KR102477885B1 - Safety analysis management server for evaluating autonomous driving roads - Google Patents

Abstract

The present invention relates to a safety grade analysis management server for evaluating a safety grade of driving on an autonomous driving road, capable of reducing an accident related to an autonomous vehicle. According to an embodiment of the present invention, a driving safety grade analysis server receives driving data provided by the autonomous vehicle and infrastructure data transmitted by an Internet of things (IoT) sensor. The driving safety grade analysis server includes: a communication unit receiving the driving data and the infrastructure data; a storage unit storing the received driving data and the infrastructure data; and a control unit drawing a road risk coefficient by road section from the driving data and the infrastructure data, which are accumulated, through a deep embedded clustering algorithm and evaluating a safety grade of driving by road section based on the road risk coefficient. The control unit controls the communication unit and the storage unit to accumulate the driving data and the infrastructure data in the storage unit for a predetermined period. Also, the control unit divides the driving data and the infrastructure data, which are accumulated, by road section and forms the divided driving data as a data set for deep embedded clustering. Moreover, the control unit learns a risk cluster classification model by using the data set and calculates the road risk coefficient and the safety grade of driving by road section. Furthermore, the control unit analyzes a driving feature of the autonomous vehicle by road section by linking the safety grade of driving with the infrastructure information.

Description

Safety level analysis management server that evaluates the driving safety level of autonomous driving roads

The present invention relates to an autonomous vehicle, and more particularly, to a driving safety analysis server that calculates a road driving risk index and a road driving safety level using driving data of the autonomous vehicle, and an operation method thereof.

Conventional road grading or safety level studies of general vehicles have been conducted in a way of calculating a safety index based on the number of accidents through various road-related variables or through sensor monitoring of an autonomous vehicle. However, as the proportion of self-driving cars gradually increases on real roads, in the case of the United States and Japan, road grading is being carried out with or without infrastructure for smooth operation of self-driving cars.
In Korea, studies have been conducted to evaluate the road environment on continuous streams (highways), but there is no road environment evaluation on intermittent streams (urban roads), where safety is most urgent in a situation where autonomous vehicles and general vehicles are mixed. Therefore, there is a need for technology to analyze driving risk factors according to road sections on intermittent roads operated by autonomous vehicles and to ensure safety. To this end, the following R&D projects were carried out.
[Supporting Organization] Next Generation Convergence Technology Research Institute
[Name of Host Organization] Next Generation Convergence Technology Research Institute
[Assignment identification number] AICT-2021-0011
[R&D Project Name] 2021 Yung Ki-won-Small and Medium Business Joint Planning Research
[R&D task name] Development of autonomous vehicle abnormal behavior discrimination technology using deep learning
[Contribution rate] 20%
[Research Period] 2021.04.05 ~ 2021.11.30

(1) Korean Patent Publication No. 10-2018-0078973 (2018.07.10) (2) Korea Registered Patent Publication No. 10-1901801 (2018.09.18)

An object of the present invention is to collect driving data in urban centers in order to collect intermittent driving data reflecting real road environments, and to calculate road risk index and evaluate road driving safety using deep learning-based algorithms. It is to provide a safety level analysis server and its analysis method.

The grading analysis server for receiving the driving data provided from the autonomous vehicle according to an embodiment of the present invention and the infrastructure data transmitted through the Internet of Things (IoT) sensor, the autonomous vehicle according to the embodiment of the present invention The driving safety analysis server that receives the driving data provided from the Internet of Things (IoT) sensor and the infrastructure data transmitted through the sensor, the driving safety analysis server, the communication unit receiving the driving data and the infrastructure data, the received driving data and the infrastructure data A storage unit that stores the accumulated driving data and the infrastructure data derives a road risk index for each road section through a deep embedded clustering algorithm, and evaluates the driving safety for each road section based on the road risk index. And a control unit that controls the communication unit and the storage unit to accumulate the operation data and the infrastructure data in the storage unit for a specific period of time, and stores the accumulated operation data and infrastructure data for each section of the road. classify, compose the classified driving data into a dataset for deep embedded clustering, learn a risk cluster classification model using the dataset, and learn the road section by road section from the learning result. A risk index and the driving safety level are calculated, and driving characteristics of the autonomous vehicle for each road section are analyzed by linking the driving safety level and the infrastructure information.
In this embodiment, the driving data includes at least one of information about speed, lateral acceleration, longitudinal acceleration, and rotation rate of the self-driving vehicle.
In this embodiment, the control unit constructs the dataset by processing the speed, the lateral acceleration, the longitudinal acceleration, and the rotation rate with a piecewise aggregate approximation (PAA) algorithm.
In this embodiment, the controller substitutes the longitudinal acceleration, the lateral acceleration, and the 'Z-score' of the rotation rate into the cumulative distribution function (CDF), and the calculated cumulative probability clusters corresponding to the upper and lower 10% risk It is judged by the driving behavior cluster.
In this embodiment, the control unit determines the road risk index and driving safety level for each road section based on the risky driving behavior cluster.

According to the above-described driving safety analysis server of autonomous driving roads and its analysis method according to an embodiment of the present invention, accidents related to autonomous vehicles can be reduced, and through analysis of road driving safety levels, industries and institutions related to road administration It can be used as an auxiliary material for policy establishment. In addition, safety can be secured in all fields related to traffic safety, as it can serve as an analysis tool to determine responsibility in the event of an autonomous vehicle-related accident.

1 is a block diagram exemplarily showing a system for acquiring and analyzing empirical big data of an autonomous vehicle according to an embodiment of the present invention.
2 is a block diagram showing the configuration of a vehicle sensor module loaded or mounted on an autonomous vehicle as an example.
3 is a block diagram showing the configuration of the Internet of Things (IoT) sensor of FIG. 1 as an example.
4 is a block diagram showing the driving safety level analysis server shown in FIG. 1 as an example.
5 is a flowchart briefly illustrating a method of performing road grading performed by the driving safety level analysis server of the present invention.
FIG. 6 is a flowchart illustrating the road driving safety evaluation procedure of FIG. 5 in more detail.
7 is a diagram showing a road map in which driving data of an autonomous vehicle is collected according to an embodiment of the present invention.
8 is a view showing a part of a precision road map according to an embodiment of the present invention.
FIG. 9 is a graph briefly showing an example of the partial aggregation approximation (PAA) algorithm of the present invention mentioned in FIG. 7 .
10 is a diagram briefly showing a method of constructing a deep embedded clustering (DEC) algorithm of the present invention.
11 is a diagram exemplarily illustrating a driving safety evaluation result of an autonomous driving road and an analysis method thereof according to an embodiment of the present invention.
12 is a graph exemplarily showing the DEC evaluation index (Generalizability) for each number of clusters according to an embodiment of the present invention.
13 is a graph exemplarily showing results of deep embedded clustering (DEC) according to an embodiment of the present invention.
14 is a table briefly showing how to calculate the risk index and the inclusion ratio of risky driving behavior clusters for each road section.
15 is a diagram visually showing a driving safety level for each road section of a final autonomous driving road based on a road risk index (RRI).
16 and 17 briefly show an example of analysis in connection with road driving safety and road infrastructure according to the present invention.

Hereinafter, some embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to components of each drawing, the same components may have the same numerals as much as possible even if they are displayed on different drawings. In addition, in describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description may be omitted.
1 is a block diagram exemplarily showing a system for acquiring and analyzing empirical big data of an autonomous vehicle according to an embodiment of the present invention. Referring to FIG. 1, a system 10 for acquiring empirical big data (autonomous vehicle operation data and infrastructure data) of an autonomous vehicle 105 includes a vehicle sensor module 100, an Internet of Things (IoT) sensor 200 ), a communication network 300, and a driving safety analysis server 400.
The vehicle sensor module 100 is installed in the self-driving vehicle 105 . The self-driving vehicle 105 may be a vehicle equipped with an autonomous driving system in various types of vehicles capable of driving on real roads for acquiring empirical big data. The vehicle sensor module 100 mounted on the autonomous vehicle 105 generates driving data of the autonomous vehicle 105 using various sensors. For example, the vehicle sensor module 100 may include a speed sensor, an acceleration sensor, a yaw rate sensor, and a position sensor of the autonomous vehicle 105 . In addition, the driving data sensed by these sensors may be transmitted to the driving safety level analysis server 400 through the communication network 300 .
The IoT sensor 200 may be a sensor or camera installed on a road or above the road to monitor road conditions, traffic flow, collision accidents, parking conditions, and the like. The IoT sensor 200 may include a sensor for detecting humidity and air quality for each road section, a road surface sensor for monitoring a road surface condition, and the like. The IoT sensor 200 may sense the state of humidity or ice and snow according to the amount of snow or precipitation on the road surface through the road surface sensor. The IoT sensor 200 may transmit data sensed for each section and lane of the road to the driving safety analysis server 400 through the communication network 300 as infrastructure data.
The communication network 300 provides a communication channel between the vehicle sensor module 100, the IoT sensor 200, and the driving safety level analysis server 400. The communication network 300 refers to a wireless or wired communication structure for exchanging information between nodes such as the vehicle sensor module 100 or the driving safety analysis server 400 . For example, the communication network 300 may include Vehicle to Everything (V2X) for exchanging information between vehicles and things such as other vehicles, mobile devices, and roads. Alternatively, the communication network 300 may include 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), World Interoperability for Microwave Access (WIMAX), Wi-Fi, 3G, 4G, 5G, 6G, etc. However, the present invention is not limited thereto.
The driving safety analysis server 400 generates a road risk index for each road section by using driving data and infrastructure data provided from the vehicle sensor module 100 and the IoT sensor 200 using a deep learning-based clustering algorithm. The driving safety level analysis server 400 may evaluate the driving safety level for each section of the road on which the autonomous vehicle 105 operates based on the road risk index of the autonomous vehicle 105 .
The driving safety level analysis server 400 may interpret the derived road grading result in association with road infrastructure such as a geometric structure or facilities of a road section. The driving safety level analysis server 400 may monitor the accident risk level for each section of the road on which the autonomous vehicle 105 is operating in association with road infrastructure, and may use the data to increase the safety level of vehicle operation. Alternatively, the road grading result can be used to supplement the driving algorithm of the self-driving vehicle 105 or to be used as insurance data or road safety management data.
According to the driving safety level analysis server 400 described above, a safety index and road safety level for each road section of the self-driving vehicle 105 may be derived based on empirical big data collected from actual roads. In addition, the derived road safety may be used to develop an operating algorithm of the self-driving vehicle 105 in connection with actual road infrastructure. In addition, road safety can be used as a data for infrastructure improvement by linking the risk section of autonomous driving roads with road infrastructure.
2 is a block diagram showing the configuration of a vehicle sensor module loaded or mounted on an autonomous vehicle as an example. Referring to FIG. 2 , the vehicle sensor module 100 may generate driving data of an autonomous vehicle (see FIG. 105 ) and transmit it to the driving safety analysis server 400 . The vehicle sensor module 100 may include a sensor unit 110 , a sensor hub 130 , and a vehicle communication unit 150 .
The sensor unit 110 recognizes motions such as speed, acceleration, position, and yaw rate or surrounding environment detected while the autonomous vehicle 105 is driving. The sensor unit 110 may include a speed sensor 111, an acceleration sensor 112, a position sensor 113, a rotation rate sensor 114, and the like to sense motion characteristics of the autonomous vehicle 105. Here, the acceleration sensor 112 may include a longitudinal acceleration sensor for detecting longitudinal acceleration of the self-driving vehicle 105 and a lateral acceleration sensor for detecting lateral acceleration. In addition, the sensor unit 110 may include a lidar sensor 115 or a radar sensor 116 for recognizing objects or situations around the autonomous vehicle.
The sensor hub 130 receives and processes sensing data provided from the plurality of sensors 111 to 116 included in the sensor unit 110 . The sensor hub 130 may periodically or aperiodically receive sensing data randomly transmitted from a plurality of sensors. Sensing data of each of the plurality of sensors 111 to 116 is collected by the sensor hub 130 . The sensor hub 130 may convert sensing data transmitted from the sensor unit 110 into a data format for efficient transmission. The sensor hub 130 may be implemented using a processor or various computing cores.
The vehicle communication unit 150 transmits sensing data provided from the sensor hub 130 through the communication network 200 . For example, the vehicle communication unit 150 may include a wired/wireless communication module supporting vehicle to machine communication (V2X).
It will be well understood that the types of sensors included in the vehicle sensor module 100 are not limited to the illustrated types. That is, the vehicle sensor module 100 may include sensors for detecting various motions of the self-driving vehicle 105 or environment information, including an altitude sensor or a temperature sensor. The vehicle sensor module 100 described above may be integrally or modularly mounted to the self-driving vehicle 105 .
3 is a block diagram showing the configuration of the Internet of Things (IoT) sensor of FIG. 1 as an example. Referring to FIG. 3 , the IoT sensor 200 may include a sensor unit 210 and a communication unit 230. A plurality of IoT sensors 200 may be installed in a plurality of sections for each road section ID.
The sensor unit 210 may include a humidity sensor 211 , a road surface condition sensor 213 , and an air quality sensor 215 . The humidity sensor 211 may be installed on the road surface or roadside to sense the humidity of the road surface. The road surface condition sensor 213 senses the condition of the road surface. The road surface condition sensor 213 may sense the amount of rainwater or ice and snow according to the amount of snow or precipitation on the road surface. The air quality sensor 215 may measure air pollution or the concentration of fine dust. Alternatively, the air quality sensor 215 may sense a visibility condition due to fog or dust. In addition, the sensor unit 210 may detect various information related to road infrastructure and generate infrastructure data.
The communication unit 230 transmits the sensing data provided by the sensor unit 210 as infrastructure data through the communication network 300 . For example, the communication unit 230 supports standards such as 3rd Generation Partnership Project (3GPP), Long Term Evolution (LTE), World Interoperability for Microwave Access (WIMAX), Wi-Fi, 3G, 4G, 5G, and 6G. It may include a wired and wireless communication module that supports.
4 is a block diagram showing the driving safety level analysis server shown in FIG. 1 as an example. The driving safety analysis server (300, see FIG. 1) processes actual road-based driving data transmitted from the vehicle sensor module (100, see FIG. 1) and infrastructure data provided from the IoT sensor (200, see FIG. 1). In this way, the safety level can be evaluated for each section of the autonomous driving road. In addition, the driving safety level analysis server 400 may analyze a causal relationship between the safety level of the autonomous driving road and the road infrastructure. Referring to FIG. 4 , the driving safety analysis server 400 may include a communication unit 320, a storage unit 340, and a control unit 360.
The communication unit 320 includes a receiving unit 321 and a transmitting unit 323. The receiving unit 321 receives driving data transmitted from the vehicle sensor module 100 . The receiving unit 321 may change the driving data transmitted through the communication network 200 (see FIG. 1) into a data format processed by the controller 360. The receiver 321 will deliver the received driving data to the controller 360 . The transmitter 323 may output a causal relationship with the infrastructure of an actual road generated by analyzing the road grading result derived from the driving safety analysis server 400 .
The storage unit 340 may include a vehicle information DB 341 , a driving data DB 343 , a road information DB 345 , and a road rating information DB 347 . The storage unit 340 may be composed of storages for storing data managed by the driving safety analysis server 400 .
The vehicle information DB 341 stores information about the self-driving vehicle 105 . For example, when there are two or more self-driving vehicles transmitting driving data, information for identifying each vehicle may be stored in the vehicle information DB 341 . In addition, the vehicle information DB 341 may store vehicle characteristic information such as the model, weight, and size of the self-driving vehicle 105 .
The driving data DB 343 stores driving data transmitted from the vehicle sensor module 100 under the control of the data collecting unit 363 . The driving data may be a large amount of big data transmitted over a long period of time from the vehicle sensor module 100 in order to derive a driving pattern. Driving information of the autonomous vehicle 105 in the form of raw data may be stored in the driving data DB 343 .
The road information DB 345 may store information about road infrastructure transmitted from an identifier (ID) for each section of a road where driving data of the self-driving vehicle 105 is generated or an IoT sensor 200 (see FIG. 3). . For example, the road infrastructure may include road facilities or information that affect the operation of the autonomous vehicle 105, such as center lines, crosswalks, guide lines, overpasses, intersections, roundabouts, and increase or decrease of lanes. In addition, the road information DB 345 may store a precise road map for each road section and provide it to the control unit 360 . The driving safety level DB 347 stores the driving safety level for each section of the road generated by the control unit 360 .
The control unit 360 collects driving data and infrastructure data transmitted through the communication unit 320 and configures the collected data into a dataset that can be processed through machine learning. The controller 360 processes the generated data set with a deep learning-based clustering algorithm to generate a Road Risk Index (hereinafter referred to as RRI) for each road section. The control unit 360 calculates the driving safety level for each section of the road on which the autonomous vehicle operates based on the generated road risk index (RRI) for each section. In addition, the control unit 360 may analyze the driving safety level generated for each road section in association with road infrastructure. To this end, the controller 360 includes a processor 361, a data collection unit 363, a data set construction unit 365, a driving safety level calculation unit 367, and a driving safety level analysis unit 369. Here, the processor 361 may preferably be composed of hardware, and the data collection unit 363, the dataset construction unit 365, the driving safety level calculation unit 367, and the driving safety level analysis unit 369 are May be provided as software.
The processor 361 may control overall operations of the driving safety analysis server 400 . The processor 361 may patch operation data or infrastructure data received through the communication unit 320 or transmit a control signal. The processor 361 may access the vehicle information DB 341 , driving data DB 343 , road information DB 345 , and driving safety DB 347 of the storage unit 340 . The processor 361 may execute algorithms or program commands constituting the data collection unit 363 , the dataset construction unit 365 , the driving safety level calculation unit 367 , and the driving safety level analysis unit 369 . The processor 361 may be implemented in the form of a system-on-chip (SoC), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or the like.
The data collection unit 363 stores driving data transmitted from the vehicle sensor module 100 (see FIG. 2 ) of the autonomous vehicle 105 in the storage unit 340 . The driving data is big-data generated by driving the self-driving vehicle 105 on a target road section multiple times during a predetermined period. Accordingly, during a period for collecting driving data, driving data may be accumulated for each period and for each road identifier (ID) in the data collection unit 363 . Driving data can be managed in the form of big data accumulated over a specific meaningful period.
The dataset construction unit 365 reconstructs the collected driving data or infrastructure data into a form easy for machine learning. The dataset construction unit 365 processes operation data and infrastructure data into a form for learning with a deep learning-based clustering algorithm. The data set configuration unit 365 analyzes the configuration of driving data and selects variables that affect the driving safety of autonomous driving roads. The dataset construction unit 365 performs preprocessing using, in particular, speed, longitudinal acceleration, lateral acceleration, and yaw rate information. The criteria for selecting data that can be used during data preprocessing are as follows.
First, the dataset construction unit 365 separates driving data by date and by vehicle from the received driving data. Then, the data set construction unit 365 extracts data that satisfies the criteria in consideration of the autonomous driving mode, each sensor, and whether or not there is a communication failure. In addition, the dataset construction unit 365 removes data in which there are missing values of utilization variables from driving data. Subsequently, the dataset construction unit 365 performs normalization to prevent performance degradation of the learning model due to data deviation. As an example of normalization, min-max normalization may be applied to all driving data. That is, the driving data may be normalized to values having a minimum of '0' and a maximum of '1'.
The data set construction unit 365 constructs a data set for dangerous driving behavior of the autonomous vehicle 105 . The dataset configuration unit 365 utilizes data within the road section defined in the current precision road map to classify risk clusters for each road section. That is, the dataset construction unit 365 calculates the length of a section by using the average length of the corresponding road section and the speed in order to select the minimum section length of the road driving safety level. In addition, the dataset constructing unit 365 excludes a road section including the number of data less than the standard based on the calculated minimum road section length from constructing the dataset. In order to unify the number of data of each road section, a piecewise aggregate approximation (PAA) algorithm that compresses to a minimum road section length may be used.
The driving safety calculation unit 367 learns a risk cluster classification model using a deep learning-based clustering algorithm for the dataset constructed by the dataset construction unit 365. The driving safety calculation unit 367 uses Deep Embedded Clustering (DEC), which is a clustering algorithm based on deep learning, in the present invention. The DEC may have a structure in which clustering is performed after dimensionality is reduced by borrowing an encoder structure of an autoencoder model. It has good performance compared to existing clustering algorithms.
Generalizability can be used as an evaluation index to verify the results of clustering. Generalizability is a method of selecting the optimal number of clusters as the number of clusters in which the largest drop occurs by using the ratio of the loss values of the training dataset and the evaluation dataset. Model parameters are defined similarly to existing deep learning models, and the number of nodes in each layer is selected based on the dimension of the dataset (ex. If the length of the data is 20, the number of nodes per layer is 10→5→2 set to dog). If the value of the evaluation index (Generalizability) is low, it is judged as overfitting. Therefore, in order to derive the optimal result, the number of clusters before the value of the evaluation index decreases the most is defined as the optimal number of clusters. In the case of the dataset used in the present invention, the optimal number of clusters was determined to be 7 because the number of clusters between 7 and 8 showed the largest decrease. The driving safety level calculation unit 367 generates a road risk index and a road driving safety level based on the risk index.
The driving safety level analysis unit 369 analyzes the road driving safety level derived through the deep learning-based clustering algorithm in association with road infrastructure. The driving safety level analysis unit 369 may select factors affecting safety from the geometric structure or facilities of road sections belonging to each cluster, the number of lanes, the length of a section, road signs, traffic enforcement equipment, and various facilities. That is, the driving safety level analysis unit 369 may estimate a causal relationship between the infrastructure of the target road provided from the road information DB 345 of the storage unit 340 and each cluster.
5 is a flowchart briefly showing a method of proceeding with road driving safety evaluation performed by the driving safety analysis server of the present invention. Referring to FIG. 5 , the driving safety level analysis server 400 may accumulate driving data and infrastructure data provided from the self-driving vehicle 105 and calculate a driving safety index and road driving safety level from the accumulated data. there is.
In step S110 , the data collecting unit 363 stores driving data and infrastructure data in the driving data DB 343 of the storage unit 340 . Driving data may be accumulated by vehicle, period, and road section and stored in the driving data DB 343 .
In step S120, the dataset construction unit 365 analyzes the collected driving data stored in the driving data DB 343 and selects variables affecting the driving safety of the autonomous driving road. For example, the data set construction unit 365 may select values for the speed, longitudinal acceleration, lateral acceleration, and rotation rate of the self-driving vehicle 105 . At this time, data that do not provide sufficient measurements can be excluded. The data set construction unit 365 performs data pre-processing when selection of data is completed. That is, the dataset construction unit 365 converts the selected data into a form for learning the risk cluster classification model. The dataset configuration unit 365 utilizes data within the road section defined in the current precision road map to classify risk clusters for each road section. That is, the dataset constructing unit 365 calculates the length of a section by using the average length of the corresponding road section and the speed in order to select the minimum section length of the road driving safety evaluation. In addition, the dataset constructing unit 365 excludes a road section including the number of data less than the standard based on the calculated minimum road section length from constructing the dataset. In order to unify the number of data of each road section, a partial aggregation approximation (PAA) algorithm that compresses to a minimum road section length may be used.
In step S130, the driving safety level calculation unit 367 learns a risk cluster classification model using unsupervised learning on the constructed dataset. The driving safety calculation unit 367 may cluster the dataset using deep embedded clustering (DEC), which is a deep learning-based clustering algorithm.
In step S140, the driving safety level calculation unit 367 generates a road risk index by using the learning result of the deep learning-based clustering algorithm. The driving safety degree calculation unit 367 calculates a representative value of each cluster to define a risk cluster. The risk cluster can be defined as the presence or absence of outliers in longitudinal acceleration, lateral acceleration, and yaw rate related to sudden acceleration and sharp turn. In order to determine the presence or absence of outliers, "z-score", a representative statistical method, can be used. If the cumulative probability calculated by substituting the "z-score" into the Cumulative Distribution Function (CDF) is within the top and bottom 10%, it can be selected as a risk cluster.
In step S150, the driving safety level calculation unit 367 evaluates the driving safety level for each section of the autonomous driving road according to the generated road risk index (RRI). For example, driving safety levels such as safety, caution, and danger may be assigned to each road section according to a road risk index (RRI). In addition, the driving safety level for each road section may be analyzed by the driving safety level analysis unit 369 . The driving safety level analysis unit 369 may analyze factors affecting the driving safety level for each road section in association with infrastructure data. The driving safety level calculation unit 367 may analyze the correlation with the road driving safety level in the geometric structure or facilities of the road section belonging to each cluster, the number of lanes, the length of the section, road signs, traffic enforcement equipment, and various facilities. Analysis results can be referred to the traffic control center for real-time road monitoring and establishment of measures for road safety.
In the above, procedures for collecting data, constructing a data set, learning a risk cluster classification model, generating a road risk index (RRI), and generating a driving safety level of the driving safety analysis server 400 according to an embodiment of the present invention are briefly described. It became.
FIG. 6 is a flowchart showing a process of generating road driving safety in FIG. 5 in more detail. Referring to FIG. 6 , the driving safety analysis server 400 accumulates actual road driving data and infrastructure data provided from the self-driving vehicle 105, and calculates a road risk index (RRI) and road driving safety from the accumulated data. evaluation can be performed.
In step S210, the driving safety analysis server 400 collects driving data, infrastructure data, and precise map data of the self-driving vehicle 105. More specifically, the data collection unit 363 stores driving data transmitted from the self-driving vehicle 105 in the driving data DB 343 . Driving data is collected for a predetermined period and accumulated in the driving data DB 343 . The infrastructure data is infrastructure information necessary for road driving safety evaluation, and may include a collision accident for each road section and each lane of each road section, or a vehicle parked and stopped state. In addition, the infrastructure data may include information such as road surface conditions, humidity, and temperature for each section of the road provided from road surface sensor equipment. In addition, the data collection unit 363 may collect a precise map of an autonomous driving road on which the autonomous vehicle 105 is driving.
In steps S220 to S228, the dataset construction unit 365 transforms the collected driving data or infrastructure data stored in the driving data DB 343 into a data-set in a form suitable for a deep learning-based clustering algorithm. composed of First, in step S220, the dataset construction unit 365 removes data with errors from the collected data and extracts only normal data. In step S222, the dataset construction unit 365 classifies the extracted data by road section and merges the data on a road section basis. In step S224, the dataset construction unit 365 checks whether the number of data exceeds the reference value n. If the number of data exceeds the reference value n (direction 'yes'), the procedure moves to step S226. On the other hand, if the number of data is less than the reference value (n), the process returns to step S220 and continues data extraction. In step S226, the dataset constructing unit 365 applies a partial aggregation approximation (PAA) algorithm to each of the merged data (velocity, longitudinal acceleration, lateral acceleration, turn rate, etc.) for each road section and processes them. In step S228, the dataset construction unit 365 configures the rearranged data into time series data according to the partial aggregation approximation (PAA) algorithm into a dataset for application to the machine learning algorithm.
In steps S230 to S232, the driving safety level calculation unit 367 learns a risk cluster classification model to which Deep Embedded Clustering (DEC) is applied to the constructed dataset. In step S230, the DEC may use a structure in which the dimension is reduced by borrowing the encoder structure of the autoencoder model, and then clustering is performed. Generalizability can be used as an evaluation index to verify the results of clustering. Using generalizability, the optimal number of clusters can be selected as the number of clusters in which the ratio of loss values between the training dataset and the evaluation dataset has decreased the most. Model parameters are defined similarly to existing deep learning models, and the number of nodes in each layer is selected based on the dimension of the dataset (for example, if the length of the data is 20, the number of nodes per layer is 10 → 5 → Set to 2). If the value of the evaluation index (Generalizability) is low, it is judged as overfitting. Therefore, in order to derive the optimal result, the number of clusters before the value of the evaluation index decreases the most is defined as the optimal number of clusters. In the case of the dataset used in the present invention, the optimal number of clusters was determined to be 7 because the number of clusters between 7 and 8 showed the largest decrease. The driving safety level calculation unit 367 generates a road risk index and a road driving safety level based on the risk index. The dataset may be clustered into normal clusters and risk clusters by clustering in step S232.
In step S240, the driving safety degree calculation unit 367 calculates a road risk index (RRI:) from the results of the clustering. The driving safety degree calculation unit 367 calculates a representative value of each cluster to define a risk cluster. The risk cluster can be defined as the presence or absence of outliers in longitudinal acceleration, lateral acceleration, and yaw rate related to rapid acceleration, rapid deceleration, and sharp turns. In order to determine the presence or absence of outliers, "z-score", a representative statistical method, can be used. If the cumulative probability calculated by substituting the "z-score" into the Cumulative Distribution Function (CDF) is within the top and bottom 10%, it can be selected as a risk cluster. To calculate the road risk index (RRI), the number of each road section data included in the risk cluster and the total number of data are calculated. In addition, the ratio of the number of risk clusters to the total number of data may be calculated as a road risk index (RRI, 0 to 1).
In step S250, the driving safety level calculation unit 367 evaluates the driving safety level for each section of the autonomous driving road according to the generated road risk index (RRI). For example, driving safety levels such as safety, caution, and danger may be assigned to each road section according to a road risk index (RRI).
In the above, procedures for collecting data, constructing a data set, learning a risk cluster classification model, generating a road risk index (RRI), and generating a driving safety level of the driving safety analysis server 400 according to an embodiment of the present invention have been described. . With reference to the additionally generated driving safety level, the factors affecting the driving safety level for each road section can be analyzed in association with the infrastructure data. The driving safety level analysis unit 369 may analyze the correlation with the road grade in the geometric structure or facilities of the road section belonging to each cluster, the number of lanes, the length of the section, road signs, traffic enforcement equipment, and various facilities. Analysis results can be used by the traffic control center to monitor roads in real time and establish measures for road safety.
7 is a diagram showing a road map in which driving data of an autonomous vehicle is collected according to an embodiment of the present invention. Referring to FIG. 7 , empirical big data collected while the level 4 self-driving shuttle runs on the route shown in the road map 500 can be used as the driving data of the present invention.
Empirical big data can be generated by accumulating data driven by two self-driving shuttles equipped with sensors and communication terminals shown in FIG. 2 . In addition, infrastructure data is collected through IoT-based road condition measurement sensors built in the road map 500 . The self-driving shuttle transmits the driving data detected during operation to the integrated control center or the driving safety level analysis server 400 through means such as V2X communication. In addition, the infrastructure data sensed through the IoT-based sensors may be provided to the control center or the driving safety level analysis server 400 provided in the control center in real time.
8 is a view showing a part of a precision road map according to an embodiment of the present invention. Referring to FIG. 8, a detailed road map 510 of an actual road on which an autonomous vehicle 105 is operating, a node 511 for dividing lanes, and a road section 512 forming a cluster of the present invention are exemplified. show hostile
The road map 500 is divided into a plurality of road sections based on the precise road map 510 . The minimum road section length set to classify the risk cluster for each road section may be a value obtained by dividing the section distance by the speed. Here, the speed may be set to 25 km/h according to domestic regulations for obtaining a level 4 autonomous driving temporary license. And among the road sections of the calculated minimum road section length, values with the number of data less than the standard are excluded from the dataset configuration. In addition, the PAA algorithm will be applied to unify the number of data of each road section.
FIG. 9 is a graph briefly showing an example of the partial aggregation approximation (PAA) algorithm of the present invention mentioned in FIG. 7 . Referring to FIG. 9 , a method of applying a partial aggregation approximation (PAA) algorithm for constructing a data set for speed among variables of travel data is illustrated as an example.
The speed of the self-driving vehicle changes for each road section or over time, and therefore, the input speed may be provided in different numbers according to unit time or section. In order to construct a data set of such speed, a partial aggregation approximation (PAA) algorithm is applied as a means to divide time series into equal time intervals and use the average value in each interval as a representative value.
For example, as illustrated, speed values in a time series period of 50 seconds may be provided as 50 measured values 521 . However, when unifying 7 segments (Segments = 7) to construct a dataset, a time interval of 50 seconds is divided into 7 unit intervals, and the average value of the speed in each unit interval is 7 representative values (523 ) are mapped to In this way, the velocity, lateral acceleration, longitudinal acceleration, and rotation rates are processed into a dataset. By applying the partial aggregation approximation (PAA) algorithm to all variables constituting driving data, it can be processed into a dataset for applying a deep learning-based clustering algorithm.
10 is a diagram briefly showing a method of constructing a deep embedded clustering (DEC) algorithm of the present invention. Referring to FIG. 10 , the deep embedded clustering 545 of the present invention may be configured using an encoder of an autoencoder 530 and a hidden layer 535.
The autoencoder 530 is an algorithm widely used as an artificial neural network model for unsupervised learning. The autoencoder 530 is a model used for feature extraction through data compression and restoration.
The deep embedded clustering 545 is configured by borrowing the encoder and hidden layer 535 of the autoencoder 530 model. Clustering 550 may be performed by reducing the dimension of data using deep embedded clustering 545 . Generalizability can be used as an evaluation index for verifying the results of clustering 550 . Using generalizability, the optimal number of clusters can be selected as the number of clusters in which the ratio of loss values between the training dataset and the evaluation dataset has decreased the most.
11 is a diagram exemplarily illustrating a driving safety evaluation result of an autonomous driving road and an analysis method thereof according to an embodiment of the present invention. Referring to FIG. 11 , the DEC learning result 610 and the road driving safety evaluation result 620 may be combined and analyzed in association with the driving safety level of the autonomous driving road and the road infrastructure.
12 is a graph exemplarily showing the DEC evaluation index (Generalizability) for each number of clusters according to an embodiment of the present invention. Referring to FIG. 12, the evaluation index (Generalizability) showed a trend as shown. In the present invention, the section with the largest drop was selected as the optimal number of clusters. That is, when the number of clusters is 7, the drop to 8 becomes the maximum. This means that when the number of clusters is 7, the optimal number of clusters is obtained.
13 is a graph exemplarily showing results of deep embedded clustering (DEC) according to an embodiment of the present invention. Referring to FIG. 13 , representative values for the 7 clusters determined in FIG. 12 may be generated.
First, to define risk clusters, representative values of each cluster must be calculated. The risky driving behavior cluster can be defined as the presence or absence of outliers in longitudinal acceleration, lateral acceleration, and yaw rate related to sudden acceleration and sharp turning. In order to determine the presence or absence of outliers, 'z-score', a representative statistical method, can be used. If the cumulative probability calculated by substituting the "z-score" into the Cumulative Distribution Function (CDF) is within the top and bottom 10%, it can be selected as a risk cluster.
In the case of the dataset used in the present invention, cluster 1 showed outliers of 0.96, 0.05, and 0.01 in longitudinal acceleration, lateral acceleration, and yaw rate, respectively. And in cluster 3, the longitudinal acceleration was 0.07, and in cluster 6, the lateral acceleration was judged to be an outlier of 0.08. Therefore, cluster 1, cluster 3, and cluster 6 can be selected as risky driving behavior clusters.
14 is a table briefly showing how to calculate the risk index and the inclusion ratio of risky driving behavior clusters for each road section. Referring to FIG. 14 , a road risk index (RRI) may be determined according to the number of risky driving behavior clusters for each road section.
The road segment ID (A3LI18BA001066) includes 1 cluster 1 and 31 cluster 3. A total of 40 clusters are included in the section ID (A3LI18BA001066). Then, the road risk index (RRI) becomes a value obtained by dividing the number of risky driving behavior clusters by the total number of clusters. That is, a road risk index (RRI) of 32÷40=0.8 is derived.
In the road segment ID (A3LI18BA000063), 4 clusters 1, 2 clusters 3, and 17 clusters 6 are included among a total of 42 clusters. Therefore, the road risk index (RRI) of the road section ID (A3LI18BA000063) is derived as 23÷42=0.55.
In the above-described manner, a road risk index (RRI) may be calculated for each road section of an autonomous driving road.
15 is a diagram visually showing a driving safety level for each road section of a final autonomous driving road based on a road risk index (RRI). Referring to FIG. 15, when the road risk index (RRI) is 0.67 or more, 'danger' is evaluated, and when the road risk index (RRI) is 0.33 or more and less than 0.67, 'caution' and the road risk index (RRI) are If it is less than 0.33, it can be evaluated as 'safe'. It is possible to evaluate driving safety as shown according to the road section ID.
16 and 17 briefly show an example of analysis in connection with road driving safety and road infrastructure according to the present invention.
Referring to FIG. 16 , a road section 710 evaluated as a 'dangerous' level corresponds to a road section immediately after a left turn at an intersection. The opposite direction to the traveling direction of the road section 710 is a U-turn section, and the right 4 lanes are used as merging lanes and correspond to a narrowing road section. Therefore, it can be interpreted that the risk of driving on the road of an autonomous vehicle is high due to the mixture of vehicles turning left, U-turn vehicles, and merging vehicles through the first and second lanes of the road section 710.
Referring to FIG. 17 , a road section 810 evaluated as a 'dangerous' level corresponds to a road section immediately before a left turn at an intersection. A bus stop and a taxi stand are located on the right side of the road section 810 in the direction of travel. Therefore, it can be interpreted that the risk of driving the autonomous vehicle on the road always exists when the taxi and the bus merge into the left turn lane.
It can be applied to improving the driving safety of the self-driving car 105 by deriving a road risk index (RRI) based on the operation data obtained from the self-driving car on the actual road described above and evaluating and analyzing the driving safety level. there is. In addition, in the case of an integrated control center that monitors the self-driving vehicle 105, it is possible to prevent an accident by transmitting the risk level of the corresponding region to the self-driving vehicle 105 in advance through the road driving safety level. In addition, the road safety index (RRI) is periodically updated using driving data and infrastructure data collected in real time, and traffic measures are established to mitigate the ratio of road sections with risk ratings in the future by evaluating the factors influencing the road risk index. can contribute
What has been described above are specific embodiments for carrying out the present invention. The present invention will include not only the above-described embodiments, but also embodiments that can be simply or easily changed in design. In addition, the present invention will also include techniques that can be easily modified and practiced using the embodiments. Therefore, the scope of the present invention should not be limited to the above-described embodiments and should not be defined by the following claims as well as those equivalent to the claims of this invention.

Claims (5)

In the driving safety analysis server that receives driving data provided from an autonomous vehicle and infrastructure data transmitted through an Internet of Things (IoT) sensor:
a communication unit receiving the driving data and the infrastructure data;
a storage unit for storing the received driving data or the infrastructure data;
Including a control unit that derives a road risk index for each road section through a deep embedded clustering algorithm from the accumulated driving data or the infrastructure data, and evaluates the driving safety for each road section based on the road risk index; ,
The control unit:
controlling the communication unit and the storage unit to accumulate the operation data or the infrastructure data in the storage unit for a specific period of time;
Classifying the accumulated driving data or infrastructure data for each section of the road, configuring the classified driving data as a dataset for the Deep Embedded Clustering algorithm,
learning a risk cluster classification model with the deep embedded clustering algorithm using the dataset; and
Calculate the road risk index and the driving safety level for each road section from the learning result of the risk cluster classification model, and analyze the driving characteristics of the autonomous vehicle for each road section by linking the driving safety level and the infrastructure data,
The Deep Embedded Clustering Algorithm is configured using an encoder of an autoencoder and a hidden layer, and uses a generalizability representing a ratio of loss values between a training dataset and an evaluation dataset. It is used as an evaluation index to verify the clustering result of
The driving safety analysis server for determining the number of clusters in the section in which the generalizability has decreased the most as the number of clusters in the risk cluster classification model. According to claim 1,
Wherein the driving data includes at least one of information about speed, lateral acceleration, longitudinal acceleration, and rotation rate of the self-driving vehicle. According to claim 2,
The controller configures the dataset by processing the speed, the lateral acceleration, the longitudinal acceleration, and the rotation rate with a piecewise aggregate approximation (PAA) algorithm. According to claim 3,
The control unit substitutes the 'Z-score' of the longitudinal acceleration, the lateral acceleration, and the rotation rate into a cumulative distribution function (CDF), and assigns a cluster corresponding to the upper and lower 10% of the cumulative probability as a risky driving behavior cluster. A driving safety level analysis server that judges. According to claim 4,
Wherein the control unit determines the road risk index and rating for each road section based on the risky driving behavior cluster.

Images