A method for generating dangerous scenarios for autonomous driving based on vehicle-side and roadside data aggregation

Abstract

The present invention relates to the field of autonomous driving technology, and discloses a method for generating dangerous scenes for autonomous driving based on the aggregation of vehicle-side and road-side data. The method divides the perception area based on the type and regional characteristics of the takeover event; collects the vehicle-side data output by the current vehicle and the road-side perception data obtained from the exchange in the road-side fusion perception network intersection system in real time; determines the location of the takeover event based on the vehicle-side data; determines whether to retain the takeover event or perform data cleaning based on the location of the takeover event; aggregates the vehicle-side data set corresponding to the target takeover event, aggregates the corresponding road-side perception data set according to a preset aggregation method, and forms the dangerous scene data corresponding to the target takeover event with the vehicle-side data set and the road-side data set. The dangerous scenes generated by this solution have wider dimensions, richer information, and more realistic data, which is more conducive to the optimization of the autonomous driving function. The scene generation process is efficient, convenient, and low-cost.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside data. Background Art

[0002] While autonomous driving technology is developing rapidly, there is still great instability in practical applications. Therefore, a large amount of autonomous driving scenario data is needed to build different virtual scenarios and conduct comprehensive and rigorous simulation tests on intelligent connected vehicles. The richness, authenticity and quality value of autonomous driving scenario data play a vital role in the implementation of autonomous driving technology, the development of autonomous driving functions and test verification.

[0003] At present, there are two common ways to generate autonomous driving scene data: 1) Based on experience, subjectively arrange various road elements, including maps, pedestrians, vehicles, etc., on simulation software such as VTD to construct road scene data; 2) Use road scene truth collection vehicles. Problems in actual use include: the cost of subjective experience arrangement is low, but the scenes are limited, and the information of each road element set inevitably has subjective tendencies, resulting in distortion of road scene data and unreliable simulation results; although the generation of truth collection vehicles ensures the authenticity and diversity of the data, the cost of data acquisition is high. Whether it is the research and development of road truth collection vehicles or the manual cleaning and labeling of later data scenes, it involves large research and development investment and manpower costs.

[0004] In addition, when generating a full scene based on all ordinary data, it is inevitable that there will be data redundancy and data worthlessness. Therefore, the conventional solution is to generate and utilize dangerous scenes specifically in scene generation. However, this method of generating dangerous scenes usually adds traditional environmental risk factors within the scope of cognition to the data screening, and constructs corresponding dangerous scenes with the screened data. From the perspective of the development of autonomous driving technology and the variability and complexity of the driving environment of autonomous driving vehicles, such dangerous scenes and their data value generated based on limited cognition are not suitable for the dynamic development of autonomous driving technology, cannot effectively identify unknown dangers, and have limitations in applicability. Summary of the invention

[0005] The present invention aims to provide a method for generating dangerous scenarios for autonomous driving based on the aggregation of vehicle-side and roadside data, so as to solve the technical problems of low efficiency and high cost of existing autonomous driving scene generation methods.

[0006] The basic solution provided by the present invention is: a method for generating dangerous scenarios for autonomous driving based on the aggregation of vehicle-side and roadside data, the method comprising:

[0007] S100, determining a current vehicle and a current area where the current vehicle needs to travel;

[0008] S200, based on the type of the takeover event and the regional characteristics, using the electronic fence to divide the current area into a first area and a second area in a preset division method, to form a first area set and a second area set respectively;

[0009] S300, real-time collection of vehicle-side data output by the current vehicle and roadside data collected from the networked intersections in the roadside fusion perception system;

[0010] S400, screening the takeover events that occurred within the electronic fence, obtaining the location of the takeover event according to the vehicle-side data, and determining whether the location of the takeover event meets a preset condition. If the preset condition is met, retaining the takeover event and using it as a target takeover event, otherwise performing data cleaning;

[0011] S500, aggregating the vehicle-side data set corresponding to the target takeover event, and aggregating the corresponding roadside data set in a preset aggregation method, so that the vehicle-side data set and the roadside data set constitute the dangerous scene corresponding to the target takeover event.

[0012] Further, in S200, the preset division method is:

[0013] At each networked intersection, a first area is delineated with the intersection RSU as the center and R1 as the radius to form a first area set; if there are overlapping areas between adjacent first areas, the radius R1 is reduced until there is just no overlap;

[0014] A second area is delineated on the vehicle driving path in the current area according to the constraint conditions to form a second area set.

[0015] Furthermore, the constraint conditions include the starting and ending positions of the route travel, and the platform parking position.

[0016] Further, in S300, the vehicle-side data includes the vehicle GPS position, the vehicle-side video and the vehicle CAN signal; the vehicle-side video includes the cockpit monitoring video, the vehicle left front monitoring video and the vehicle right front monitoring video.

[0017] Further, in S300, the roadside data includes roadside perception data, signal control data and map information constituted by messages broadcast by the RSU at the networked intersection.

[0018] Further, in S400, the preset condition is that the location where the takeover event occurs is within the first area set and not within the second area set.

[0019] Further, in S500, the vehicle-side data set corresponding to the target takeover event includes a number of vehicle-side data in the time period before and after t1, centered at the time when the target takeover event occurs.

[0020] Further, in S500, the preset aggregation method is to locate the area according to the location of the target takeover event, locate the corresponding networked intersection in the area, and obtain from the networked intersection a number of roadside data in the time period t2 before and after the time of the target takeover event as the center to form a roadside data set.

[0021] Furthermore, based on the principle of minimum area overlap, the information of the current vehicle as a traffic participant is deleted from a number of roadside data.

[0022] Further, in S100, a monitoring terminal and an OBU are installed on the vehicle side, wherein the monitoring terminal is connected to a video acquisition device, a positioning device and a CAN signal; the OBU is connected to a host computer, and the OBU is used to receive messages broadcast by the RSU at the networked intersection in the roadside fusion perception system and output them to the host computer.

[0023] The working principle and advantages of the present invention are: the takeover signal of the autonomous driving vehicle drives the generation of dangerous scene data, based on the type of takeover event and the regional characteristics, the current area where the current vehicle is traveling is divided into perception areas, and then the vehicle-side data and roadside data are collected, and the location of the takeover event is determined according to the vehicle-side data; whether the takeover event is retained or cleaned is determined based on whether the location of the takeover event meets the preset conditions; the vehicle-side data set corresponding to the target takeover event is aggregated, and the corresponding roadside data set is aggregated according to the preset aggregation method, and the vehicle-side data set and the roadside data set form the dangerous scene corresponding to the target takeover event.

[0024] Compared with the existing technology, this solution breaks through the technical bias of the existing technology in recognizing dangerous scenes, and opens up a new angle and idea to recognize the value of scene data. It changes the definition of dangerous scenes from the traditional result event type to starting from the vehicle side, and judging the scene danger and data value based on the takeover situation directly fed back by the vehicle side. After collecting all the data that can be collected before and after the relative time with the vehicle side takeover event as the collection point, the dangerous scenes that may affect the vehicle are restored and extracted. The real restoration is high and more targeted, which is more in line with the characteristics of the development of autonomous driving technology itself. In order to realize the distinction of dangerous scene data based on the takeover directly fed back by the vehicle side, the regional characteristics of the takeover event are deeply mined, and the data source is regionalized. Driven by the takeover and regional characteristics, the dangerous factors are identified based on the feedback of vehicle behavior, and the identification of dangerous factors is handed over to the vehicle side to generalize the danger of the scene, avoiding the limitations of subjective cognition, and at the same time, it can expand the effective identification of dangerous factors, especially unknown dangerous factors. The dangerous scenes generated by this solution are more conducive to the optimization of autonomous driving functions.

[0025] This solution breaks through the limitations of the existing independent use of vehicle-side data or roadside data to generate scene data. On the basis of the regional scenario of the data source, it realizes the method of aggregating vehicle-side data and roadside data in the real road environment to generate dangerous scenes, aggregates data within a certain period of time to form a dynamic dangerous scene, and makes the scene dynamic, so that the generated dangerous scene has a wider dimension, richer information, and more real data. In specific applications, the segmentation of dangerous scenes and subsequent technical optimization analysis have higher data value; based on this method, the equipment can be built using existing vehicle-side and road equipment, data acquisition is convenient, the investment cost is low, and subjective cooperation work is reduced. In addition, this solution also breaks through the conventional back-end full collection and targeted scene annotation method adopted to ensure the integrity of the data, and is optimized to be directly screened and cleaned at the front end of data collection, which not only ensures the value of the data, but also reduces the amount of data, and is also conducive to the operation of the collection system, and constructs a dangerous scene data set with real data, high quality, and appropriate data volume. The entire dangerous scene generation process is simple, efficient, convenient, and low-cost, which is beneficial in terms of time, economy, and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A flow chart of a method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside data provided by an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of the first area demarcation of the method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside data provided in an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of the second area demarcation of the method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside data provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following is a further detailed description through specific implementation methods:

[0030] The embodiment is basically as shown in the attached Figure 1 As shown: A method for generating dangerous scenarios for autonomous driving based on the aggregation of vehicle-side and road-side perception data, the method comprises the following steps:

[0031] S100, determining a current vehicle and a current area where the current vehicle needs to travel.

[0032] Specifically, the current vehicle is a vehicle with an autonomous driving function. In order to complete the aggregation of scene data sources, a platform is built. The platform includes the current vehicle and a host computer. At the same time, data is exchanged with the roadside fusion perception system to collect perception data of the roadside fusion perception system. Finally, the scene data source can be obtained from the current vehicle and the host computer. Compared with conventional technologies, this method automatically aggregates data, the data comes from public roads, and the autonomous driving vehicles and data collection host computer software put into operation are reused. Long-term collection can be achieved without the need to arrange full-time personnel for processing, saving manpower and material costs.

[0033] The monitoring terminal is connected to the video acquisition device, the positioning device and the CAN signal.

[0034] The video acquisition device includes a cockpit monitoring camera, and the left and right front cameras of the vehicle. The cockpit monitoring camera is used to monitor the driving behavior of the safety officer, and a product with a pixel of 720P and a focal length of 2.5 mm can be preferably selected. The left and right front cameras of the vehicle are used to monitor the situation in front of the vehicle, and a product with a pixel of 720P and a focal length of 4.0 mm can be preferably selected. The camera installation position and angle are adjusted to make the monitoring field of view meet the requirements. The field of view can be selected to cover 100 meters in front of the vehicle, and the monitoring video is saved in real time.

[0035] The positioning device receives the positioning signal. The GPS antenna can be installed in the center of the vehicle to ensure good signal access. The monitoring terminal can also be connected to the RTK positioning module at the same time to achieve high-precision positioning. The positioning accuracy reaches the centimeter level, thereby improving the accuracy of the positioning data.

[0036] The host computer sets the mapping relationship of CAN signals, including vehicle control status, gear position, steering wheel, throttle travel, manual takeover reminder and other signals, and sends the configuration information and DBC file to the monitoring terminal. The monitoring terminal accesses the current vehicle's automatic driving control CAN signal through the OBD port or DB9 port of the automatic driving vehicle. The parsed CAN signals include vehicle control status, gear position, steering wheel, throttle travel, manual takeover reminder and other control signals. CAN signals play an important role in the judgment of takeover events and the division of regions in this solution, which will be described in detail in the subsequent steps.

[0037] The monitoring terminal collects and saves video, positioning and CAN signals to the local hard disk in real time. The vehicle-side data includes the vehicle GPS position, vehicle-side video, and vehicle CAN signal. The vehicle-side video includes the cockpit monitoring video, the vehicle left front monitoring video, and the vehicle right front monitoring video. The sampling frequency of CAN signal and satellite positioning is 1HZ. Based on such a platform, it is possible to obtain appropriate and clear data.

[0038] Due to the functional limitations of the autonomous driving monitoring terminal equipment, it is impossible to directly output the structured data of traffic participants on the public road, such as the GPS position, speed, heading angle, geometric size, type, etc. of vehicles, the GPS position, speed, walking direction, etc. of pedestrians; while the intersection fusion perception system can perceive the GPS position, speed and other kinematic information of traffic participants such as vehicles and pedestrians. Therefore, the aggregation of roadside messages can make up for the lack of data collected by the autonomous driving monitoring terminal, and form structured road scene data, so that the generated scene dimension is wider, the information is richer, and the data is more real.

[0039] The current area is an intelligent connected road area, which is equipped with a roadside fusion perception system and several connected intersections. The roadside fusion perception system perceives the target object data of road traffic participants at the connected intersections and pushes the perception results to the RSU at the connected intersection; the RSU broadcasts in the PC5 network segment in the form of RSM messages, MAP messages and SPAT messages.

[0040] The current vehicle is equipped with an OBU, which is connected to the host computer. When the autonomous vehicle passes through each networked intersection, the vehicle OBU receives the messages broadcast by the RSU of each networked intersection of the roadside fusion perception system and transmits them to the host computer. The host computer collects and saves the RSM messages, MAP messages and SPAT messages broadcast by the RSU in real time through the OBU. Among them, the RSM message constitutes the roadside perception data (i.e., the roadside traffic participant data), the MAP message constitutes the map information (i.e., the roadside static information), and the SPAT message constitutes the signal control data (i.e., the signal light status data). The above data together constitute the roadside data and are saved on the local hard disk. The aggregated data includes the signal control data and the map information, so that the data set of each dangerous scene is more comprehensive, which is conducive to improving the value of the scene data.

[0041] Through the method of obtaining the above vehicle-side data and roadside data in this solution, the data that needs to be aggregated can be obtained conveniently, in an appropriate amount, and with accuracy.

[0042] S200, based on the type of the takeover event and the regional characteristics, using the electronic fence to divide the current area into a first area and a second area in a preset division method, to form a first area set and a second area set respectively.

[0043] Specifically, not all vehicle-side data and roadside data are valuable. In order to ensure the integrity of the data, the conventional method is usually to collect all data on the back end and distinguish the value through subjective labeling. This method will result in a large amount of back-end data, strong randomness in value judgments, and cannot truly reflect the value of the data. The conventional cognition of dangerous scenarios is also based on subjectively identified dangerous elements, which will lead to limitations in the cognition of danger. The dangerous scenarios generated by the data selected in this way are also distorted, and unknown dangers cannot be effectively identified.

[0044] After analysis, the CAN signal received by the vehicle-side monitoring terminal shows that the vehicle control status is divided into automatic driving mode and manual takeover mode. In automatic driving mode, the vehicle is controlled by the automatic driving system; in manual takeover mode, the vehicle is controlled by the safety officer. The vehicle switches from automatic driving mode to manual driving mode, which is called a takeover event. Takeover events can be divided into two situations: 1) Planned takeover, that is, the vehicle takes over according to the driving plan, such as the end of the route, stopping at the station, etc. 2) Unplanned takeover, when more complex road traffic conditions occur during driving, which exceed the processing capacity of the automatic driving system, or continue to be in automatic driving state, the risk is greater and traffic accidents may occur.

[0045] In fact, when an unplanned takeover occurs, the road traffic conditions are more complicated. Road scenes in complex environments are often of high value for optimizing autonomous driving algorithms. Therefore, judging the value of data based on planned and unplanned takeovers is in line with the characteristics of the development of autonomous driving technology itself, and on the other hand, it can ensure the construction of a real and valuable data set. It can also generalize the danger, that is, not judging the value of data based on the existing conventional subjective risk factors, but identifying risk factors based on vehicle behavior feedback. The identification of risk factors is left to the vehicle side to avoid the limitations of subjective cognition. At the same time, it can expand the effective identification of risk factors, especially unknown risk factors.

[0046] In view of the difference in importance of the two takeover methods, it is necessary to distinguish the two takeover events. However, the CAN signal itself does not have the identification of the takeover attribute, that is, it is impossible to distinguish whether it is an unplanned takeover or a planned takeover. Therefore, this solution uses external conditions to distinguish the takeover types. Since planned takeover events such as vehicle entry have obvious regional characteristic attributes, based on the regional characteristics of the takeover event, the regional division method can, on the one hand, achieve the distinction of data value, and at the same time, it can also collect the corresponding roadside data of the vehicle-side data based on the divided regional positioning, and then aggregate the vehicle-side data and its corresponding roadside data, so as to form a rich-dimensional and data-rich scene data source.

[0047] Specifically, Figure 2 As shown, the first area defined by the first type of electronic fence is the effective coverage area of ​​the RSU message. The upper computer sets an electronic fence at the networked intersection to filter data, and defines a first area for each networked intersection to form a first area set U 0The setting of the electronic fence in the first area refers to the standard requirements for the coverage of the RSU message at the networked intersection. The electronic fence in the first area covers the range with the RSU at the networked intersection as the center and R as the radius. If there is an overlapping area between adjacent first areas, the radius R1 is reduced until there is no overlap. R can be controlled within 280-340 meters, and 300 meters can be preferred. The selection of this data range can well match the existing autonomous driving road area conditions, while reducing the overlapping parts of the area, reducing the adjustment frequency, and improving the construction efficiency. The first area is the coverage range of fusion perception. When an unplanned takeover event occurs within this range, the scene data aggregation is driven.

[0048] On the driving path of the autonomous driving vehicle in the current area, according to the regional characteristics of the planned takeover, the electronic fence perception range of the area where the planned takeover event occurs is set, and all takeover events occurring within the electronic fence are regarded as planned takeovers. Planned takeover events are not used to drive the aggregation of scene data.

[0049] Specifically, Figure 3 As shown, the second type of electronic fence demarcates the second area, which is regarded as the planned takeover area and is demarcated according to the constraint conditions to form the second area set U 1 , it is approximately considered that it occurs in the second region set U 1 Any element in the plan is a planned takeover; the constraints include the starting and ending positions of the driving route, and the platform parking position, which can meet the identification of conventional planned takeover events and realize the reasonable division of takeover events.

[0050] S300 collects in real time the vehicle-side data output by the current vehicle and the roadside data collected from the networked intersections in the roadside fusion perception system.

[0051] Specifically, the current vehicle is traveling in the current area, and the data collected by the current vehicle and the host computer are stored in real time. The data can be acquired in real time or offline for processing, including but not limited to automatic real-time acquisition and periodic acquisition for centralized generation of dangerous scenarios. In this solution, since the generation method of each dangerous scenario is centered on the occurrence of a takeover event and has the independence and integrity of the event, scene generation can be carried out regardless of real-time processing or offline processing, and the continuity of the generation of dangerous scenarios is not affected. This solution is more flexible in terms of scene generation and data acquisition based on takeover events.

[0052] S400, filter the takeover events that occur within the electronic fence, obtain the location of the takeover event based on the vehicle-side data, and determine whether the location of the takeover event meets the preset conditions. If the preset conditions are met, retain the takeover event and use it as the target takeover event, otherwise perform data cleaning.

[0053] Specifically, the location of the takeover event is determined based on the vehicle-side data; the vehicle-side data includes the vehicle's GPS location, vehicle-side video, and vehicle CAN signal. The occurrence of the takeover event is determined by changes in the vehicle's CAN signal, and the location of the takeover event is determined by the vehicle's GPS location at the moment the vehicle's CAN signal changes.

[0054] The preset condition is that the location where the takeover event occurs is within the first region set and not within the second region set. If the preset condition is met, the takeover event is retained and used as the target takeover event, otherwise data cleaning is performed.

[0055] The division of the first area and the second area is based on the regional characteristics of the takeover event. Therefore, planned and unplanned takeover events can be distinguished according to whether the event occurs within the first type of electronic fence or the second type of electronic fence. According to the reason for the aforementioned division of areas, takeover events occurring in the second area are considered planned takeover events and are not used to drive scene data aggregation. This part of the data is cleaned; takeover events occurring in the first area but not in the second area are considered unplanned takeover events, and can receive reliable RSU broadcast messages for driving scene data aggregation. This part of the data is retained; takeover events may also occur in non-intersection areas or non-networked intersections, that is, not within the scope of any designated networked intersection electronic fence. Due to the lack of roadside data, such takeover events are not used to generate road scene data and are directly cleaned.

[0056] Simply put, based on the location (GPS) where the takeover event occurred, traverse all the electronic fences in the RSU coverage area of ​​the fusion perception network intersection to meet the necessary condition: the location where the takeover event occurred is within a first area. Based on the location (GPS) where the takeover event occurred, traverse all the second areas to meet the necessary condition: the location where the takeover event occurred cannot exist in any second area. Both of the above necessary conditions are met to drive the aggregation of data on dangerous scenarios for autonomous driving.

[0057] Thus, a set of unplanned takeover events E that occurs within the electronic fence of the networked sensing intersection, that is, within the first area set but not within the second area set, can be formed. 0 ; Through regional division, takeover events that were previously unclassifiable are classified, and data screening and cleaning are achieved based on the scope. The whole process is simple and effective, achieving efficient data processing. The system computing power requirements are not high, which is conducive to system construction.

[0058] S500, aggregating the vehicle-side data set corresponding to the target takeover event, and aggregating the corresponding roadside data set in a preset aggregation method, so that the vehicle-side data set and the roadside data set constitute the dangerous scene corresponding to the target takeover event.

[0059] Specifically, the vehicle-side data set corresponding to the target takeover event includes several vehicle-side data in the time period t1 before and after the time when the target takeover event occurs as the center, that is, the time when the CAN signal state of the driving mode changes is taken as the center, and the GPS position data, video data and CAN signal data before and after each t1 time can be aggregated to constitute the vehicle-side data of the target takeover event. t1 can be 60s, that is, 60s before and after the occurrence time, a total of 120s of vehicle-side data. Reasonable selection of this time period can obtain an appropriate amount of effective vehicle-side data and avoid data redundancy.

[0060] The preset aggregation method is to locate the area where the target takeover event occurs according to the location, locate the corresponding networked intersection in the area, and obtain a number of roadside data in the time period t2 before and after the target takeover event as the center from the networked intersection to form a roadside data set.

[0061] Traverse the unplanned takeover event set E 0 For each takeover event element, the GPS location where the takeover event occurs is matched to the electronic fence, and the RSM message data set of the networked intersection corresponding to the electronic fence is retrieved, and the MAP message and SPAT message of the intersection are retrieved at the same time. t2 can be 60 seconds, that is, 60 seconds before and after the takeover event, a total of 120 seconds of RSM message data are retrieved, and the roadside data set of the same time period of the vehicle-side data is aggregated to form a valid data set belonging only to the dangerous scene. The data set is used to form a dynamic dangerous scene, which is conducive to the later application of dangerous scenes, so that there can be a wider range of related data for comprehensive analysis and comprehensive utilization, and a more complete evaluation result and utilization value.

[0062] The RSM message includes information about one or more traffic participants, which may include the autonomous driving vehicle itself. It is necessary to delete the information about the current vehicle as a traffic participant from the RSM messages in several roadside data based on the principle of minimum area overlap. Specifically, based on the timestamp of the RSM message, retrieve the data frame with the smallest time difference from the takeover event data set, obtain the GPS position of the autonomous driving vehicle at that moment, and calculate the distance between the position of the autonomous driving vehicle and each traffic participant in the RSM message. Select the target with the smallest distance and an absolute value of less than 0.5 meters, and regard it as the autonomous driving vehicle with the monitoring terminal installed, and delete the corresponding target information from the RSM message. The processed RSM message set constitutes the road traffic participant data at the time of the takeover event.

[0063] The processed RSM message data constitutes roadside perception data (i.e., roadside traffic participant data), MAP messages constitute map information, and SPAT messages constitute signal control data. The above data together constitute the roadside data of the target takeover event, and are aggregated with the vehicle-side data set corresponding to the target takeover event to constitute a road scene simulation data set for the target takeover event, generating a dangerous scene corresponding to the target takeover event.

[0064] The method for generating dangerous scenarios for autonomous driving based on the aggregation of vehicle-side and roadside data provided in this embodiment has a different understanding of the value of dangerous scenario data from that of existing dangerous scenarios during the scenario generation process. This solution starts from the vehicle-side and judges the danger of the scenario and the value of the data based on the takeover situation directly fed back by the vehicle-side, which is in line with the characteristics of the development of autonomous driving technology itself. At the same time, in order to achieve the distinction of dangerous scenario data based on the takeover situation directly fed back by the vehicle-side, the regional characteristics of the takeover event are deeply mined, the data source is scenario-based, and then the danger of the scenario is generalized, which is not limited to the conventional danger factors in the existing active cognition, but is expanded to the unknown danger factors within the regional range. The dangerous scenarios generated by this solution are more conducive to the optimization of the autonomous driving function.

[0065] This solution breaks through the limitations of the existing independent use of vehicle-side data or roadside data to generate scene data. On the basis of regional scenario of data source, it realizes the method of aggregating vehicle-side data and roadside data in real road environment to generate scenes, so that the generated scenes have wider dimensions, richer information and more realistic data. Based on this method, the equipment can be set up using existing vehicle-side and road equipment, which makes data acquisition convenient, low-cost and reduces subjective cooperation work.

[0066] In addition, this solution also breaks through the conventional method of back-end comprehensive collection and targeted scene annotation to ensure data integrity, and optimizes it to direct screening and cleaning at the front end of data collection, which not only ensures the value of the data, but also reduces the amount of data, and is beneficial to the operation of the collection system. It constructs a scene data set with real data, high quality and appropriate data volume. The entire scene generation process is simple, efficient, convenient and low-cost, and is beneficial in terms of time, economy and quality.

[0067] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field are aware of all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement the scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for the ordinary technicians in the relevant field to implement this application. It should be pointed out that for the technicians in this field, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for generating dangerous scenarios for autonomous driving based on the aggregation of vehicle-side and roadside perception data, characterized in that: The method comprises the following steps: S100, determining a current vehicle and a current area where the current vehicle needs to travel; S200, based on the type of the takeover event and the regional characteristics, using the electronic fence to divide the current area into a first area and a second area in a preset division method, to form a first area set and a second area set respectively; S300, real-time collection of vehicle-side data output by the current vehicle and roadside data collected from the networked intersections in the roadside fusion perception system; S400, screening the takeover events that occurred within the electronic fence, obtaining the location of the takeover event based on the vehicle-side data, and determining whether the location of the takeover event meets the preset conditions. If the preset conditions are met, retain the takeover event and use it as the target takeover event, otherwise perform data cleaning; S500, aggregating the vehicle-side data set corresponding to the target takeover event, and aggregating the corresponding roadside data set according to the preset aggregation method, so that the vehicle-side data set and the roadside data set constitute the dangerous scene corresponding to the target takeover event; In S200, the preset division method is: At each networked intersection, a first area is delineated with the intersection RSU as the center and R1 as the radius to form a first area set; if there are overlapping areas between adjacent first areas, the radius R1 is reduced until there is just no overlap; Delimiting a second area on the vehicle driving path in the current area according to the constraint conditions to form a second area set; In S400, the preset condition is that the location where the takeover event occurs is within the first area set and not within the second area set; Among them, takeover events occurring in the second area are all regarded as planned takeover events and are not used to drive dangerous scene data aggregation; takeover events occurring in the first area but not in the second area are all regarded as unplanned takeover events, and reliable RSU broadcast messages can be received to drive dangerous scene data aggregation.

2. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: The constraints include the starting and ending positions of the driving route, and the platform parking position.

3. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: In S300, the vehicle-side data includes the vehicle GPS position, vehicle-side video and vehicle CAN signal; the vehicle-side video includes the cockpit monitoring video, the vehicle left front monitoring video and the vehicle right front monitoring video.

4. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: In S300, the roadside data includes roadside perception data, signal control data and map information constituted by messages broadcast by RSUs at networked intersections.

5. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: In S500, the vehicle-side data set corresponding to the target takeover event includes a number of vehicle-side data in the time period t1 before and after the time when the target takeover event occurs as the center.

6. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: In S500, the preset aggregation method is to locate the area where the target takeover event occurs according to the location, locate the corresponding networked intersection in the area, and obtain from the networked intersection a number of roadside data in the time period t2 before and after the time of the target takeover event as the center to form a roadside data set.

7. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 6, characterized in that: Based on the principle of minimum area overlap, the information of the current vehicle as a traffic participant is deleted from several roadside data.

8. The method for generating dangerous scenarios for autonomous driving based on aggregation of vehicle-side and roadside perception data according to claim 1, characterized in that: In S100, the current vehicle is equipped with a monitoring terminal and an OBU, wherein the monitoring terminal is connected to a video acquisition device, a positioning device and a CAN signal; the OBU is connected to a host computer, and the OBU is used to receive messages broadcast by the RSU at the networked intersection in the roadside fusion perception system and output them to the host computer.

Images