direct lane distance. The lane change cost is estimated, a virtual lane change route is generated, and the distance of the virtual lane change route is calculated and multiplied by a weight:

[0025] Cost=dst*p Formula 1

[0026] Cost is the cost of changing lanes, dst is the distance of the virtual lane change route, and p is the weight;

[0027] Smoothing of lane change jumps: A solution that uses a cosine function to generate a virtual lane at the tangential jump of the lane centerline. By continuously translating the points at the tangent plane of the existing road section to connect the two tangentially translated lanes, the strength of the jump is reduced;

[0028] Lane trajectory level path planning triggering methods include:

[0029] (a1) When the length of a single lane trajectory level path planning exceeds 2 km, the first trigger condition is that the vehicle is about to reach the critical value of the result data of a single lane trajectory level path planning, which automatically triggers the lane trajectory level path planning to further plan;

[0030] (a2) When the vehicle leaves the recommended lane of the current lane trajectory level path planning by changing lanes, the lane trajectory level path planning is automatically triggered;

[0031] 2) Driving behavior decision-making method guided by high-precision maps: First, the action decision items under structured road surface conditions are analyzed and expressed through the state switching mechanism of the finite state machine. Then, the switching conditions between the various state items of the state machine are analyzed. The important state machine switching conditions include: using the map to map obstacles, predicting the behavior of obstacles through the map, and classifying left and right lane changes as lateral decisions, and the remaining decision items as longitudinal decisions; for lateral decisions, the lane change intention generation starting point judgment and lane change intention generation and lane change condition judgment are divided into three steps. Among them, the absolute difference between the expected vehicle speed and the current vehicle speed is greater than a certain critical value as the starting point for the lane change intention generation, and the lane change intention is generated by the accumulated driving efficiency loss greater than a certain critical value. Finally, the lane change condition is judged by combining the lane change intention with the map data, and finally a decision is generated;

[0032] Action decision items based on structured road traffic regulations:

[0033] (b1) Lane Straightening - Conventional decision, keep driving in the current lane: Regardless of whether there is an obstacle in the current lane, the vehicle keeps driving in the lane until a new decision causes a state switch;

[0034] (b2) Left turn at intersection - conventional decision, in the case of intersections, the following will be encountered: after the intersection, the vehicle enters a slow operation, adapting to the faster planning frequency and perception frequency, which brings more calculation delays, and some other attributes of the turning process will gradually be added and applied;

[0035] (b3) Right turn at intersection - conventional decision, in the case of intersections, the following will be encountered: after the vehicle at the intersection enters the predetermined lane, it no longer judges the traffic light information and directly enters the lane change state;

[0036] (b4) Go straight at the intersection - conventional decision, in the case of an intersection, the following will be encountered: the vehicle will enter the intersection and move slowly, adapting to the more computational delay caused by faster planning and perception frequency;

[0037] (b5) Parking at an intersection - conventional decision-making. When approaching an intersection, you will encounter: vehicles entering a slow-moving, following-the-car state;

[0038] (b6) Left lane change—conventional decision, encountered while keeping the lane: the weight of the left road is increased;

[0039] (b7) Right lane change - conventional decision, encountered while keeping lane: the weight of the right road is increased;

[0040] (b8) U-turn driving—conventional decision: distinguish U-turn lanes and increase the weight of U-turn lanes;

[0041] The obstacle behavior prediction method combined with the lane attribute information of the map is as follows: the speed of the lane direction and the lane tangent is decomposed according to the vehicle speed, and the tangential speed Vx is obtained; at the same time, the center point of the vehicle is matched with the center point of the lane, and the matching distance D1 is obtained. The tangential distance between the center point of the vehicle and the lane boundary is half the lane width minus D1, and the calculation formula for the lateral crossing lane time is obtained as follows:

[0042]

[0043] T is the time of lateral lane crossing, V is the actual speed of the vehicle, V x represents the tangential component of the vehicle speed V on the road, δ represents the angle between the vehicle speed direction and the lane direction, LanePt.Heading represents the lane direction, Vehicle.Heading represents the vehicle driving direction, LanePt represents the matching point of the vehicle center on the lane centerline, and Width represents the lane width; the time to cross the lane line is predicted to be between 1.5 and 2 seconds for small vehicles and 2.5 seconds for large vehicles. It is determined that the vehicle has lateral movement and constraints are imposed on the decision;

[0044] Lateral decision of lane-changing vehicles: Lane-changing decisions include: first, the generation of lane-changing intentions, second, the timing of lane-changing intentions or the conditions for abandoning them, and third, the judgment of lane-changing decisions;

[0045] (c1) Generate lane change intention: Based on the description of the cumulative driving efficiency loss LCDT, the lane change intention is generated when the current vehicle speed is less than the expected speed; the time integral of the value of the current speed being less than the expected speed is compared with the pre-set critical value to determine whether the generation condition of the lane change intention is met;

[0046] (c2) Lane change intention start time or abandonment condition: Determine the time point to start accumulating the driving efficiency loss cumulative value. When the difference between the expected speed and the current speed is greater than the set critical value, the driving efficiency loss cumulative calculation starts. When the difference between the expected speed and the current driving speed is less than the set critical value, the calculated cumulative value is abandoned and reset to zero, waiting for the next start;

[0047] (c3) Lane change decision judgment: When the cumulative value of driving efficiency loss is greater than the set critical value, the lane change feasibility judgment is initiated;

[0048] 3) Local path planning and selection method guided by high-precision maps: First, the local path is planned, and then the center line of the lane planned at the map lane trajectory level is extracted as the planning baseline, and the boundary of the drivable area is used as the maximum margin of the trajectory cluster. In the case that the planning curve crosses the drivable areas of different widths, the narrowest drivable area boundary is used as the maximum margin of the trajectory cluster. Finally, the decision cost, road cost and comfort cost that affect the driving curve selection are integrated through cost calculation, and the trajectory line is selected by weighted summation instead of the minimum cost method, where the road cost is obtained by extracting the lane line type;

[0049] Real-time local path trajectory cluster planning:

[0050] The first step is to obtain the planning base baseline: obtain a center line base baseline of the road followed by the trajectory, which acts as a baseline. All trajectory lines in the trajectory cluster use the trajectory center line as the planning baseline, and then generate a series of trajectory lines near the baseline. The baseline is both the core line of the trajectory cluster and provides the movement direction of the end of the trajectory line;

[0051] The second step is baseline smoothing: each trajectory is generated to ensure that it is as smooth as possible and the curvature changes linearly. The baseline is composed of a series of hash points, and the trajectory generation needs to be further expressed by a function.

[0052] The third step is to obtain the tangential edge constraint conditions: the tangential divergence of the trajectory cluster is limited effectively, and the boundary is reasonable to ensure the effectiveness of the planning and controllable calculation amount;

[0053] Step 4: Generation of trajectory clusters: Through the side deviation function, superimposed with the baseline, and adjusting the maximum side deviation distance, a series of trajectory data with a certain lateral spacing are generated. Each trajectory of these trajectory clusters satisfies the vehicle kinematic constraints and is related to the current vehicle position and posture;

[0054] Generate trajectory clusters: q left and q right Represent the maximum side deviation distances on the left and right sides respectively, and obtain the center line of the lane as the base baseline of the trajectory cluster and the side deviation distance limit parameter q of the vehicle. left and q right After that, the vehicle's drivable area is obtained, and a series of drivable trajectories are generated in this space. The generation of these trajectories ensures that the vehicle has sufficient maneuverability and makes driving behavior safer and more comfortable. The steps for generating trajectory clusters include:

[0055] Step 1: Positioning and matching with the baseline: The vehicle is mapped from the Cartesian coordinate system to the curved coordinate system of the baseline through the matching process, using the heading information of the base baseline;

[0056] Step 2: Determine the length of a single local path planning: determine the distance within which the vehicle needs to correct the heading and offset errors between the current vehicle and the baseline;

[0057] Step 3: construct the side deviation change function and generate trajectory clusters;

[0058] The length of a single path planning consists of two parts. The first part is the adjustment stage of the heading deviation and the lateral distance deviation. A cubic spline curve is used to simulate the curve of the adjustment stage and is solved by the following formula:

[0059]

[0060]

[0061]

[0062] Δs=s-si Formula 7

[0063] Among them, q(s) is the cubic spline function in the adjustment stage, s i represents the mapping point on the baseline obtained by matching the vehicle, s f represents the arc length during the adjustment phase, q f is the lateral offset value of a certain track line in the track cluster, calculated by the same lane; q i It represents the lateral distance between the vehicle and the baseline calculated by matching the current point, a, b, c represent the corresponding parameters, and s represents the function arc length variable;

[0064] Comfort cost path selection: Using the cost calculation method, a cost is calculated for all the alternative trajectories in the trajectory cluster, and multiple factors affecting trajectory selection are normalized to form a unified standard. Normalizing each factor is to give each factor affecting the cost a quantitative measurement value and assign weights respectively, so as to finally obtain the cost calculation function.

[0065] Preferably, the boundary constraint conditions are extracted based on the high-precision map: the high-precision map is introduced, and after the lane trajectory-level path planning guided by the high-precision map, a lane trajectory-level path planning result and similar lanes in each lane section that can be used for lane change and obstacle avoidance are obtained. The acquisition of the same lane is the maximum deviation of the lateral path planning;

[0066] The specific method is as follows: After extracting the planned lane centerline hash point string from the lane ID obtained through lane trajectory level path planning each time, the similar lanes of each lane trajectory level planned lane Lane ID are obtained through the results of lane trajectory level path planning. These similar lanes all support lanes for vehicle lane change. The maximum side deviation distance is determined by the similar lane data. Assuming that the width of each normal lane is the same, the maximum side deviation distance q left / right Calculation:

[0067] q left / right =NumofBrotherLanes left / right *Width formula 3

[0068] NumofBrotherLaneS left / right Indicates the number of lanes of the same type, Width indicates the lane width, and the maximum side deviation distance q left / right It must be the minimum lateral distance within a single planning distance, centered on the lane trajectory level planning.

[0069] Preferably, the comfort cost path selection is: the curvature on the entire trajectory is cumulatively calculated and used as the comfort cost of the trajectory. Based on the fact that the influence factor of curvature is exponential, the square of the curvature value is used as the integration condition:

[0070]

[0071] In the above formula, C k [i] is the comfort cost, k i represents the curvature of the ith point, and s is the arc length along the trajectory;

[0072] Finally, assuming that the three types of costs are cost1, cost2, and cost3, and the factors affecting the final choice are different, a weight should be added to obtain the final cost calculation formula:

[0073] cost=q1*cost1+q2*cost2+q3*cost3 Equation 9

[0074] cost is the final cost, q1, q2, q3 are the weights corresponding to the cost, and the final cost diagram is obtained.

[0075] Compared with the prior art, the innovations and advantages of this application are:

[0076] First, this application is based on the large amount of rich and accurate road static information provided by real high-precision maps. Based on the navigation application mode, it combines the path planning decision method based on the perception system and the simple element map, and proposes a path planning decision system framework, modules and methods based on high-precision maps suitable for smart cars, and focuses on three aspects of improvement: First, the lane trajectory level path planning method guided by high-precision maps: improve the Dijkstra method to perform high-precision map lane trajectory level path planning, use the correspondence between roads and lanes in the map to reversely calculate the drivable area of ​​the same type of lane expression plan to meet the vehicle's obstacle avoidance and lane change needs, construct a virtual lane through the cosine function to perform tangential smoothing on the planned lane to solve the lane jump problem caused by map division in the lane trajectory level planning, and meet the path planning requirements. The first is the need for smoothness of the lane, and the triggering methods of the lane trajectory level path planning method include insufficient planning distance and lane change driving; the second is the driving behavior decision method guided by high-precision maps: it is expressed through the state switching mechanism of the finite state machine, and then the switching conditions between the various state items of the state machine are analyzed, and the left and right lane changes are classified as lateral decisions, and the remaining decision items are classified as longitudinal decisions; the third is the local path planning and selection method guided by high-precision maps: the center line of the lane planned at the lane trajectory level of the map is extracted as the planning baseline, the boundary of the drivable area is used as the maximum margin of the trajectory cluster, and the narrowest drivable area boundary is used as the maximum margin of the trajectory cluster. The decision cost, road cost and comfort cost that affect the selection of driving curves are integrated through cost calculation, and the trajectory line is selected by weighted summation instead of the minimum cost method. Through the above three improvements, the reliability, safety, intelligence and economy of autonomous driving have been greatly improved, which has great practical value.

[0077] Second, the current autonomous driving technology cannot achieve sufficient safety. It only relies on the perception system without the assistance of maps, and its stability is insufficient. After combining high-precision maps with the multi-level advanced autonomous driving system, the role of high-precision maps can not only assist the vehicle's perception system, but also the positioning system can assist the vehicle's path planning and decision-making system, which has greatly upgraded the stability of autonomous driving. Specifically, through lane trajectory-level path planning guided by high-precision maps, behavioral decision-making methods guided by high-precision maps, and local path planning and selection guided by high-precision maps, the safety, scientificity and authority of autonomous driving decisions have been greatly improved. Tests have proved that the autonomous driving of this application, as a substitute for manual driving, can not only significantly reduce the incidence of traffic accidents, but also greatly improve the operating efficiency of urban highway traffic. It has huge market value and social value, and effectively solves the most direct safety problem that restricts the commercialization of autonomous driving technology.

[0078] Third, in response to the problem that conventional lane trajectory-level path planning methods cannot express the complete drivable area required by smart cars, this application proposes the concept and calculation method of similar lanes, and improves the conventional smart car lane trajectory-level path planning method based on perception systems and simple element maps. The actual drivable area is expressed by combining similar lanes. On the one hand, unconventional lanes such as emergency lanes are included in the drivable area of ​​smart cars on conventional roads, so that smart cars can choose to drive or park in emergency lanes in emergency situations; on the other hand, in the scenario of road forks, some normal lanes on the same road section that are different from the lane trajectory-level planning driving direction are filtered out, thereby avoiding interference, which is conducive to improving the overall order of road traffic, and helps to greatly accelerate the popularization and application of this technology and provide better travel services.

[0079] Fourth, in order to address the risk of wrong decision-making caused by the lack of lane attributes, incorrect detection of marking attributes, and incorrect lane right of way calculation caused by errors in ID mapping of the lane to which obstacles belong when purely relying on intelligent vehicle perception systems and simple feature maps for drivable area detection, this application reduces the risk of misjudgment of decision-making conditions by extracting the geometry and lane and marking attribute information of high-precision maps. At the same time, a high-precision map-guided obstacle mapping method and a method for calculating expected right of way changes developed based on the national standard LDW warning conditions are proposed. This method can not only adapt well to the obstacle lane trajectory-level mapping problem on large curved road sections, but also calculate in real time the changes in road driving conditions caused by the lane changing behavior of obstacle vehicles, thereby improving the intelligence and flexibility of the entire autonomous driving, reducing the negative and uncertain impacts on road traffic, and making autonomous driving more mature and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is a schematic diagram of the lane trajectory-level path planning results guided by high-precision maps.

[0081] Figure 2 It is the lane-level planning result diagram after the topological connection relationship between multiple lanes and the matching of start and end points.

[0082] Figure 3 This is a schematic diagram of lane centerline jump.

[0083] Figure 4 This is the flow chart of the tangential smoothing algorithm for the lane centerline jump.

[0084] Figure 5 It is a schematic diagram of obstacle lateral motion prediction.

[0085] Figure 6 It is a schematic diagram of the cumulative loss of driving efficiency in this application.

[0086] Figure 7 It is an abstract diagram of the intersection decision problem.

[0087] Figure 8 It is a schematic diagram of the local path planning process guided by high-precision maps.

[0088] Fig. 9 It is a schematic diagram of the relationship between the baseline and the generated trajectory clusters.

[0089] Fig.10 It is a schematic diagram of road boundary restriction parameters.

[0090] Fig.11 Schematic diagram of the generated trajectory clusters and comfort costs.

[0091] Fig.12 It is a flow chart for high-precision map surveying and preparation.

[0092] Fig.13 It is a schematic diagram of simulation verification path selection when there are no obstacles.

[0093] Fig.14 It is a schematic diagram of simulation verification path selection when obstacles are added. DETAILED DESCRIPTION

[0094] The following, in conjunction with the accompanying drawings, further describes the technical solution of the high-precision map-guided autonomous driving path planning decision method provided in the present application, so that technical personnel in the field can better understand the present application and implement it.

[0095] With the breakthrough development of AI and big data related technologies in recent years, autonomous driving technology, as a hot technology with significant commercial potential, has attracted more and more attention from the society. The core of autonomous driving technology is to solve how to make the machine have the ability of a driver, including the ability to perceive and recognize the environment, the ability to position and determine its own posture, and the ability to plan and make decisions. High-precision maps are one of the necessary conditions for achieving advanced autonomous driving. At present, autonomous driving is developed using a simplified version of high-precision maps, similar to RNDF road network files, but because the road network files contain very little information, there is still room for further exploration of the application potential of high-precision maps. This application is based on the large amount of rich and accurate road static information provided by real high-precision maps, and based on the navigation application mode, proposes a path planning decision system guided by high-precision maps suitable for smart cars.

[0096] This application first constructs the elements and data model of high-precision maps, and analyzes the data modeling method of high-precision maps through typical map data model examples. Then, the path planning decision system guided by high-precision maps is divided into three layers: the first layer is the lane trajectory level path planning layer, the second layer is the behavior decision layer, and the third layer is the local path planning layer. Each layer of the method analyzes the path planning decision of the high-precision map data in the direction of autonomous driving application, which are:

[0097] (1) Lane trajectory level path planning method guided by high-precision maps: First, the Dijkstra method is improved to perform lane trajectory level path planning based on high-precision maps. Then, the correspondence between roads and lanes in the map is used to reversely calculate the drivable area of ​​the same type of lane expression plan to meet the vehicle's obstacle avoidance and lane change requirements. Then, a virtual lane is constructed through a cosine function to perform tangential smoothing on the planned lane to solve the lane jump problem caused by map division in lane trajectory level planning and meet the requirements of path planning smoothness. Finally, based on the advantage of minimal resource consumption in short-distance map topological relationship calculation, the lane trajectory level path planning method triggering mode includes two items: insufficient planning distance and lane change driving.

[0098] (2) Driving behavior decision-making method guided by high-precision maps: First, the action decision items under structured road surface conditions are analyzed and expressed through the state switching mechanism of the finite state machine. Then, the switching conditions between the various state items of the state machine are analyzed. The important state machine switching conditions include: using the map to map obstacles and predicting the behavior of obstacles through the map. Finally, left and right lane changes are classified as lateral decisions, and the remaining decision items are classified as longitudinal decisions. For lateral decisions, the starting point judgment of lane change intention generation and lane change intention generation and lane change condition judgment are divided into three steps. Among them, the absolute difference between the expected vehicle speed and the current vehicle speed is greater than a certain critical value as the starting point of lane change intention generation, and the lane change intention is generated when the accumulated driving efficiency loss is greater than a certain critical value. Finally, the lane change condition is judged by combining the lane change intention with the map data, and finally a decision is generated.

[0099] (3) Local path planning and selection method guided by high-precision maps: First, the steps of the local path planning method are proposed. Then, the center line of the lane planned at the map lane trajectory level is extracted as the planning baseline, and the boundary of the drivable area is used as the maximum margin of the trajectory cluster. When the planning curve crosses drivable areas of different widths, the narrowest drivable area boundary is used as the maximum margin of the trajectory cluster. Finally, the three different indicators that affect the selection of driving curves, namely decision cost, road cost and comfort cost, are integrated through cost calculation, and the trajectory line is selected by weighted summation instead of the minimum cost method, where the road cost is obtained by extracting the lane line type.

[0100] 1. Lane trajectory level path planning method guided by high-precision map

[0101] The traditional navigation method is that the navigation system calculates the road-level calculation results. When the vehicle drives to a specific area, the navigation system broadcasts the corresponding voice information to remind the driver to change lanes and then change roads. However, in the autonomous driving system, traditional semantic information is not sufficient for vehicles to adopt well. Multiple lanes on the same road lead to different directions. Only the road-level navigation results may cause the autonomous driving vehicle to go to the wrong lane and then the wrong road. Therefore, in order to meet the needs of autonomous driving, data information down to the lane trajectory level and lane trajectory-level navigation calculation results based on data are indispensable.

[0102] 1. Lane trajectory level planning method

[0103] On the basis of the optimal lane planning, a comprehensive plan for multiple suboptimal alternative lanes is obtained. The plan includes three parts. The first part is based on the road-level planning, and the lane trajectory-level topology calculation is performed based on the road-level planning LinkID sequence and the one-to-many inclusion relationship between the road LinkID and the lane LaneID to ensure connectivity; the second part is to completely obtain the lane trajectory-level LaneID sequence to ensure lane connectivity, and then backtrack layer by layer to calculate the same type of lanes using the differences in attributes between lanes; the third part is based on the high-precision map data itself, and the segmentation of the road segment is based on the tangent of the road. The result of the lane trajectory-level planning is likely to cause lateral steps of the lane, so a virtual lane connection is made for the step part to smooth it. Figure 1 The complete lane trajectory level planning result diagram is shown in the figure. The closer to the intersection, the closer the lane planned by the lane trajectory level is to the rightmost lane until it enters the rightmost lane. When the vehicle drives to the interchange area about to exit the ramp, even if there are still several lanes vacant on the left side of the lane planned by the lane trajectory level, they will not be counted as the same lane, because the left lane change is no longer allowed here.

[0104] This application refines lane trajectory level path planning in two aspects:

[0105] (1) Similar lanes are provided. In addition to the lane ID sequence obtained by lane trajectory-level planning, similar lanes are also provided to ensure the needs of local path planning and lane change, and reflect the actual road conditions to the vehicle for reference;

[0106] (2) A tangent point search was performed on the search results and a tangent translation was performed using the cosine function to ensure a smooth connection between the center lines of the two lanes.

[0107] 1. High-precision map lane trajectory level topology calculation

[0108] Through high-precision positioning, the current s_lane and the end point e_lane are matched. Each long-distance planning is completed by road-level planning. Each lane trajectory-level planning only needs to be responsible for calculating the data of the front 1 to 2 km, which greatly reduces the complexity of the calculation.

[0109] Lane trajectory level topology information is provided by Figure 2 (a) It can be seen that when the number of lanes remains unchanged, there is no one-to-many relationship between the lanes in different road sections, and lane changes are not supported in terms of topology. However, in actual planning, lane changes must be achieved by connecting multiple LaneIDs belonging to the same LinkID. Figure 2 shown.

[0110] The lane change itself is converted into a distance cost like the lane length and is involved in the path calculation. The lane cost is the direct lane distance. The lane change cost is estimated, a virtual lane change route is generated, and the distance of the virtual lane change route is calculated and multiplied by a weight:

[0111] Cost=dst*p Formula 1

[0112] Assuming that the length of the virtual lane change route is the product of the driving speed and the lane change time, the default value used in this application is 120km / h, 6s lane change time, that is, 216m, and the weight is adjustable. The default value used in this application is 1.5. Figure 2 (b) shows the lane planning calculation result.

[0113] 2. Smooth lane change

[0114] like Figure 3 As shown in the figure, the result of lane trajectory level topology calculation cannot avoid lane changing, and the need to change lanes will cause tangential translation of the road centerline data.

[0115] This application proposes a solution to generate a virtual lane using a cosine function at the tangential jump of the lane centerline, and to connect two tangentially translated lanes by continuously translating the points at the tangent plane of the existing road section, thereby reducing the strength of the jump. The method flow chart is shown in Figure 4 shown.

[0116] 3. Lane trajectory level path planning trigger mechanism

[0117] (1) The single length of lane trajectory level path planning exceeds 2 km. It is a plan based on static information and has no time-varying characteristics. The first trigger condition is that the vehicle is about to reach the critical value of the result data of a single lane trajectory level path planning, which automatically triggers the lane trajectory level path planning to further plan;

[0118] (2) Lane trajectory level path planning, which derives the recommended lane, plays an important role in the decision-making of the planning module and local path planning. Lane trajectory level path planning adopts virtual lane smoothing for the tangential translation of the recommended lane. Considering the above two reasons, when the vehicle leaves the recommended lane of the current lane trajectory level path planning by changing lanes, the lane trajectory level path planning is automatically triggered.

[0119] 2. Driving behavior decision-making method guided by high-precision maps

[0120] (I) Behavioral decision items based on structured road traffic regulations

[0121] (1) Lane Straightening - Conventional decision, keep driving in the current lane: regardless of the longitudinal content, that is, whether there is an obstacle in the current lane (even if the obstacle speed is 0), the vehicle keeps driving in the lane until a new decision causes a state switch;

[0122] (2) Turn left at the intersection - conventional decision. At the intersection, the following will be encountered: the vehicle will enter a slow operation after the intersection, adapting to the faster planning frequency and perception frequency, which brings more calculation delays. In addition, some other attributes of the turning process will also be gradually added and applied.

[0123] (3) Right turn at intersection - conventional decision, in the case of intersections, the following will be encountered: after the vehicle at the intersection enters the predetermined lane, it no longer judges the traffic light information (unless the right turn lane also has a light attribute), and directly enters the lane change state;

[0124] (4) Go straight at the intersection - conventional decision, in the case of an intersection, the vehicle will enter the intersection and move slowly, adapting to the more computational delay caused by faster planning and perception frequency;

[0125] (5) Parking at an intersection - conventional decision-making. When approaching an intersection, you may encounter: vehicles entering a slow-moving or following state;

[0126] (6) Left lane change—conventional decision, encountered while keeping the lane: the weight of the left road is increased;

[0127] (7) Right lane change - conventional decision, encountered while keeping lane: the weight of the road on the right is increased.

[0128] (8) U-turn—conventional decision: Distinguish U-turn lanes and increase the weight of U-turn lanes.

[0129] (II) Obstacle behavior prediction method based on map lane attribute information

[0130] In addition to predicting the behavior of obstacles through the lane trajectory-level attributes of the HD map, its geometric information is further used to estimate the lateral movement of obstacles and detect sudden lane changes as soon as possible. Figure 5 As shown, the lateral motion estimation of the obstacle vehicle is performed.

[0131] like Figure 5 As shown in the figure, the lateral motion of the vehicle is modeled, the speed of the lane direction and the lane tangent is decomposed according to the vehicle speed, and the tangential speed Vx is obtained; at the same time, the center point of the vehicle is matched with the center point of the lane, and the matching distance D1 is obtained. The tangential distance between the center point of the vehicle and the lane boundary is half the lane width minus D1, and the calculation formula for the lateral crossing lane time is obtained as follows:

[0132]

[0133] V represents the real speed of the vehicle, V x It represents the tangential component of the vehicle speed V on the road, δ represents the angle between the vehicle speed direction and the lane direction, LanePt.Heading represents the lane direction, Vehicle.Heading represents the vehicle's driving direction, LanePt represents the matching point of the vehicle center on the lane centerline, and Width represents the lane width. The time to cross the lane line is predicted to be between 1.5 and 2 seconds for small vehicles and 2.5 seconds for large vehicles. The vehicle is considered to have lateral movement and constraints are imposed on the decision.

[0134] (III) Lateral decision-making of vehicles changing lanes

[0135] Lane change decision affects the lane selection of the vehicle in the tangential direction. The behavior of lateral lane change is an inevitable but high-risk decision during vehicle driving. Lane change decision includes: first, the generation of lane change intention, second, the timing of lane change intention calculation or abandonment conditions, and third, the judgment of lane change decision.

[0136] (1) Generate lane change intention: Based on the description of driving efficiency loss accumulation LCDT, first, the lane change intention is generated when the current vehicle speed is lower than the expected speed; second, this inefficiency needs to accumulate for a period of time; the integral of the time when the current speed is lower than the expected speed is compared with the pre-set critical value to determine whether the generation condition of the lane change intention is met. Figure 6 Shown is a schematic diagram of the cumulative loss of driving efficiency.

[0137] (2) Lane change intention start time or abandonment condition: Determine the time point to start accumulating the driving efficiency loss cumulative value. When the difference between the expected speed and the current speed is greater than the set critical value, the driving efficiency loss cumulative calculation starts. When the difference between the expected speed and the current driving speed is less than the set critical value, the calculated cumulative value is abandoned and reset to zero, waiting for the next start.

[0138] (3) Lane change decision judgment: When the cumulative value of driving efficiency loss is greater than the set critical value, the lane change feasibility judgment is initiated.

[0139] (IV) Speed ​​control longitudinal decision

[0140] Keeping the lane only affects the speed control, such as Figure 7 As shown in the figure, the high-precision map constructs the topological connection relationship at the lane trajectory level for the intersection, and further constructs the virtual lane geometry information based on the logical relationship.

[0141] like Figure 7 Lane L6 is a virtual lane, and the LaneID sequence of lane trajectory level planning is extracted. There is no essential difference from the non-intersection situation. The only difference is that lane L6, as a virtual lane, has some special lane attributes in the intersection situation, but the vehicle has no lane change capability.

[0142] 3. Local Path Planning and Selection Methods Guided by HD Maps

[0143] The discrete optimization method is combined with the prior information of the high-precision map to connect with the high-precision map and obtain an application method of the high-precision map. The specific calculation process is as follows Figure 8 shown.

[0144] 1. Real-time local path trajectory cluster planning

[0145] Fig. 9 This is a schematic diagram of the baseline and the generated trajectory clusters. The trajectory clusters are generated as follows:

[0146] The first step is to obtain the planning base baseline: obtain a center line of the road followed by the trajectory, which serves as the baseline. All trajectory lines in the trajectory cluster use the trajectory center line as the planning baseline, and then generate a series of trajectory lines near the baseline according to the method. The baseline is both the core line of the trajectory cluster and provides the movement direction of the end of the trajectory line. Fig. 9 Shown is a schematic diagram of the relationship between the baseline and trajectory clusters;

[0147] The second step is baseline smoothing: each trajectory is generated to ensure that it is as smooth as possible and the curvature changes linearly. The baseline is composed of a series of hash points, and the trajectory generation needs to be further expressed by a function.

[0148] The third step is to obtain the tangential edge constraint conditions: the tangential divergence of the trajectory cluster is limited effectively, and the boundary is reasonable to ensure the effectiveness of the planning and controllable calculation amount;

[0149] The fourth step is the generation of trajectory clusters: through the lateral deviation function and superimposing it with the baseline, while adjusting the maximum lateral deviation distance, a series of trajectory data with a certain lateral spacing are generated. Each trajectory of these trajectory clusters satisfies the vehicle kinematic constraints and is related to the current vehicle position and posture.

[0150] (II) Selection and extraction of lane centerline based on HD map

[0151] First, planning is performed based on a base baseline of a guiding direction. The baseline is not only to ensure that the vehicle drives on the road and does not deviate from the lane, but also to be guidable and have smooth curvature changes. Each lane line in the high-precision map has high-precision road centerline data to reproduce the curvature of the real road. The lane centerline point column data of the high-precision map replaces the road network point column of RNDF as the baseline for trajectory cluster generation. The lane centerline point column data of the lane ID where the current lane trajectory-level path planning is located can be directly extracted.

[0152] 3. Extracting boundary constraints based on high-precision maps

[0153] like Fig.10 As shown in the figure, the road boundary restriction parameter diagram. The traditional method of using RNDF files to define the maximum side deviation parameter q has the following problems: First, the value of the side deviation parameter q is relatively lagging and cannot change very timely with the change of lanes. Second, during driving, even if some road surfaces are wide enough, these drivable areas are often not places where vehicles can actually drive. When you are about to exit the ramp, it is not recommended to drive on the middle lane or even the overtaking lane on the road section at the ramp. You should change lanes in time to the rightmost lane track-level planning road.

[0154] This application introduces high-precision maps. By using lane trajectory-level path planning guided by high-precision maps, a lane trajectory-level path planning result and similar lanes in each lane that can be used for lane changing and obstacle avoidance are obtained. The acquisition of similar lanes is the maximum deviation of the lateral path planning.

[0155] The specific method is as follows: After extracting the planned lane centerline hash point string each time through the LaneID obtained by the lane trajectory level path planning, the similar lanes of each lane trajectory level planning LaneID are further obtained through the results of the lane trajectory level path planning. Some of these similar lanes are on the left side of the lane trajectory level planning lane, and some are on the right side, but all support lanes for vehicle lane change. The maximum side deviation distance is determined by the similar lane data. Assuming that the width of each normal lane is the same, the calculation of the maximum side deviation distance q is:

[0156] q left / right =NumofBrotherLaneS left / right *Width formula 3

[0157] NumofBrotherLaneS left / right It indicates the number of lanes of the same type, Width indicates the lane width, and the maximum lateral deviation distance q must be the minimum lateral distance within a single planning distance, centered on the lane trajectory level planning.

[0158] (IV) Generate trajectory clusters

[0159] Get the center line of the lane as the base baseline of the trajectory cluster and the vehicle's side deviation distance limit parameter q left and q right After that, the vehicle's drivable area is obtained, and a series of drivable trajectories are generated in this space. The generation of these trajectories ensures that the vehicle has sufficient maneuverability and makes driving behavior safer and more comfortable. The steps for generating trajectory clusters include:

[0160] Step 1: Positioning and matching with the baseline: The vehicle is mapped from the Cartesian coordinate system to the curved coordinate system of the baseline through the matching process, using the heading information of the base baseline;

[0161] Step 2: Determine the length of a single local path planning: determine the distance within which the vehicle needs to correct the heading and offset errors between the current vehicle and the baseline;

[0162] Step 3: Construct the side deviation change function and generate trajectory clusters

[0163] The length of a single path planning consists of two parts. The first part is the adjustment stage of the heading deviation and the lateral distance deviation. A cubic spline curve is used to simulate the curve of the adjustment stage and is solved by the following formula:

[0164]

[0165]

[0166]

[0167] Δs=ss i Formula 7

[0168] Among them, q(s) is the cubic spline function in the adjustment stage, s i represents the mapping point on the baseline obtained by matching the vehicle, s f represents the arc length during the adjustment phase, q f is the lateral offset value of a certain track line in the track cluster, calculated by the same lane; q i It represents the lateral distance between the vehicle and the baseline calculated by matching the current point. a, b, c represent the corresponding parameters, and s represents the arc length variable of the function.

[0169] (V) Comfort cost path selection

[0170] The cost calculation method is adopted to calculate a cost for all the alternative trajectories in the trajectory cluster, normalize the multiple factors affecting the trajectory selection, and finally form a unified standard. The normalization of each factor is to give each factor affecting the cost a quantitative measurement value and assign weights respectively, so as to finally obtain the cost calculation function.

[0171] The comfort of vehicle driving depends on whether the trajectory of local path planning is smooth enough. The radius of curvature of each point on the trajectory. The larger the curvature, the smaller the turning radius of the vehicle and the sharper the turn. In order to comprehensively consider the curvature of the entire trajectory, this application uses the cumulative calculation of the curvature of the entire trajectory line as the comfort cost of the trajectory line. Based on the fact that the influence factor of curvature is exponential, the square of the curvature value is used as the integral condition. Fig.11 The figure shows selecting a straight line segment, generating a trajectory cluster, and then calculating the cost.

[0172]

[0173] In the above formula, k i represents the curvature of the ith point, and s is the arc length along the trajectory;

[0174] Finally, assuming that the three types of costs are cost1, cost2, and cost3, and the factors affecting the final choice are different, a weight should be added to obtain the final cost calculation formula:

[0175] cost=q1*cost1+q2*cost2+q3*cost3 Equation 9

[0176] Get the final cost diagram.

[0177] 4. Experimental testing and analysis

[0178] (I) Test background

[0179] In the early stage, MATLAB simulation was used to verify the path planning method. In the later stage, the method was transplanted to the self-driving car as a verification platform to verify the selected solution. The verification site was selected as the large circular runway in the traffic test field of Beijing Highway Research Institute. The large outer ring of the site is a formal road with four lanes in both directions. Common roads such as intersections and one-way streets are also built in the middle, and a variety of different traffic signs are set up for verification and testing.

[0180] The preparation of high-precision maps of the experimental site used a professional high-precision map collection vehicle to survey the site, and the data was processed in layers and blocks according to the format of high-precision maps, such as Fig.12 .

[0181] (II) Test plan items

[0182] (1) MATLAB simulation verification method: The center line of a lane is randomly extracted from the HD map data, and the local path planning method is verified based on the center line. After the trajectory cluster generation method can be implemented, some obstacles are randomly introduced to check whether the trajectory selection method after cost calculation is feasible. Fig.13 and Fig.14 shown.

[0183] (2) Smart car field verification platform: A test vehicle modified from Dongfeng AX7 was used. The test vehicle used high-precision combined positioning equipment and RTK differential services. The centerline trajectory data of the center lane was collected through manual driving. Then, a dedicated map editing tool was used to perform editing operations such as translation, interruption, attribute addition, and lane division on the centerline data to generate a high-precision map data. The verification was carried out at the Tongzhou Highway Test Site.

[0184] (III) Test analysis

[0185] In the Matlab indoor simulation test, the road baseline is obtained by extracting the center line of a lane in the high-precision map. Fig.13As shown in the figure, the center line generated by the trajectory cluster has a lot of jagged driving after zooming in and is not as smooth as imagined. In order to reduce the amount of lane centerline data, the data extracted from the centerline data of the high-precision map was extracted in the experiment, which further increased the unevenness of the baseline. Since the baseline is not smooth enough, the trajectories of the generated trajectory clusters all have a tendency similar to overfitting, and often cannot well reflect the vehicle running trajectory in real scenarios. In addition, from Fig.14 As shown in the figure, when no obstacles are introduced, the trajectory of the lane line finally selected is basically the trajectory parallel to the center line of the trajectory. This is due to the comprehensive reflection of the comfort cost and the cost of driving along the center line of the trajectory when calculating the cost. Fig.14 As shown, the trajectory is selected correctly after introducing obstacles of different shapes and sizes.

[0186] (IV) Test Summary

[0187] This application conducts simulation and field tests on the path planning decision-making scheme guided by high-precision maps, and verifies the feasibility of high-precision map-assisted planning in terms of process. In the actual process, with the assistance of high-precision maps, the adoption and flexibility of the overall system have been further improved. In particular, for specific scenarios, it is completely achievable to use the assistance of high-precision maps to meet the stability and reliability required by the autonomous driving system under limited conditions.

Claims

1. A high-precision map-guided autonomous driving path planning decision method, characterized in that: First, construct the elements and model data of the high-precision map: Based on the optimal lane planning, obtain multiple suboptimal alternative lane comprehensive solutions. The solution includes three parts. The first part is to perform lane trajectory level topology calculation based on the one-to-many inclusion relationship between road LinkID and lane LaneID on the basis of road-level planning to ensure connectivity; The second part is to obtain the LaneID, ensure the lane connectivity, and then trace back layer by layer to calculate the same type of lanes using the differences in attributes between lanes. The third part is to segment the road segments based on the HD map data itself, and make virtual lane connections for smoothing the step parts according to the tangent direction of the road. The lane trajectory level path planning is refined in two aspects: (1) Constructing similar lanes. In addition to the LaneID obtained by lane trajectory-level planning, similar lanes are also provided to ensure local path planning and lane change requirements, reflecting the actual road conditions to the vehicle; (2) Perform tangent point search on the search results and use the cosine function to perform tangent translation to ensure smooth connection between the center lines of the two lanes; Then, the HD map-guided path planning decision system is divided into three layers: the first layer is the lane trajectory level path planning layer, the second layer is the behavior decision layer, and the third layer is the local path planning layer. Each layer method analyzes the HD map data for the path planning decision in the direction of autonomous driving application. 1) Lane trajectory level path planning method guided by high-precision map: First, the Dijkstra method is improved to perform lane trajectory level path planning for high-precision maps. Then, the correspondence between roads and lanes in the map is used to reversely calculate the drivable area of ​​the same type of lane expression plan to meet the vehicle's obstacle avoidance and lane change requirements. Then, a virtual lane is constructed through a cosine function to perform tangential smoothing on the planned lane to solve the lane jump problem caused by map division in lane trajectory level planning and meet the requirements of path planning smoothness. Finally, based on the advantage of minimal resource consumption in short-distance map topological relationship calculation, a lane trajectory level path planning trigger method is constructed, including two items: insufficient planning distance and lane change driving. High-precision map lane trajectory level topology calculation: Through high-precision positioning, the current lane s_lane and the end lane e_lane are matched. Each long-distance planning is completed by the road-level planning calculation. Each lane trajectory level planning only needs to calculate the data of the first 1 to 2 km. Based on the lane trajectory level topology information, the lane change action must be achieved by connecting multiple lane LaneIDs belonging to the same road LinkID; The lane change itself is the same as the lane length, which is converted into a distance cost and involved in the path calculation. The lane cost is the direct lane distance. The cost of changing lanes is obtained by estimation, generating a virtual lane change route, and calculating the distance of the virtual lane change route and multiplying it by a weight: Cost=dst*p Formula 1 Cost is the cost of changing lanes, dst is the distance of the virtual lane change route, and p is the weight; Smoothing of lane change jumps: A solution that uses a cosine function to generate a virtual lane at the tangential jump of the lane centerline. By continuously translating the points at the tangent plane of the existing road section to connect the two tangentially translated lanes, the strength of the jump is reduced; Lane trajectory level path planning triggering methods include: (a1) When the length of a single lane trajectory level path planning exceeds 2 km, the first trigger condition is that the vehicle is about to reach the critical value of the result data of a single lane trajectory level path planning, which automatically triggers the lane trajectory level path planning to further plan; (a2) When the vehicle leaves the recommended lane of the current lane trajectory level path planning by changing lanes, the lane trajectory level path planning is automatically triggered; 2) Driving behavior decision-making method guided by high-precision maps: First, the action decision items under structured road surface conditions are analyzed and expressed through the state switching mechanism of the finite state machine. Then, the switching conditions between the various state items of the state machine are analyzed. The important state machine switching conditions include: using the map to map obstacles, predicting the behavior of obstacles through the map, and classifying left and right lane changes as lateral decisions, and the remaining decision items as longitudinal decisions; for lateral decisions, the lane change intention generation starting point judgment and lane change intention generation and lane change condition judgment are divided into three steps. Among them, the absolute difference between the expected vehicle speed and the current vehicle speed is greater than a certain critical value as the starting point for the lane change intention generation, and the lane change intention is generated by the accumulated driving efficiency loss greater than a certain critical value. Finally, the lane change condition is judged by combining the lane change intention with the map data, and finally a decision is generated; Action decision items based on structured road traffic regulations: (b1) Lane Straightening - Conventional decision, keep driving in the current lane: Regardless of whether there is an obstacle in the current lane, the vehicle keeps driving in the lane until a new decision causes a state switch; (b2) Left turn at intersection - conventional decision, in the case of intersections, the following will be encountered: after the intersection, the vehicle enters a slow operation, adapting to the faster planning frequency and perception frequency, which brings more calculation delays, and some other attributes of the turning process will gradually be added and applied; (b3) Right turn at intersection - conventional decision, in the case of intersections, the following will be encountered: after the vehicle at the intersection enters the predetermined lane, it no longer judges the traffic light information and directly enters the lane change state; (b4) Go straight at the intersection - conventional decision, in the case of an intersection, the following will be encountered: the vehicle will enter the intersection and move slowly, adapting to the more computational delay caused by faster planning and perception frequency; (b5) Parking at an intersection - conventional decision-making. When approaching an intersection, you will encounter: vehicles entering a slow-moving, following-the-car state; (b6) Left lane change—conventional decision, encountered while keeping the lane: the weight of the left road is increased; (b7) Right lane change - conventional decision, encountered while keeping lane: the weight of the right road is increased; (b8) U-turn driving—conventional decision: distinguish U-turn lanes and increase the weight of U-turn lanes; The obstacle behavior prediction method combined with the lane attribute information of the map is as follows: the speed of the lane direction and the lane tangent is decomposed according to the vehicle speed, and the tangential speed Vx is obtained; at the same time, the center point of the vehicle is matched with the center point of the lane, and the matching distance D1 is obtained. The tangential distance between the center point of the vehicle and the lane boundary is half the lane width minus D1, and the calculation formula for the lateral crossing lane time is obtained as follows: T is the time of lateral lane crossing, V is the actual speed of the vehicle, V x represents the tangential component of the vehicle speed V on the road, δ represents the angle between the vehicle speed direction and the lane direction, LanePt.Heading represents the lane direction, Vehicle.Heading represents the vehicle driving direction, LanePt represents the matching point of the vehicle center on the lane centerline, and Width represents the lane width; the time to cross the lane line is predicted to be between 1.5 and 2 seconds for small vehicles and 2.5 seconds for large vehicles. It is determined that the vehicle has lateral movement and constraints are imposed on the decision; Lateral decision of lane-changing vehicles: Lane-changing decisions include: first, the generation of lane-changing intentions, second, the timing of lane-changing intentions or the conditions for abandoning them, and third, the judgment of lane-changing decisions; (c1) Generate lane change intention: Based on the description of the cumulative driving efficiency loss LCDT, the lane change intention is generated when the current vehicle speed is less than the expected speed; the time integral of the value of the current speed being less than the expected speed is compared with the pre-set critical value to determine whether the generation condition of the lane change intention is met; (c2) Lane change intention start time or abandonment condition: Determine the time point to start accumulating the driving efficiency loss cumulative value. When the difference between the expected speed and the current speed is greater than the set critical value, the driving efficiency loss cumulative calculation starts. When the difference between the expected speed and the current driving speed is less than the set critical value, the calculated cumulative value is abandoned and reset to zero, waiting for the next start; (c3) Lane change decision judgment: When the cumulative value of driving efficiency loss is greater than the set critical value, the lane change feasibility judgment is initiated; 3) Local path planning and selection method guided by high-precision maps: First, the local path is planned, and then the center line of the lane planned at the map lane trajectory level is extracted as the planning baseline, and the boundary of the drivable area is used as the maximum margin of the trajectory cluster. In the case that the planning curve crosses the drivable areas of different widths, the narrowest drivable area boundary is used as the maximum margin of the trajectory cluster. Finally, the decision cost, road cost and comfort cost that affect the driving curve selection are integrated through cost calculation, and the trajectory line is selected by weighted summation instead of the minimum cost method, where the road cost is obtained by extracting the lane line type; Real-time local path trajectory cluster planning: The first step is to obtain the planning base baseline: obtain a center line base baseline of the road followed by the trajectory, which acts as a baseline. All trajectory lines in the trajectory cluster use the trajectory center line as the planning baseline, and then generate a series of trajectory lines near the baseline. The baseline is both the core line of the trajectory cluster and provides the movement direction of the end of the trajectory line; The second step is baseline smoothing: each trajectory is generated to ensure that it is as smooth as possible and the curvature changes linearly. The baseline is composed of a series of hash points, and the trajectory generation needs to be further expressed by a function. The third step is to obtain the tangential edge constraint conditions: the tangential divergence of the trajectory cluster is limited effectively, and the boundary is reasonable to ensure the effectiveness of the planning and controllable calculation amount; Step 4: Generation of trajectory clusters: Through the side deviation function, superimposed with the baseline, and adjusting the maximum side deviation distance, a series of trajectory data with a certain lateral spacing are generated. Each trajectory of these trajectory clusters satisfies the vehicle kinematic constraints and is related to the current vehicle position and posture; Generate trajectory clusters: q left and q right Represent the maximum side deviation distances on the left and right sides respectively, and obtain the center line of the lane as the base baseline of the trajectory cluster and the side deviation distance limit parameter q of the vehicle. left and q right After that, the vehicle's drivable area is obtained, and a series of drivable trajectories are generated in this space. The generation of these trajectories ensures that the vehicle has sufficient maneuverability and makes driving behavior safer and more comfortable. The steps for generating trajectory clusters include: Step 1: Positioning and matching with the baseline: The vehicle is mapped from the Cartesian coordinate system to the curved coordinate system of the baseline through the matching process, using the heading information of the base baseline; Step 2: Determine the length of a single local path planning: determine the distance within which the vehicle needs to correct the heading and offset errors between the current vehicle and the baseline; Step 3: construct the side deviation change function and generate trajectory clusters; The length of a single path planning consists of two parts. The first part is the adjustment stage of the heading deviation and the lateral distance deviation. A cubic spline curve is used to simulate the curve of the adjustment stage and is solved by the following formula: Δs=ss i Formula 7 Among them, q(s) is the cubic spline function in the adjustment stage, s i represents the mapping point on the baseline obtained by matching the vehicle, s f represents the arc length during the adjustment phase, q f is the lateral offset value of a certain track line in the track cluster, calculated by the same lane; q i It represents the lateral distance between the vehicle and the baseline calculated by matching the current point, a, b, c represent the corresponding parameters, and s represents the function arc length variable; Comfort cost path selection: Using the cost calculation method, a cost is calculated for all the alternative trajectories in the trajectory cluster, and multiple factors affecting trajectory selection are normalized to form a unified standard. Normalizing each factor is to give each factor affecting the cost a quantitative measurement value and assign weights respectively, so as to finally obtain the cost calculation function.

2. The high-precision map-guided automatic driving path planning decision method according to claim 1, characterized in that: Extract boundary constraints based on HD maps: HD maps are introduced. After lane trajectory-level path planning guided by HD maps, a lane trajectory-level path planning result and similar lanes in each lane that can be used for lane change and obstacle avoidance are obtained. The acquisition of similar lanes is the maximum deviation of the lateral path planning. The specific method is as follows: After extracting the planned lane centerline hash point string from the lane ID obtained through lane trajectory level path planning each time, the similar lanes of each lane trajectory level planned lane Lane ID are obtained through the results of lane trajectory level path planning. These similar lanes all support lanes for vehicle lane change. The maximum side deviation distance is determined by the similar lane data. Assuming that the width of each normal lane is the same, the maximum side deviation distance q left / right Calculation: q left / right = NumofBrotherLans left / right * Width Formula 3 NumofBrotherLanes left / right Indicates the number of lanes of the same type, Width indicates the lane width, and the maximum side deviation distance q left / right It must be the minimum lateral distance within a single planning distance, centered on the lane trajectory level planning.

3. The high-precision map-guided automatic driving path planning decision method according to claim 1, characterized in that: Comfort cost path selection: The curvature of the entire trajectory is cumulatively calculated and used as the comfort cost of the trajectory. Based on the fact that the influence of curvature is exponential, the square of the curvature value is used as the integration condition: In the above formula, C k [i] is the comfort cost, k i represents the curvature of the ith point, and s is the arc length along the trajectory; Finally, assuming that the three types of costs are cost1, cost2, and cost3, and the factors affecting the final choice are different, a weight should be added to obtain the final cost calculation formula: cost=q1*cost1+q2*cost2+q3*cost3 Equation 9 cost is the final cost, and q1, q2, and q3 are the weights corresponding to the cost.

Images