TWI750762B - Hybrid planniing method in autonomous vehicles and system thereof - Google Patents

Abstract

A hybrid planning method in autonomous vehicles is proposed to plan a best trajectory function of a host vehicle. The hybrid planning method in autonomous vehicles includes a parameter obtaining step, a learning-based scenario deciding step, a learning-based parameter optimizing step and a rule-based trajectory planning step. The parameter obtaining step includes driving a sensing unit to sense a surrounding scenario of the host vehicle to obtain a parameter group to be learned and storing the parameter group to be learned to a memory. The learning-based scenario deciding step includes driving a processing unit to receive the parameter group to be learned from the memory and deciding one of a plurality of scenario categories that matches the surrounding scenario of the host vehicle according to the parameter group to be learned and a learning-based model. The learning-based parameter optimizing step includes driving the processing unit to execute the learning-based model with the parameter group to be learned to generate a key parameter group. The rule-based trajectory planning step includes driving the processing unit to execute a rule-based model with the one of the scenario categories and the key parameter group to plan the best trajectory function. Therefore, the present disclosure can not only process a plurality of multi-dimensional variables, but also be equipped with learning capabilities and conform to the continuity of trajectory planning.

Description

Hybrid decision-making method and system for self-driving cars

The present invention relates to a decision-making method and system for self-driving cars, in particular to a hybrid decision-making method and system for self-driving cars.

The development of self-driving cars has flourished in recent years, and many automakers have invested a lot of resources to prepare for the advent of the self-driving era. They have planned to use driverless cars to operate the transportation system, and have allowed experimental self-driving cars.

Self-driving cars use a variety of sensors, such as Lidar and Radar, to continuously perform large-scale sensing. During driving, the self-driving car needs to use the reference information of the body and the environment as the system input, so as to plan a safe vehicle driving trajectory.

The current decision-making of autonomous vehicles for obstacle avoidance mostly adopts two methods: rule-based model or artificial intelligence-based model (AI-based model). The evaluation is only applicable to the scene within the constraints, and the trajectory based on the learning model will be discontinuous, and the trajectory and speed will not be stable. It can be seen that there is currently a lack of a hybrid decision-making method and system for self-driving cars that can deal with multi-dimensional variables at the same time, has learning ability, high stability, and conforms to the continuity of trajectory planning and vehicle dynamic constraints. seek its solution.

Therefore, the purpose of the present invention is to provide a hybrid decision-making method and system for self-driving cars, which first use a learning-based model to learn driving obstacle avoidance behavior, and then integrate rule-based path planning to construct hybrid decision-making. The trajectory planned by the rule-based model of the hybrid decision through a specific scene category and a specific key parameter is already the optimal trajectory, which can solve the problem that multiple trajectories need to be generated to select one additional screening action in the prior art.

According to an embodiment of the method aspect of the present invention, there is provided a hybrid decision-making method for self-driving cars, which is used to determine the optimal trajectory function of the vehicle. steps, learning-based parameter optimization steps, and rule-based path planning steps. The parameter obtaining step is to drive the sensing unit to sense the surrounding scene of the vehicle to obtain the parameter set to be learned, and store the parameter set to be learned in the memory. The learning-based scene decision-making step is to drive the computing processing unit to receive the to-be-learned parameter set from the memory, and to determine one of the scene types that matches the surrounding scenes of the vehicle according to the to-be-learned parameter set and a learning model from a plurality of scene types . In addition, the learning-based parameter optimization step drives the arithmetic processing unit to execute the learning model-based parameter set to be learned to generate a key parameter set. The rule-based path planning step drives the computing processing unit to execute a rule-based model for one of the scene categories and key parameter groups to plan the optimal trajectory function.

Thereby, the hybrid decision-making method for self-driving cars of the present invention learns the driving obstacle avoidance behavior based on the learning model, and then integrates the rule-based path planning to construct hybrid decision-making, which can not only process multi-dimensional variables at the same time, but also enable the system to learn capability, and comply with the continuity of trajectory planning and vehicle dynamics constraints.

Other examples of the aforementioned embodiments are as follows: the aforementioned parameter set to be learned may include the vehicle road width, relative distance, obstacle length and obstacle lateral distance, wherein the vehicle road width represents the width of the road where the vehicle is located. The relative distance represents the distance between the vehicle and the obstacle. Obstacle length represents the length of the obstacle. The obstacle lateral distance represents the distance of the obstacle from the center line of the lane.

Other examples of the aforementioned embodiments are as follows: the aforementioned parameter obtaining step may include an information sensing step, and the information sensing step includes a vehicle dynamic sensing step, an obstacle sensing step, and a lane sensing step. The vehicle dynamic sensing step is to drive the vehicle dynamic sensing device to locate the current position of the vehicle and the stop line at the intersection according to the map information, and to sense the current heading angle, current speed and current acceleration of the vehicle. The obstacle sensing step is to drive the obstacle sensing device to sense obstacles within a predetermined distance from the vehicle to generate obstacle information corresponding to the obstacles and a plurality of drivable space coordinate points corresponding to the vehicle. The obstacle information includes the current position of the obstacle, the speed of the obstacle and the acceleration of the obstacle. The lane sensing step is to drive the lane sensing device to sense the lane spacing and road curvature of the vehicle.

Other examples of the aforementioned implementation manner are as follows: the aforementioned parameter obtaining step may further include a data processing step, and the data processing step is configured and implemented by an arithmetic processing unit. The data processing step includes a cutting step, and the cutting step is to cut the current position, current heading angle, current speed, current acceleration, obstacle information, etc. The cutting data is generated from the coordinates of the drivable space, the distance between the lane lines of the vehicle and the curvature of the road. In addition, there is a collision time interval between the vehicle and the obstacle, and the vehicle has a yaw rate; when the collision time interval is less than or equal to the preset time interval, the cutting step is started; when the change of the yaw rate is less than or equal to the preset yaw rate When the rate of change is changed, the cropping step is stopped.

Other examples of the aforementioned embodiments are as follows: the aforementioned data processing step may further include a grouping step, and the grouping step is to group the cropping data into plural groups according to the plural preset acceleration ranges and the plural opposite obstacle information. The preset acceleration ranges include a conservative preset acceleration range and a normal preset acceleration range, the opposing obstacle information includes opposing obstacle information and opposing obstacle-free information, and these groups include conservative groups and normal groups . The conservative preset acceleration range and the information without obstacles in the opposite direction correspond to the conservative group, and the normal preset acceleration range and the information with obstacles in the opposite direction correspond to the normal group.

Other examples of the aforementioned embodiments are as follows: the aforementioned data processing step may further include a mirroring step, and the mirroring step mirrors the own vehicle trajectory function along the traveling direction of the vehicle according to each scene category to generate the mirrored own vehicle trajectory function, The parameter group to be learned includes the vehicle trajectory function after mirroring.

Other examples of the aforementioned embodiments are as follows: the aforementioned learning-based parameter optimization step may include a learning-based driving behavior generating step and a key parameter generating step, wherein the learning-based driving behavior generating step is based on learning a set of parameters to be learned based on a learning model. A learned behavior parameter group is generated. The parameter group to be learned includes a driving route parameter group and a driving acceleration/deceleration behavior parameter group. The key parameter generation step is to calculate the system actuation parameter group of the learned behavior parameter group to obtain the system actuation time point, and to calculate the system actuation time point, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, and the speed of the vehicle. and the target speed combination to form a key parameter group.

Other examples of the aforementioned embodiments are as follows: the aforementioned learned behavior parameter set may include a system action parameter set, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, the vehicle speed and the target speed. The system action parameter group includes vehicle speed, vehicle acceleration, steering wheel angle, yaw rate, relative distance and obstacle lateral distance.

Other examples of the aforementioned embodiments are as follows: the aforementioned optimal trajectory function may include a plane coordinate curve equation, tangential velocity and tangential acceleration, wherein the plane coordinate curve equation represents the optimal trajectory of the vehicle in plane coordinates. The tangential speed represents the speed of the vehicle at the tangent point of the plane coordinate curve equation. Tangent acceleration represents the acceleration of the vehicle at the tangent point. The optimal trajectory function is updated according to the sampling time of the arithmetic processing unit.

Other examples of the foregoing implementation manner are as follows: the foregoing scene categories may include obstacle occupation scenes, intersection scenes, and entry and exit scenes. The obstacle occupied scene includes the obstacle occupied percentage, the obstacle occupied scene represents that there are obstacles and roads in the surrounding scene, and the obstacle occupied percentage represents the proportion of the obstacle occupied by the road. The intersection scene represents that there are intersections in the surrounding scene. Inbound and outbound scenarios represent surrounding scenes with inbound and outbound stations.

According to one embodiment of the structural aspect of the present invention, a hybrid decision-making system for self-driving cars is provided, which is used to decide the optimal trajectory function of the own vehicle. The hybrid decision-making system of the self-driving car includes a sensing unit, a memory and an arithmetic processing unit, wherein the sensing unit is used for sensing the surrounding scene of the vehicle to obtain a set of parameters to be learned. The memory is used to access the parameter set to be learned, the plural scene categories, the learning-based model and the rule-based model. In addition, the arithmetic processing unit is electrically connected to the memory and the sensing unit, and the arithmetic processing unit is configured to implement operations including the following steps: a learning-based scene decision step, a learning-based parameter optimization step, and a rule-based path planning step. The learning-based scene decision-making step is to determine one of the scene categories matching the surrounding scenes of the vehicle from the scene categories according to the parameter set to be learned and the learning model. The learning-based parameter optimization step is to execute the learning model-based parameter set to be learned to generate a key parameter set. The rule-based path planning step is to plan an optimal trajectory function by executing a rule-based model on one of the scene categories and key parameter groups.

Thereby, the hybrid decision-making system for self-driving cars of the present invention uses the learning-based model to learn the driving obstacle avoidance behavior, and then integrates the rule-based path planning to construct hybrid decision-making, which can simultaneously process multi-dimensional variables and enable the system to learn capability, and comply with the continuity of trajectory planning and vehicle dynamics constraints.

Other examples of the aforementioned embodiments are as follows: the aforementioned parameter set to be learned may include the vehicle road width, relative distance, obstacle length and obstacle lateral distance, wherein the vehicle road width represents the width of the road where the vehicle is located. The relative distance represents the distance between the vehicle and the obstacle. Obstacle length represents the length of the obstacle. The obstacle lateral distance represents the distance of the obstacle from the center line of the lane.

Other examples of the aforementioned embodiments are as follows: the aforementioned memory can be used to access a map information, and the map information is related to the route traveled by the vehicle. The sensing unit includes a vehicle dynamic sensing device, an obstacle sensing device and a lane sensing device, wherein the vehicle dynamic sensing device locates the current position of the vehicle according to the map information, and senses the current heading angle, current speed and current acceleration. The obstacle sensing device senses obstacles within a predetermined distance from the vehicle to generate obstacle information corresponding to the obstacle and a plurality of drivable space coordinate points corresponding to the vehicle. The obstacle information includes the current position of the obstacle, the speed of the obstacle and the acceleration of the obstacle. The lane sensing device senses the distance between the lane lines of the vehicle and the curvature of the road.

Other examples of the aforementioned embodiments are as follows: the aforementioned arithmetic processing unit is configured to perform a data processing step, the data processing step including a cutting step. This cutting step is to cut the current position, current heading angle, current speed, current acceleration, obstacle information, these drivable space coordinate points, the current position of the vehicle according to the preset time interval and the preset yaw rate of change. Lane line spacing and road curvature to generate clipping data. In addition, there is a collision time interval between the vehicle and the obstacle, and the vehicle has a yaw rate. When the collision time interval is less than or equal to the preset time interval, the cutting step is started; when the change of the yaw rate is less than or equal to the preset yaw change rate, the cutting step is stopped.

Other examples of the aforementioned embodiments are as follows: the aforementioned data processing step may further include a grouping step, and the grouping step is to group the cropping data into plural groups according to the plural preset acceleration ranges and the plural opposite obstacle information. The preset acceleration ranges include a conservative preset acceleration range and a normal preset acceleration range, the opposite obstacle information includes a pair of information with obstacles in the direction and a pair of information without obstacles, and the groups include a Conservative groups and a normal group. The conservative preset acceleration range and the information without obstacles in the opposite direction correspond to the conservative group, and the normal preset acceleration range and the information with obstacles in the opposite direction correspond to the normal group.

Other examples of the aforementioned embodiments are as follows: the aforementioned data processing step may further include a mirroring step, which mirrors the own vehicle trajectory function along the vehicle traveling direction according to each scene category to generate the mirrored own vehicle trajectory function. The parameter group to be learned includes the vehicle trajectory function after mirroring.

Other examples of the aforementioned embodiments are as follows: the aforementioned learning-based parameter optimization step may include a learning-based driving behavior generating step and a key parameter generating step, wherein the learning-based driving behavior generating step is based on learning a set of parameters to be learned based on a learning model. A learned behavior parameter group is generated. The parameter group to be learned includes a driving route parameter group and a driving acceleration/deceleration behavior parameter group. The key parameter generation step is to calculate the system operation parameter group of the learned behavior parameter group to obtain the system operation time point, and calculate the system operation time point, the vertical distance of the target point, the horizontal distance of the target point, the curvature of the target point, the speed of the vehicle and the The target speed combination forms the key parameter group.

Other examples of the aforementioned embodiments are as follows: the aforementioned learned behavior parameter set may include a system action parameter set, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, the vehicle speed and the target speed. The system action parameter group includes vehicle speed, vehicle acceleration, steering wheel angle, yaw rate, relative distance and obstacle lateral distance.

Other examples of the aforementioned embodiments are as follows: the aforementioned optimal trajectory function may include a plane coordinate curve equation, tangential velocity and tangential acceleration, wherein the plane coordinate curve equation represents the optimal trajectory of the vehicle in plane coordinates. The tangential speed represents the speed of the vehicle at the tangent point of the plane coordinate curve equation. Tangent acceleration represents the acceleration of the vehicle at the tangent point. The optimal trajectory function is updated according to the sampling time of the arithmetic processing unit.

Other examples of the aforementioned embodiments are as follows: the aforementioned scene categories may include obstacle occupied scenes, intersection scenes, and entry and exit scenes, wherein the obstacle occupied scene includes the obstacle occupied percentage, and the obstacle occupied scene represents the surrounding scene with obstacles and roads. , the obstacle occupancy percentage represents the proportion of the road occupied by obstacles. The intersection scene represents that there are intersections in the surrounding scene. Inbound and outbound scenarios represent surrounding scenes with inbound and outbound stations.

Several embodiments of the present invention will be described below with reference to the drawings. For the sake of clarity, many practical details are set forth in the following description. It should be understood, however, that these practical details should not be used to limit the invention. That is, in some embodiments of the present invention, these practical details are unnecessary. In addition, for the purpose of simplifying the drawings, some well-known and conventional structures and elements will be shown in a simplified and schematic manner in the drawings; and repeated elements may be denoted by the same reference numerals.

In addition, when a certain element (or unit or module, etc.) is "connected" to another element herein, it may mean that the element is directly connected to another element, or it may also mean that a certain element is indirectly connected to another element , that is, there are other elements interposed between said element and another element. When it is expressly stated that an element is "directly connected" to another element, it means that no other element is interposed between the element and the other element. The terms first, second, third, etc. are only used to describe different elements, and do not limit the elements themselves. Therefore, the first element can also be renamed as the second element. And the combination of elements/units/circuits in this article is not a commonly known, conventional or well-known combination in this field, and it cannot be determined whether the combination relationship of the elements/units/circuits is well-known or not easily understood by those in the technical field. Usually the knowledgeable can do it easily.

Please refer to FIG. 1. FIG. 1 is a schematic flowchart of a hybrid decision-making method 100 for self-driving cars according to a first embodiment of the present invention. The hybrid decision-making method 100 for self-driving cars is used to determine the optimal trajectory function 108 of the vehicle. The hybrid decision-making method 100 for self-driving cars includes a parameter obtaining step S02, an AI-based scene decision step S04, and a learning-based decision-making step S04. The parameter optimization step S06 and the rule-based path planning step S08.

The parameter obtaining step S02 is to drive the sensing unit to sense the surrounding scene of the vehicle to obtain the parameter group 102 to be learned, and store the parameter group 102 to be learned in the memory. The learning-based scene decision step S04 is to drive the arithmetic processing unit to receive the to-be-learned parameter set 102 from the memory, and to discriminate from the plurality of scene categories 104 according to the to-be-learned parameter set 102 and an AI-based model. One of the scene categories 104 of the surrounding scenes of the vehicle. In addition, the learning-based parameter optimization step S06 drives the arithmetic processing unit to execute the learning model-based parameter set 102 to be learned to generate the key parameter set 106 . The rule-based path planning step S08 drives the arithmetic processing unit to execute a rule-based model on one of the scene categories 104 and the key parameter group 106 to plan the optimal trajectory function 108 . Thereby, the hybrid decision-making method 100 for self-driving cars of the present invention learns the driving obstacle avoidance behavior based on the learning model, and then integrates the rule-based path planning to construct hybrid decision-making, which can not only process multi-dimensional variables at the same time, but also enable the system to have Learning ability, and comply with the continuity of trajectory planning and vehicle dynamic constraints. The details of the above steps will be described below through more detailed embodiments.

Please refer to FIG. 2 to FIG. 9 together, wherein FIG. 2 is a schematic flowchart of the hybrid decision-making method 100a for a self-driving car according to the second embodiment of the present invention; A schematic diagram of the information sensing step S122 of the hybrid decision-making method 100a; FIG. 4 is a schematic diagram of the input data and output data 101 of the information sensing step S122 of the hybrid decision-making method 100a of the self-driving car in FIG. 2; FIG. 5 is a schematic diagram Figure 2 is a schematic diagram of the data processing step S124 of the hybrid decision-making method 100a for self-driving cars; Figure 6 is a schematic diagram illustrating the application of the hybrid decision-making method 100a for self-driving cars in Figure 2 to avoid obstacles in the same lane; Figure 7 Fig. 2 is a schematic diagram of the hybrid decision-making method 100a of the self-driving car applied to an obstacle occupancy scene; Fig. 8 is a schematic diagram illustrating the application of the hybrid decision-making method 100a of the self-driving car in Fig. 2 to lane changing; and Fig. 9 The figure is a schematic diagram of the rule-based path planning step S18 of the hybrid decision-making method 100a of the self-driving car in FIG. 2; as shown in the figure, the hybrid decision-making method 100a of the self-driving car is used to decide the optimal trajectory function of the HV of the vehicle 108. The hybrid decision-making method 100a for self-driving cars includes a parameter acquisition step S12, a learning-based scenario decision step S14, a learning-based parameter optimization step S16, a rule-based path planning step S18, a diagnosis step S20, and a control step S22.

The parameter obtaining step S12 is to drive the sensing unit to sense the surrounding scene of the HV of the vehicle to obtain the parameter group 102 to be learned, and store the parameter group 102 to be learned in the memory. In detail, the parameter group 102 to be learned includes the road width LD of the own vehicle, the relative distance RD , the length of the obstacle L _obj , and the lateral distance D _{obj of the} obstacle. The own vehicle road width LD represents the width of the road where the own vehicle HV is located. The relative distance RD represents the distance between the vehicle HV and the obstacle Obj. The obstacle length L _obj represents the length of the obstacle Obj. The obstacle lateral distance D _obj represents the distance of the obstacle Obj from the center line of the lane. Furthermore, the parameter obtaining step S12 includes an information sensing step S122 and a data processing step S124.

The information sensing step S122 includes a vehicle dynamic sensing step S1222 , an obstacle sensing step S1224 and a lane sensing step S1226 . The vehicle dynamic sensing step S1222 is to drive the vehicle dynamic sensing device to locate the current position of the vehicle HV and the stop line at the intersection according to the map information, and sense the current heading angle, current speed and current acceleration of the vehicle HV. The obstacle sensing step S1224 is to drive the obstacle sensing device to sense the obstacle Obj within a predetermined distance from the vehicle HV, so as to generate obstacle information corresponding to the obstacle Obj and a plurality of drivable spaces corresponding to the vehicle HV Coordinate point. The obstacle information includes the current position of the obstacle corresponding to the obstacle Obj, the speed of the obstacle v _obj and the acceleration of the obstacle. The lane sensing step S1226 is to drive the lane sensing device to sense the lane line spacing and road curvature of the vehicle. In addition, it can be seen from FIG. 4 that the input data of the information sensing step S122 includes map information, Global Positioning System (GPS) data, image data, Lidar data, Radar data, and inertial measurement Unit (Inertial Measurement Unit, IMU) information. The output data 101 includes the current position, the current heading angle, the intersection stop line, the current position of the obstacle, the speed of the obstacle v _obj , the acceleration of the obstacle, the coordinates of the drivable space, the distance between the lanes of the vehicle and the curvature of the road.

，g代表重力加速度，但本發明不以此為限。藉此，分群步驟S1244之目的在於區分駕駛行為的差異(保守或者常態)，其可改善後續之基於學習模型所訓練的成效。此外，分群步驟S1244可方便切換模型或參數，且能讓系統在可執行之範圍內切換加速程度，或者避開障礙物Obj。另外，鏡射步驟S1246係依據各場景類別104沿一車輛行進方向(例如：Y軸)將一本車軌跡函數鏡射而產生一鏡射後本車軌跡函數。待學習參數組102包含鏡射後本車軌跡函數。本車軌跡函數為本車HV行駛之軌跡，其代表駕駛行為資料。藉此，鏡射步驟S1246之本車軌跡函數與鏡射後本車軌跡函數均可供後續之基於學習模型訓練，以增加蒐集資料的多樣性，進而避免基於學習模型因為資料多樣性不足而無法有效分辨場景類別104的問題。 The data processing step S124 is configured and implemented by an arithmetic processing unit. The data processing step S124 includes a cropping step S1242 , a grouping step S1244 and a mirroring step S1246 . The cutting step S1242 is to cut the current position, current heading angle, current speed, current acceleration, obstacle information, these drivable space coordinate points, A clipping data is generated based on the distance between the lane lines of the vehicle and the curvature of the road. There is a collision time interval between the vehicle HV and the obstacle Obj, and the vehicle HV has a yaw rate; when the collision time interval is less than or equal to the preset time interval, the cutting step S1242 is started; when the yaw rate changes When it is less than or equal to the preset yaw change rate, the cropping step S1242 is stopped. The above preset time interval can be 3 seconds, the preset yaw rate of change can be 0.5, and the change of yaw rate can be comprehensively judged for multiple consecutive data (for example: the change of yaw rate of 5 consecutive records is less than or equal to 0.5), but the present invention is not limited to this. In addition, the grouping step S1244 is to group the cropping data into a plurality of groups according to a plurality of predetermined acceleration ranges and a plurality of opposite obstacle information, and the predetermined acceleration ranges include a conservative predetermined acceleration range and a normal predetermined acceleration The information on the opposite obstacles includes a pair of information with obstacles and a pair of information without obstacles, and these groups include a conservative group and a normal group. The conservative preset acceleration range and the information without obstacles in the opposite direction correspond to the conservative group, while the normal preset acceleration range and the information with obstacles in the opposite direction correspond to the normal group. The above conservative preset acceleration range may be -0.1g to 0.1g, and the normal preset acceleration range may be -0.2g to -0.3g and 0.2g to 0.3g, namely

, g represents the gravitational acceleration, but the present invention is not limited to this. Therefore, the purpose of the grouping step S1244 is to distinguish differences in driving behaviors (conservative or normal), which can improve the results of subsequent training based on the learning model. In addition, the grouping step S1244 can facilitate the switching of models or parameters, and enables the system to switch the acceleration level within an executable range, or avoid obstacles Obj. In addition, the mirroring step S1246 is to mirror a vehicle trajectory function along a vehicle traveling direction (eg, Y axis) according to each scene category 104 to generate a mirrored vehicle trajectory function. The to-be-learned parameter group 102 includes the vehicle trajectory function after mirroring. The vehicle trajectory function is the trajectory of the vehicle HV driving, which represents the driving behavior data. In this way, both the own vehicle trajectory function in the mirroring step S1246 and the own vehicle trajectory function after mirroring can be used for subsequent training based on the learning model, so as to increase the diversity of collected data, thereby avoiding the failure of the learning model based on insufficient data diversity. The problem of effectively distinguishing scene categories 104.

%，三分之一車身為0.7m)；第三場景1043代表障礙物Obj有二分之一車身佔據車道(障礙物佔據百分比=50%，二分之一車身為1.05m)；第四場景1044代表障礙物Obj有三分之二車身佔據車道(障礙物佔據百分比=66.

%，三分之二車身為1.4m)；第五場景1045代表障礙物Obj全車身佔據車道(障礙物佔據百分比=100%，全車身為2.1m)。此外，路口場景代表周圍場景中具有一路口。當其中一個場景類別104為路口時，車輛動態感測裝置透過地圖資訊得到一路口停止線。再者，進出站場景代表周圍場景中具有一進出站。藉此，基於學習之場景決策步驟S14可得到符合周圍場景之場景類別104，以供後續之基於規則之路徑規劃步驟S18使用。 The learning-based scene decision step S14 is to drive the arithmetic processing unit to receive the to-be-learned parameter set 102 from the memory, and to determine the surrounding scene that matches the HV of the vehicle from the plurality of scene categories 104 according to the to-be-learned parameter set 102 and the learning model-based One of the scene categories 104. In detail, the learning-based model is based on the probability statistics method and is trained by collecting real driving behavior data, which may include end-to-end or sampling based planning. The scene category 104 may include an obstacle occupied scene, an intersection scene, and an entry and exit scene, wherein the obstacle occupied scene includes an obstacle occupied percentage, the obstacle occupied scene represents that there are obstacles Obj and roads in the surrounding scene, and the obstacle occupied percentage represents an obstacle Object Obj occupies the proportion of the road. Taking FIG. 7 as an example, the scene category 104 is an obstacle occupied scene, which may include a first scene 1041 , a second scene 1042 , a third scene 1043 , a fourth scene 1044 and a fifth scene 1045 . The first scene 1041 represents that the obstacle Obj does not occupy the lane (the percentage of obstacle occupancy=0%); the second scene 1042 represents that the obstacle Obj has one-third of the vehicle body occupying the lane (the percentage of obstacle occupancy=33.3.

%, one-third of the body is 0.7m); the third scene 1043 represents that the obstacle Obj has one-half of the body occupying the lane (obstacle occupancy percentage = 50%, and one-half of the body is 1.05m); the fourth scene 1044 means that the obstacle Obj has two-thirds of the vehicle body occupying the lane (obstacle occupancy percentage = 66.

%, two-thirds of the body is 1.4m); the fifth scene 1045 represents that the obstacle Obj occupies the lane with the whole body (the percentage of obstacle occupancy=100%, and the whole body is 2.1m). Furthermore, the intersection scene represents that there is an intersection in the surrounding scene. When one of the scene categories 104 is an intersection, the vehicle dynamic sensing device obtains a stop line at the intersection through the map information. Furthermore, the inbound and outbound scene represents that there is an inbound and outbound station in the surrounding scene. In this way, the learning-based scene decision step S14 can obtain the scene category 104 that matches the surrounding scene, which can be used in the subsequent rule-based path planning step S18.

The learning-based parameter optimization step S16 is to drive the arithmetic processing unit to execute the learning model-based parameter set 102 to be learned to generate the key parameter set 106 . In detail, the learning-based parameter optimization step S16 includes a learning-based driving behavior generating step S162 and a key parameter generating step S164, wherein the learning-based driving behavior generating step S162 is generated based on the learning model-based learning of the to-be-learned parameter set 102 A learned behavior parameter group 103 . The learned behavior parameter group 103 includes a system action parameter group, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, the vehicle speed v _h and the target speed. The driving route parameter set ( _xi , y _j ) and the driving acceleration/deceleration behavior parameter set can be obtained through the information sensing step S122 ; in other words, the parameter set 102 to be learned includes the driving route parameter set ( _xi , y _j ) and the driving Acceleration and deceleration behavior parameter group. Furthermore, the key parameter generation step S164 is to calculate the system operation parameter group of the learned behavior parameter group 103 to obtain a system operation time point, and calculate the system operation time point, the vertical distance of the target point, the horizontal distance of the target point, and the target point. The curvature, host vehicle speed v _h and target speed combine to form a key parameter group 106 . The system action parameter group includes the vehicle speed v _h , the vehicle acceleration, the steering wheel angle, the yaw rate, the relative distance RD and the obstacle lateral distance D _obj .

The rule-based path planning step S18 drives the arithmetic processing unit to execute one of the scene categories 104 and the key parameter group 106 to plan the optimal trajectory function 108 based on the rule model. Specifically, one of the scene categories 104 corresponds to the current surrounding scene of the HV of the vehicle. A rule-based model formulates rules based on deterministic behavior, and the decision results depend on sensor information, including polynomials or interpolation curves. Furthermore, the rule-based path planning step S18 includes a target point generation step S182, a coordinate conversion step S184, and a trajectory generation step S186. The target point generation step S182 is to drive the operation processing unit to generate a plurality of target points TP according to the operation of the scene type 104 and the key parameter group 106 . The coordinate conversion step S184 is to drive the arithmetic processing unit to convert the target points TP into a plurality of target two-dimensional coordinates according to the drivable space coordinate points. The trajectory generation step S186 is to drive the arithmetic processing unit to connect the two-dimensional coordinates of the targets to generate the optimal trajectory function 108 . Taking Fig. 9 as an example, the target point generation step S182 will generate three target points TP, and then the coordinate conversion step S184 will generate three target two-dimensional coordinates corresponding to the three target points TP, and finally the trajectory generation step S186 is based on the target two-dimensional coordinates. The coordinates produce the optimal trajectory function 108 . In addition, the optimal trajectory function 108 includes the plane coordinate curve equation BTF, the tangential velocity and the tangential acceleration, wherein the plane coordinate curve equation BTF represents the optimal trajectory of the vehicle HV in the plane coordinates, that is, the coordinate equation of the optimal trajectory function 108 . The tangential speed represents the speed of the vehicle HV at the tangent point of the plane coordinate curve equation BTF. The tangential acceleration represents the acceleration of the vehicle HV at the tangent point. It is also worth mentioning that the to-be-learned parameter set 102 can be updated according to the sampling time of the operation processing unit, and then the optimal trajectory function 108 can be updated; in other words, the optimal trajectory function 108 can be updated according to the sampling time of the operation processing unit.

The diagnosis step S20 is to diagnose whether the future driving trajectory of the vehicle HV and the current surrounding scene (for example: the current road curvature, the distance between the vehicle lanes or the relative distance RD ) are all maintained within a safe tolerance error, and generate a diagnosis result by This judges whether the route of automatic driving is safe, and at the same time, the judgment equation can also directly judge the parameters that need to be corrected in the future driving trajectory and make corrections to improve the safety of automatic driving.

The control step S22 is to control the automatic driving parameters of the HV of the vehicle according to the diagnosis result, the details of which are known in the art, and are not repeated here.

Thereby, the hybrid decision-making method 100a for self-driving cars of the present invention learns the driving obstacle avoidance behavior based on the learning model, and then integrates the rule-based path planning to construct hybrid decision-making, which can simultaneously process multi-dimensional variables and enable the system to have Learning ability, and comply with the continuity of trajectory planning and vehicle dynamic constraints.

Please refer to FIG. 2 to FIG. 10 together, wherein FIG. 10 is a block diagram illustrating a hybrid decision-making system 200 for self-driving cars according to a third embodiment of the present invention. The hybrid decision-making system 200 of the self-driving car is used to determine the optimal trajectory function 108 of the HV of the vehicle. The hybrid decision-making system 200 of the self-driving car includes a sensing unit 300 , a memory 400 and an arithmetic processing unit 500 .

The sensing unit 300 is used for sensing the surrounding scene of the HV of the vehicle to obtain a parameter set 102 to be learned. Specifically, the sensing unit 300 includes a vehicle dynamic sensing device 310 , an obstacle sensing device 320 and a lane sensing device 330 , wherein the vehicle dynamic sensing device 310 , the obstacle sensing device 320 and the lane sensing device 330 are all Installed on the HV of the vehicle. The vehicle dynamic sensing device 310 locates the current position of the vehicle HV according to the map information, and senses the current heading angle, current speed and current acceleration of the vehicle HV. The vehicle dynamic sensing device 310 includes a GPS, a gyroscope (Gyroscope), an odometer (Odemeter), a speed meter (Speed Meter) and an inertial measurement unit (IMU). Furthermore, the obstacle sensing device 320 senses an obstacle Obj within a predetermined distance from the vehicle HV to generate obstacle information corresponding to the obstacle Obj and a plurality of drivable space coordinate points corresponding to the vehicle HV. The obstacle information includes the current position of the obstacle, the speed of the obstacle and the acceleration of the obstacle corresponding to the obstacle Obj. In addition, the lane sensing device 330 senses the distance between the lane lines of the vehicle and the curvature of the road. The above-mentioned obstacle sensing device 320 and lane sensing device 330 include Lidar, Radar and camera. The details of its structure are known in the prior art, so they will not be repeated here.

The memory 400 is used to access the to-be-learned parameter set 102 , the plural scene categories 104 , the learning model and the rule-based model, and the memory 400 is used to access map information, which is related to the route driven by the HV of the vehicle.

The arithmetic processing unit 500 is electrically connected to the memory 400 and the sensing unit 300. The arithmetic processing unit 500 is configured to implement the hybrid decision-making methods 100 and 100a of the self-driving car, and it can be a microprocessor or an electronic control unit (ECU; ECU). ), computer, mobile device or other computing processor.

Thereby, the hybrid decision-making system 200 for self-driving cars of the present invention uses the learning model to learn the driving obstacle avoidance behavior, and then integrates the rule-based path planning to construct a hybrid decision, which can deal with multi-dimensional variables at the same time and has the ability to learn. And comply with the vehicle dynamic constraints and the continuity of trajectory planning.

It can be seen from the above-mentioned embodiments that the present invention has the following advantages: firstly, by learning the driving obstacle avoidance behavior based on the learning model, and then integrating the path planning based on the rules to construct a mixed decision-making, it can not only deal with multi-dimensional variables at the same time, but also has the ability to learn capability, and comply with the continuity of trajectory planning and vehicle dynamics constraints. Second, the trajectory planned by the rule-based model through specific scene categories and specific key parameters is already the best trajectory, which can solve the problem of generating multiple trajectories to select one additional screening action in the prior art. Third, the parameter set to be learned can be updated at any time according to the sampling time of the arithmetic processing unit, and then the optimal trajectory function can be updated at any time, thereby greatly improving the safety and practicability of automatic driving.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be determined by the scope of the appended patent application.

100, 100a: Hybrid decision-making method for self-driving cars 101: Output data 102: Parameter group to be learned 103: Learning behavior parameter group 104: Scene category 1041: First scene 1042: Second scene 1043: Third scene 1044: Fourth scene 1045: Fifth scene 106: Key parameter group 108: Best trajectory function 200: Hybrid decision-making system for self-driving cars 300: Sensing unit 310: Vehicle dynamic sensing device 320: Obstacle sensing device 330: Lane sensing device 400 : memory 500: arithmetic processing unit S02, S12: parameter acquisition steps S04, S14: learning-based scene decision-making steps S06, S16: learning-based parameter optimization Steps S08, S18: rule-based path planning Step S122: information sensing Step S1222: Vehicle Dynamic Sensing Step S1224: Obstacle Sensing Step S1226: Lane Sensing Step S124: Data Processing Step S1242: Cropping Step S1244: Grouping Step S1246: Mirroring Step S162: Learning-Based Driving Behavior Generation Step S164 : key parameter generation step S182: target point generation step S184: coordinate conversion step S186: trajectory generation step S20: diagnosis step S22: control step BTF: plane coordinate curve equation D _obj : obstacle lateral distance HV: own vehicle LD : this Road width L _obj : obstacle length Obj: obstacle RD : relative distance TP: target point v _h : own vehicle speed v _obj : obstacle speed x _i , y _j : driving route parameter group

FIG. 1 is a schematic flowchart of a hybrid decision-making method for self-driving cars according to a first embodiment of the present invention; FIG. 2 is a schematic flowchart of a hybrid decision-making method for self-driving cars according to a second embodiment of the present invention; FIG. 3 is a schematic diagram illustrating the information sensing steps of the hybrid decision-making method of the self-driving car of FIG. 2; FIG. 4 is a schematic diagram showing the input data and output data of the information sensing step of the hybrid decision-making method of the self-driving car of FIG. 2; FIG. 5 is a schematic diagram showing the data processing steps of the hybrid decision-making method of the self-driving car of FIG. 2; FIG. 6 is a schematic diagram illustrating the application of the hybrid decision-making method of the self-driving car in FIG. 2 to avoid obstacles in the same lane; FIG. 7 is a schematic diagram illustrating the application of the hybrid decision-making method of the self-driving car in FIG. 2 to an obstacle occupation scene; FIG. 8 is a schematic diagram illustrating the application of the hybrid decision-making method of the self-driving car in FIG. 2 to lane changing; FIG. 9 is a schematic diagram illustrating the rule-based path planning steps of the hybrid decision-making method of the self-driving car of FIG. 2; and FIG. 10 is a block diagram illustrating a hybrid decision-making system for self-driving cars according to a third embodiment of the present invention.

100: A Hybrid Decision-Making Approach for Self-Driving Vehicles

102: Parameter group to be learned

104: Scene Category

106:Key parameter group

108: Best Trajectory Function

S02: Parameter acquisition step

S04: Scenario decision-making steps based on learning

S06: Parameter optimization steps based on learning

S08: Rule-based path planning steps

Claims (20)

A hybrid decision-making method for self-driving cars is used to decide an optimal trajectory function for a car. The hybrid decision-making method for self-driving cars includes the following steps: a parameter obtaining step, driving a sensing unit to sense a surrounding scene of the vehicle to obtain a parameter set to be learned, and storing the parameter set to be learned in a memory; A learning-based scene decision-making step, which drives an arithmetic processing unit to receive the to-be-learned parameter set from the memory, and discriminates from a plurality of scene categories according to the to-be-learned parameter set and a learning-based model, which is suitable for the vehicle. one of the scene categories of the surrounding scenes; a learning-based parameter optimization step, which drives the operation processing unit to execute the learning-based model on the to-be-learned parameter set to generate a key parameter set; and A rule-based path planning step drives the arithmetic processing unit to execute a rule-based model on the one of the scene categories and the key parameter group to plan the optimal trajectory function. The hybrid decision-making method for self-driving cars according to claim 1, wherein the parameter group to be learned includes: The width of a vehicle road represents one of the widths of the road where the vehicle is located; A relative distance, representing a distance between the vehicle and an obstacle; an obstacle length, representing a length of the obstacle; and A lateral distance of an obstacle, representing a distance of the obstacle from the centerline of a lane. The hybrid decision-making method for self-driving cars according to claim 1, wherein the parameter obtaining step comprises: An information sensing step, including: A vehicle dynamic sensing step is to drive a vehicle dynamic sensing device to locate a current position of the vehicle and a stop line at an intersection according to a map information, and sense a current heading angle, a current speed and a stop line of the vehicle. current acceleration; An obstacle sensing step is to drive an obstacle sensing device to sense an obstacle within a predetermined distance from the vehicle, so as to generate obstacle information corresponding to the obstacle and a plurality of numbers corresponding to the vehicle. The drivable space coordinate point, the obstacle information includes the current position of an obstacle, an obstacle speed and an obstacle acceleration corresponding to the obstacle; and In a lane sensing step, a lane sensing device is driven to sense a distance between a lane line of a vehicle and a road curvature. The hybrid decision-making method for self-driving cars according to claim 3, wherein the parameter obtaining step further comprises: A data processing step is configured and implemented by the arithmetic processing unit, and the data processing step includes: a cutting step, according to a preset time interval and a preset yaw rate of change, cutting the current position, the current heading angle, the current speed, the current acceleration, the obstacle information, the some drivable space coordinate points, the own vehicle lane line spacing and the road curvature to generate a cropping data; Wherein, there is a collision time interval between the vehicle and the obstacle, and the vehicle has a yaw rate; when the collision time interval is less than or equal to the preset time interval, the cutting step is started; when the yaw rate When the rate of change is less than or equal to the preset yaw rate of change, the cutting step is stopped. The hybrid decision-making method for self-driving cars according to claim 4, wherein the data processing step further comprises: A grouping step includes grouping the cropped data into a plurality of groups according to a plurality of predetermined acceleration ranges and a plurality of opposite obstacle information, the predetermined acceleration ranges include a conservative predetermined acceleration range and a normal predetermined acceleration range, the Some of the opposite obstacle information includes a pair of opposite obstacle-free information and a pair of opposite obstacle-free information, the groups include a conservative group and a normal group, and the conservative preset acceleration range is matched with the opposite obstacle-free information. The group should be conservative, and the normal preset acceleration range and the opposite obstacle information correspond to the normal group. The hybrid decision-making method for self-driving cars according to claim 4, wherein the data processing step further comprises: In a mirroring step, a vehicle trajectory function is mirrored along a vehicle traveling direction according to each scene category to generate a mirrored vehicle trajectory function, and the to-be-learned parameter group includes the mirrored vehicle trajectory function. The hybrid decision-making method for self-driving cars as claimed in claim 1, wherein the learning-based parameter optimization step comprises: A learning-based driving behavior generating step is to learn the to-be-learned parameter group according to the learning-based model to generate a learned behavior parameter group, and the to-be-learned parameter group includes a driving route parameter group and a driving acceleration/deceleration behavior parameter group; and A key parameter generating step is to calculate a system actuation parameter group of the learned behavior parameter group to obtain a system actuation time point, and calculate the system actuation time point, a vertical distance of a target point, a horizontal distance of a target point, A target point curvature, a vehicle speed and a target speed are combined to form a key parameter group. The hybrid decision-making method for self-driving cars according to claim 7, wherein, the learned behavior parameter set includes the system actuation parameter set, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, the vehicle speed and the target speed; and The system operation parameter group includes the vehicle speed, the acceleration of the vehicle, a steering wheel angle, a yaw rate, a relative distance and a lateral distance of an obstacle. The hybrid decision-making method for self-driving cars as claimed in claim 1, wherein the optimal trajectory function comprises: a plane coordinate curve equation representing an optimal trajectory of the vehicle in a plane coordinate; tangential speed, representing the speed of the vehicle at the tangent point of the equation of the plane coordinate curve; and Tangential acceleration, representing the acceleration of the vehicle at the tangent point; Wherein, the optimal trajectory function is updated according to a sampling time of the arithmetic processing unit. The hybrid decision-making method for self-driving cars according to claim 1, wherein the scene categories include: an obstacle occupancy scene, including an obstacle occupancy percentage, the obstacle occupancy scene represents an obstacle and a road in the surrounding scene, and the obstacle occupancy percentage represents a proportion of the obstacle occupied by the road; An intersection scene, representing an intersection in the surrounding scene; and An inbound and outbound scene means that there is an inbound and outbound station in the surrounding scene. A hybrid decision-making system for self-driving cars is used to decide an optimal trajectory function for a car. The hybrid decision-making system for self-driving cars includes: a sensing unit for sensing a surrounding scene of the vehicle to obtain a parameter set to be learned; a memory for accessing the to-be-learned parameter set, the plurality of scene categories, a learning-based model and a rule-based model; and an arithmetic processing unit electrically connected to the memory and the sensing unit, the arithmetic processing unit is configured to perform operations including the following steps: a learning-based scene decision-making step, which is to determine one of the scene categories that conform to the surrounding scenes of the own vehicle from the scene categories according to the parameter set to be learned and the learning-based model; a learning-based parameter optimization step of executing the learning-based model on the to-be-learned parameter set to generate a key parameter set; and A rule-based path planning step is to execute the rule-based model on the one of the scene categories and the key parameter group to plan the optimal trajectory function. The hybrid decision-making system for self-driving cars as claimed in claim 11, wherein the parameter group to be learned includes: The width of a vehicle road represents one of the widths of the road where the vehicle is located; A relative distance, representing a distance between the vehicle and an obstacle; an obstacle length, representing a length of the obstacle; and A lateral distance of an obstacle, representing a distance of the obstacle from the centerline of a lane. The hybrid decision-making system for self-driving cars as claimed in claim 11, wherein, the memory is used to access a map information related to a route traveled by the own vehicle; and The sensing unit includes: a vehicle dynamic sensing device for locating a current position of the own vehicle according to the map information, and sensing a current heading angle, a current speed and a current acceleration of the own vehicle; an obstacle sensing device for sensing an obstacle within a predetermined distance from the vehicle to generate obstacle information corresponding to the obstacle and a plurality of drivable space coordinate points corresponding to the vehicle, the obstacle The object information includes the current position of an obstacle, an obstacle velocity and an obstacle acceleration corresponding to the obstacle; and A lane sensing device senses the distance between a lane line of a vehicle and a road curvature. The hybrid decision-making system for self-driving cars as claimed in claim 13, wherein the arithmetic processing unit is configured to implement a data processing step, the data processing step comprising: a cutting step, according to a preset time interval and a preset yaw rate of change, cutting the current position, the current heading angle, the current speed, the current acceleration, the obstacle information, the some drivable space coordinate points, the own vehicle lane line spacing and the road curvature to generate a cropping data; Wherein, there is a collision time interval between the vehicle and the obstacle, and the vehicle has a yaw rate; when the collision time interval is less than or equal to the preset time interval, the cutting step is started; when the yaw rate When the rate of change is less than or equal to the preset yaw rate of change, the cutting step is stopped. The hybrid decision-making system for self-driving cars as claimed in claim 14, wherein the data processing step further comprises: A grouping step includes grouping the cropped data into a plurality of groups according to a plurality of predetermined acceleration ranges and a plurality of opposite obstacle information, the predetermined acceleration ranges include a conservative predetermined acceleration range and a normal predetermined acceleration range, the Some of the opposite obstacle information includes a pair of opposite obstacle-free information and a pair of opposite obstacle-free information, the groups include a conservative group and a normal group, and the conservative preset acceleration range is matched with the opposite obstacle-free information. The group should be conservative, and the normal preset acceleration range and the opposite obstacle information correspond to the normal group. The hybrid decision-making system for self-driving cars as claimed in claim 14, wherein the data processing step further comprises: In a mirroring step, a vehicle trajectory function is mirrored along a vehicle traveling direction according to each scene category to generate a mirrored vehicle trajectory function, and the to-be-learned parameter group includes the mirrored vehicle trajectory function. The hybrid decision-making system for self-driving cars as claimed in claim 11, wherein the learning-based parameter optimization step comprises: A learning-based driving behavior generating step is to learn the to-be-learned parameter group according to the learning-based model to generate a learned behavior parameter group, and the to-be-learned parameter group includes a driving route parameter group and a driving acceleration/deceleration behavior parameter group; and A key parameter generating step is to calculate a system actuation parameter group of the learned behavior parameter group to obtain a system actuation time point, and calculate the system actuation time point, a vertical distance of a target point, a horizontal distance of a target point, A target point curvature, a vehicle speed and a target speed are combined to form a key parameter group. The hybrid decision-making system for self-driving cars as described in claim 17, wherein, the learned behavior parameter set includes the system actuation parameter set, the longitudinal distance of the target point, the lateral distance of the target point, the curvature of the target point, the vehicle speed and the target speed; and The system operation parameter group includes the vehicle speed, the acceleration of the vehicle, a steering wheel angle, a yaw rate, a relative distance and a lateral distance of an obstacle. The hybrid decision-making system for self-driving cars as claimed in claim 11, wherein the optimal trajectory function comprises: a plane coordinate curve equation representing an optimal trajectory of the vehicle in a plane coordinate; tangential speed, representing the speed of the vehicle at the tangent point of the equation of the plane coordinate curve; and Tangential acceleration, representing the acceleration of the vehicle at the tangent point; Wherein, the optimal trajectory function is updated according to a sampling time of the arithmetic processing unit. The hybrid decision-making system for self-driving cars as claimed in claim 11, wherein the scene categories include: an obstacle occupancy scene, including an obstacle occupancy percentage, the obstacle occupancy scene represents an obstacle and a road in the surrounding scene, and the obstacle occupancy percentage represents a proportion of the obstacle occupied by the road; An intersection scene, representing an intersection in the surrounding scene; and An inbound and outbound scene means that there is an inbound and outbound station in the surrounding scene.

Images