TWI813971B - Individually adapted road risk prediction system, method and user equipment - Google Patents

Abstract

The present invention relates to a individually adapted road risk prediction system, comprising; a system server comprising a traffic risk prediction master computer program product; and a user equipment communicatively connected with the system server and providing a front-end computer program product for a user to operate to enable a traffic risk prediction machine learning model program component well-trained and included in the traffic risk prediction master computer program product, in which the traffic risk prediction machine learning model program component is performed to implement a prediction for computing a risk map and selectively compute a personal risk indicator and accordingly provides at least one personally recommended route for the user.

Description

Personalized road risk prediction system, method and user equipment

The present invention relates to a road risk prediction system, method and user equipment, in particular to a system, method and user equipment that predict road risks based on a machine learning model and consider personal risk tolerance index.

Due to the narrow and densely populated area of Taiwan, the population prefers to concentrate in metropolitan areas, resulting in complex urban traffic conditions and frequent traffic accidents. Especially in recent years, the rise of emerging logistics industries, such as motorcycle logistics, food delivery, etc., has resulted in endless traffic accidents, not only causing Property damage often results in casualties.

Take Taoyuan City as an example. According to statistics from the Taoyuan City Government, there were a total of 15,286 A2 traffic accidents in Taoyuan City in 2009. A2 traffic accidents refer to accidents involving injury or death for more than 24 hours, resulting in a total of 20,048 injuries. However, By 2018, the number of accidents had increased to 30,689, resulting in 40,595 injuries. In just ten years, both the number of accidents and the number of injuries have approximately doubled, which proves that Taiwan's traffic safety problems are becoming increasingly serious.

There are many factors that cause traffic accidents, such as driver factors, weather, road conditions, etc., which may lead to traffic accidents. These factors are often difficult to predict and quantify, making the prevention of traffic accidents doubly difficult. In recent years, in addition to government units continuing to Amend regulations and improve In addition to road design, there are also many research and invention attempts to solve the problem of urban traffic accidents from different aspects.

For example, some scholars have proposed using regression analysis models to establish a basic relationship model between traffic accidents and road congestion, and provide information to drivers to avoid driving on congested road sections. Some scholars have also proposed using vehicle-to-vehicle -to-vehicle, V2V) communication, automatically adjusts vehicle speed and distance. In addition to reducing the probability of accidents, it also makes highway traffic flow smoother. Fuzzy theory is also used to analyze accident hot spots and causes, or deep learning is used to analyze weather conditions and traffic accidents. relationship.

U.S. Invention Patent No. US 8,188,887 B2 discloses a vehicle monitoring system installed on a vehicle, which can be coupled with the on-board computer to obtain the current real-time status data of the vehicle and connect to the central server through the Internet. Electronic maps and real-time traffic information on the driving route, such as road speed limits, etc., the system can also combine vehicle condition data and road condition information to issue driving prompts to users; U.S. Invention Patent Publication No. US 2016/0061625 A1, has been disclosed Technology that stores traffic accidents on a central server according to location and type, provides users with real-time access, and can issue warnings to users based on their location.

Taiwan Invention Patent No. I 702563 has disclosed a method and system for real-time notification of road status. When the user device enters a certain road section, a registration message is sent to the management server to register the group containing the road section. When using When the user device sends a notification message to the management server, the management server can send the notification message to all registered user devices in the group containing the road segment where the user device is located. When the user device leaves the road segment, an unregistration message is sent. Go to the management server to unregister the corresponding group that at least includes the leaving road section. This technology can quickly and effectively notify the vehicle of unexpected road conditions ahead, and can instantly transmit traffic events. message to all relevant vehicles on the road.

The inventor of this case disclosed in Taiwan Invention Patent Application No. 109104727 "Map-based adaptive variable range traffic condition detection method and device and computer program product thereof" that dynamically adjust events based on the user's position, forward direction and moving speed. The detection range technology, combined with the detection of a possible road incident ahead, improves the audio and video quality of the video stream, or triggers the cloud driving recorder to start, and records longer seconds and better image quality, so as to To prevent key accident scenes from being missed, if there are no road incidents, the video and audio quality can be lowered to save network traffic and power consumption.

However, most of the above-mentioned research and inventions focus on proposing novel theories or algorithm architectures. The solutions are often too theoretical and ideal, or require the execution of a large number of mathematical operations, the use of specific hardware equipment, or the construction cost is too high. High or poor accuracy makes it difficult to reduce to practice in real life, and almost all of these solutions or inventions lack the function of predicting the risk of future traffic accidents.

If a set of technology can be proposed, it can be implemented in real life. It can effectively predict future traffic risk accident hot spots, traffic volume or average speed based on the driver's personalized background information, and predict different traffic risk accidents. The risk probability of traffic accidents along the route can also be notified to the driver in advance, so the driver can select a route that meets his or her own risk tolerance in advance, and the driver will be personally configured according to the driver's adaptability. Setting, or in response to the possible risk of traffic accidents, actively slowing down, increasing alertness, paying more attention to the surrounding road conditions, or driving on detours, or adopting defensive driving will definitely effectively reduce the probability of traffic accidents as a whole.

For this reason, in view of the shortcomings in the conventional technology, the inventor finally conceived the "Personalized Road Risk Prediction" of this case after careful attempts and research, and a spirit of perseverance. "Measurement system, method and user equipment" can overcome the above shortcomings. The following is a brief description of the present invention.

In view of the shortcomings of conventional technologies, the present invention proposes to use the road events provided by the government's open data platform and road events uploaded by users of the self-made application in a self-made vehicle application to continuously utilize the history of past traffic conditions. information, and apply machine learning technology to predict road risk (road risk) traffic conditions at the next point in time, as well as risk map prediction, and then merge the personalized information provided by the user with the risk map obtained by machine learning to calculate Appropriate recommended routes for users to choose as a reference.

The system of the present invention focuses on predicting and analyzing future events, predicts future events through past and current information, gives users a more accurate road risk tolerance index, and recommends road sections with lower risks for users to drive on.

Accordingly, the present invention proposes a personalized road risk prediction system, which includes: a system server, which includes a traffic risk prediction main control computer program product; and user equipment, which is communicated with the system server and provides a front-end computer The program product is operated by the user to drive the trained traffic risk prediction machine learning model program component included in the traffic risk prediction master computer program product. After execution, the traffic risk prediction machine learning model program component implements the prediction of the risk map and Optionally calculate a personalized risk index and provide at least one personalized recommendation path to the user based on this.

The present invention further proposes a personalized road risk prediction method, which includes: installing a traffic risk prediction main control computer program product on the back-end system server; executing the front-end computer program product on the user device for the user to operate to drive the Traffic risk prediction main control circuit The brain program product contains a trained traffic risk prediction machine learning model program component; after execution, the traffic risk prediction machine learning model program component implements the prediction of the risk map and selectively calculates a personalized risk index, and calculates at least one personalized risk index accordingly. Recommended path; and displaying the at least one personalized recommended path to the user through the user device.

The present invention further proposes a personalized road risk prediction user equipment, which is communicated with a system server including a traffic risk prediction main control computer program product. The personalized road risk prediction user equipment includes: a display unit; and a processing unit. A server unit that executes a front-end computer program product for user operation to drive a trained traffic risk prediction machine learning model program component included in the traffic risk prediction master computer program product. The traffic risk prediction machine learning model program component is executed Then, the prediction of the risk map and the selective calculation of the personalized risk index are implemented, and at least a personalized recommended path is generated based on this, and displayed to the user through the display unit.

The above summary is intended to provide a simplified summary of the disclosure to provide readers with a basic understanding of the disclosure. This summary is not a complete description of the disclosure, and is not intended to identify important/critical elements of the embodiments of the disclosure. Define the scope of the invention.

10: Personalized road risk prediction system of the present invention

100: User device

101: Processor unit

103: Communication module

105:Storage unit

107:Display unit

109:Front-end program

110: Vehicle telematics device

115:Mobile device

120:Desktop computer

125:Laptop

130:Smartphone

135:Tablet device

200:System server

210: Traffic risk prediction main control program

211: Traffic risk prediction machine learning model program component

213: Image space feature extraction model

214: Time series dependency learning model

250:Traffic Risk Database

300:Internet

S: starting point

E: End point

A:Path

B:Path

C: path

500: Personalized road risk prediction method of the present invention

501~507: Implementation steps

Figure 1 shows a system architecture view of the personalized road risk prediction system of the present invention;

Figure 2 shows a functional component view of the hardware network equipment of the remote system server and user equipment included in the personalized road risk prediction system of the present invention;

Figure 3 shows a schematic diagram of the hierarchical architecture of the CNN model used in the personalized road risk prediction system of the present invention;

Figure 4 shows a schematic diagram of the internal logical architecture of the LSTM model used in the personalized road risk prediction system of the present invention;

Figure 5 shows a schematic diagram of the multi-layer architecture of the traffic risk CNN-LSTM model used in the personalized road risk prediction system of the present invention;

Figure 6 shows a schematic diagram of the data sorting method before and after processing in the traffic risk CNN-LSTM model used by the personalized road risk prediction system of the present invention after transposition layer and reshaping layer;

Figure 7 shows a schematic diagram of the construction process of the traffic risk CNN-LSTM model of the present invention;

Figure 8 shows a schematic diagram of the input matrix used by the traffic risk CNN-LSTM model of the present invention and the matrix row and column structure of the input matrix;

Figure 9 shows a schematic diagram of the suggested path given by the Google Map navigation function displayed in the front-end program operation interface of the present invention;

Figure 10 shows a schematic diagram of the risk map calculated by the traffic risk prediction machine learning model for the grid prediction range displayed in the front-end application operating interface of the present invention;

Figure 11 shows a schematic diagram of the traffic risk prediction machine learning model superimposed on Google Map recommended routes and risk maps displayed in the front-end application operating interface of the present invention;

Figure 12 shows a schematic diagram of the road risk prediction results calculated by the traffic risk prediction machine learning model displayed in the front-end application operating interface of the present invention based on Google Map recommended routes and risk maps and incorporating personalized risk indexes;

Figure 13 is a schematic diagram showing the road risk prediction results displayed in the web browser by the personalized road risk prediction system of the present invention using a web browser as a front-end program;

Figure 14 shows that the personalized road risk prediction system of the present invention uses a dedicated application for vehicle telematics. A schematic diagram of road risk prediction results displayed in an in-vehicle telematics application using the program as a front-end program; and

Figure 15 shows the flow chart of the implementation steps of the personalized road risk prediction method.

The present invention will be fully understood from the following examples, so that those skilled in the art can implement it. However, the implementation of the present invention is not limited to its implementation form by the following examples; the drawings of the present invention are not There are no limitations on the size, dimensions and scale, and the size, dimensions and scale of the actual implementation of the present invention are not limited by the drawings of the present invention.

The word "preferably" used in this article is non-exclusive and should be understood as "preferably but not limited to". Any steps described or recorded in any specification or claim can be performed in any order, and are not limited to the claims. The order described, the scope of the present invention should only be determined by the appended claims and their equivalents, and should not be determined by the examples of implementation examples; when the word "comprising" and its changes appear in the description and claims, , is an open-ended term with no restrictive meaning and does not exclude other features or steps.

Figure 1 shows a system architecture view of the personalized road risk prediction system of the present invention; Figure 2 shows a view of the hardware network equipment functional components of the remote system server and user equipment included in the personalized road risk prediction system of the present invention; The personalized road risk prediction system 10 of the present invention includes one or more system servers 200 configured at the back-end or server end using a client-server model, and Multiple user equipment (UE) 100 configured at the front-end, client end or presentation layer, the system server 200 and the multiple user equipment 100 are connected via Wired The (wired) or wireless (wireless) Internet 300 establishes a communication link to establish an upload and download communication link for two-way communication and data exchange.

The user equipment 100 is preferably a telematics device 110, a mobile device 115, a desktop computer 120, a notebook computer 125, a smart phone 130 or a tablet device 135, etc. The user equipment 100 is at least Built-in hardware units such as a processor unit 101, a communication module 103, a storage unit (storage) 105 and a display unit 107, which are electrically connected to each other, and a software unit of a front-end program 109. The front-end program 109 is installed on the user device 100 And is loaded and executed through the processor unit 101, preferably such as but not limited to a web browser (web browser), application (App) or micro application (micro App). The user operates the program on the user device 100. The front-end program 109 operates the traffic risk prediction main control program 210 on the server 200, and the display unit 107 is preferably a monitor, an external monitor, or a touch screen.

The traffic risk prediction main control program 210 is installed on the system server 200. The traffic risk prediction main control program 210 is preferably a back-end logic execution and management background, which mainly includes the traffic risk prediction machine learning model program component 211, and also includes Peripheral logic processing components, management program components, interface program components and data exchange program components, etc., are provided to the user to log in and operate the system through the front-end program installed on the user device 100 as a front-end operating interface. The server 200 is communicated with the traffic risk database 250. The traffic risk database 250 can be built and stored directly on the system server 200, or it can be built and stored in another configuration separate from the system server 200 but maintained. In the data layer of the communication link or in a dedicated data server.

When the front-end program is provided as a front-end user interface for user operation in the form of a dedicated application or micro-application, it is best to be built on several mainstream mobile operating systems. Operate under mobile OS, such as but not limited to: Palm mobile operating system, Windows mobile operating system, Android operating system, Apple iOS, Blackberry operating system, etc., especially the two main operating systems of mobile devices, Android system and iOS system. In the Android system part, the application is developed using the Android Studio design editor. The remote server program and software system part preferably use, for example, but not limited to, PHP scripts to access the correlation database MySQLDB to transmit and To receive data, MySQLDB is a multi-user, multi-thread SQL database server that can effectively arrange, file, and tabulate a database software for more efficient subsequent query, sorting, transmission, and reception. Data; For iOS devices, the application is preferably developed using basic components such as but not limited to Xcode and UIKit APIs, and is developed using Objective-C programming language or Swift programming language.

When the front-end program is provided as a front-end user interface for user operation in the form of a car infotainment-specific application, it is better to run it under several mainstream car info operating systems, such as but not limited to QNX Car 2.0, iOS in the Car, CarPlay, Android Auto, Windows CE Automotive and other vehicle information operating systems, or other applications running under the PikeOS embedded vehicle information operating system, or other closed vehicle information operating systems. The vehicle information communication device 110 is preferably mounted on a vehicle, can run the vehicle information operating system and execute applications on the vehicle information operating system, and is equipped with an intelligent human machine interface (HMI), such as voice control and full touch control. Smart devices with information processing, computing and communication capabilities, such as car consoles, graphical car dashboards and gesture controls.

The personalized road risk prediction system and related program components proposed by the present invention, in addition to building the front-end program as a specialized application program, can also be deployed externally (off-premises) using Software as a Service (SaaS) or Platform as a Service. Built as a service (PaaS) cloud technology and provided for use The web browser executed on the user device 100 serves as a front-end program and is directly provided to the user for operation. In the case of using SaaS or PaaS services, the user only needs to obtain permission to access and use the program online. In the system of the invention, the user does not need to install applications on the user device 100 at the same time. For the vehicle telematics device 110 that can execute a web browser, the user does not need to install an application program on the vehicle telematics device 110 .

In one embodiment, the mobile terminal preferably uses the Android platform as the development basis, the backend is preferably developed using the Node.js system, combined with the lightweight data transmission framework express.js, and the database uses the open source system MariaDB, and the entire system is built It is placed in the Kubernetes cloud container management system architecture and achieves system load balancing, resource allocation and uninterrupted service quality through automated deployment, expansion and management of the container application service system.

The processor unit 101 inside the user equipment 100 will receive the instructions issued by the user through the front-end program, and transmit the instructions to the traffic risk prediction main control program 210 on the system server 200 through the communication module 103. The control program 210 will execute instructions accordingly, including for example but not limited to: accessing the traffic risk database 250, including reading the required data or writing the data, and downloading it to the user of the client. On the device 100, the processor unit 101 also refers to an electronic control unit (ECU) or a microcontroller unit (MCU) with a relatively simple structure. Between the front-end program 109 of the user device 100 and the traffic risk database 250 of the system server 200, it is preferable to use a communication protocol such as but not limited to HTTP/HTTPS and cooperate with a JSON format for data exchange and two-way communication.

The traffic risk prediction main control program 210, the traffic risk prediction machine learning model program component 211, the front-end program 109 or other program components, and the remote server program or software system can use any programming language to program (programming) and compile ( compiling), for example But not limited to C, C++, C#, Java, VBScript, Macromedia Cold Fusion, COBOL, Microsoft Dynamic Server Web, assembly language, PERL, PHP, awk, Python, Visual Basic, SQL stored procedures, PL/SQL, any UNIX command Description language, Extensible Markup Language (XML) with algorithms implemented in any combination of data structures, objects, procedures, routines, or other program elements.

The personalized road risk prediction system of the present invention and the included program components are composed of various hardware in the physical equipment layer and application programs/software platforms/computer program products in the application program layer. According to the international open system Interconnection Communication Reference Model (OSI/RM) architecture, program components are software application services that execute and operate on layer 7 (application layer) of the OSI/RM architecture. Software application services at layer 7 can independently choose layer 4. Various communication protocols in the transport layer form data packets and determine the transmission path at the layer 3 network layer. After adding logical link control (LLC) and media access control (MAC) through the layer 2 data connection layer, and Various devices located on the layer 1 physical layer, such as but not limited to multiple user devices 100 and system servers 200, etc., establish required communication links.

The traffic risk prediction machine learning model program component 211 of the present invention is preferably composed of two main components such as an image space feature extraction model 213 and a time series dependency learning model 214. The image space feature extraction model 213 is preferably selected from, for example, but Not limited to: convolutional neural network (CNN) model, regional convolutional neural network (R-CNN) model, fast regional convolutional neural network (Faster R-CNN) model, deep convolutional neural network (DCNN) Model, Recurrent Neural Network (RNN) model, Convolutional Recurrent Neural Network (CRNN) model, Deep Recurrent Neural Network (DRNN) model, Fully Convolutional Neural Network (FCN) model, Multi-column convolutional neural network (MCNN) model, bidirectional neural network (BRNN) model or deep neural network (DNN) model, etc.

The time series dependency learning model 214 is preferably selected from, for example but not limited to: long-term Short-term memory model (LSTM), unsupervised learning model, supervised learning model, deep learning (deep learning) model, autoencoder (AutoEncoder) model, neural network (ANN) model, multi-layer perception (MLP) model, Deep neural network (DNN) model, ensemble learning (ensemble learning) model, etc.

In this embodiment, the traffic risk prediction machine learning model program component 211 of the present invention is preferably composed of a convolutional neural network (Convolution Neural Network, CNN) model combined with a long short-term memory (Long Short-Term Memory, LSTM) model. The traffic risk CNN-LSTM model is taken as an example as the core explanation of the calculation.

Figure 3 shows a schematic diagram of the hierarchical architecture of the CNN model used in the personalized road risk prediction system of the present invention; the CNN model has the following characteristics: (1) it can directly use a two-dimensional matrix as input; (2) weights sharing ) to reduce model training parameters and shorten training time; (3) The spatial feature extraction effect is good. In this embodiment, the CNN model includes a convolution layer, a pooling layer, and a full layer as shown in Figure 3. A fully connected layer, in which the convolution layer performs convolution operations on the input image from left to right and top to bottom with a fixed-size square filter (kernel) to capture the features of the input image. The operation formula of the convolutional layer is listed below:

表示第l層的第q個特徵圖(feature map)，

為第l層的濾波器，

為偏移量(bias)，M _q 為輸入特徵圖，*為卷積運算，g為激勵函數(activation function)，在本發明系統實施例，較佳是選擇使用線性整流單位(ReLU)函數做為激勵函數，具有收斂速度快、效率高等特性。 in

Represents the q- th feature map of the l-th layer,

is the filter of the lth layer,

is the offset (bias), M _q is the input feature map, * is the convolution operation, and g is the activation function. In the system embodiment of the present invention, it is preferable to use the rectified linear unit (ReLU) function. It is an excitation function and has the characteristics of fast convergence speed and high efficiency.

The pooling layer is used in CNN to reduce the amount of data and retain important information. Common pooling methods include max pooling, mean pooling, etc., but pooling will destroy the order. relationship, and since the convolutional layer calculation results need to be transmitted to the LSTM layer later, it is not suitable to add a pooling layer. Therefore, the CNN used in the system of the present invention does not add a pooling layer to prevent temporal features from being lost in the pooling process.

The fully connected layer is a neural network (NN) layer. The two-dimensional feature map can be converted into a one-dimensional output result through the activation function. The operation formula of the fully connected layer is as follows, where w ^l is the weight matrix ( weight matrix), b ^l is the offset vector, g is the excitation function:

x ^l = g ( w ^l x ^{l -1} + b ^l )

In the system of the present invention, the feature extraction part will be handled by the convolutional layer in CNN. In addition to the pooling layer that is eliminated in order to retain the time sorting features, the fully connected layer will also be moved to the end of the entire model architecture for output use.

Figure 4 shows a schematic diagram of the internal logical structure of the LSTM model used in the personalized road risk prediction system of the present invention; the LSTM model enhances long-term memory through the use of memory units, such as adding input gates, forget gates and The configuration of the output gate allows the model to learn on its own whether the data stream is input, memorized and output, and filter out important data accordingly to solve the problem that occurs when the Recurrent Neural Network (RNN) memorizes a long time series. The gradient vanishing or gradient exploding problem can effectively capture the dependencies of longer time series data.

The internal calculation formulas of the LSTM model are as follows:

f _t = σ ( w _f [ h _{t -1} , X _t ]+ b _f )

i _t = σ ( w _i [ h _{t -1} , X _t ]+ b _i )

o _t = σ ( w _o [ h _{t -1} , X _t ]+ b _o )

C _t = f _t * C _{t -1} + i _t * tanh( w _c [ h _{t -1} , X _t ]+ b _c )

h _t = o _t * tanh( C _t )

Among them, f _t , i _t , o _t represent the forget gate, input gate and output gate respectively, h _{t -1} is the previous state, X _t is the input, C _{t -1} and C _t are the previous and current memory units respectively. State, w _f , w _i and w _o are weight matrices, b _f , b _i and b _o are bias vectors, σ is a sigmoid function, * is element-wise Multiplication operator.

The objective function of the CNN-LSTM model is defined by a loss function (loss function) to minimize the loss (loss) and improve the accuracy as much as possible as the model training goal. Commonly used loss functions such as mean square error (MSE) function or mean absolute error (MAE) function, etc., when using MSE as the loss function, because of the square operation, it is more sensitive to outliers (outlier) than MAE. When the data contains outliers, the loss will be amplified, prompting the model to Correcting the model to improve the amplified error will actually reduce the accuracy of other items. MAE does not amplify outliers. The weight of each result is the same, and the model will not change the model significantly for one outlier.

The system of the present invention considers that low loss does not mean high accuracy, and the original data of road incident risk data easily contains outliers. Considering that the model does not want to be excessively biased towards outliers, the system of the present invention selects MAE as the CNN-LSTM. The objective function of the model.

Figure 5 shows a schematic diagram of the multi-layer architecture of the traffic risk CNN-LSTM model used in the personalized road risk prediction system of the present invention; the present invention proposes to combine the CNN model and the LSTM model as the main machine learning model, with a permute layer, a re- The reshape layer, the flatten layer and the dense layer constitute the machine learning model used in the present invention. The complete model architecture is shown in Figure 5.

Figure 6 shows a schematic diagram of the data sorting method before and after processing by the transposition layer and the reshaping layer in the traffic risk CNN-LSTM model used in the personalized road risk prediction system of the present invention; the CNN model can retain image position information Under the conditions, the extracted image is the feature of the input matrix, and the LSTM model retains the temporal correlation of the input data. However, because the CNN model is a neural network that extracts spatial features, and the LSTM model is a neural network that analyzes time series, both They cannot be directly connected without processing, so the system of the present invention adds a transposition layer and a reshaping layer between the CNN model and the LSTM model to convert the data sorting dimension into time sorting, and finally connects the flat layer to convert it into a data flow sequence, providing LSTM model is trained.

In order to avoid overfitting of the model, and to avoid the rapid increase in training time and hardware computing cost when the number of neural network layers is too large, but there is no corresponding improvement in accuracy, after trial and error, the system of the present invention The maximum layers of the CNN model and LSTM model are set to 5 and 3 layers respectively.

Figure 7 shows a schematic diagram of the construction process of the traffic risk CNN-LSTM model of the present invention; the original data used by the system of the present invention is obtained from the government open data platform, including Taoyuan City's A2 traffic accident data in 2018 and 2019, A2 Traffic accidents refer to accidents that cause injuries or death for more than 24 hours, and are obtained from other private road incidents or traffic map platform service providers. The obtained raw data are converted into CNN through data preprocessing. -The data form that the LSTM model can process. Data pre-processing includes steps such as data preprocessing, generating input and output matrices, and segmenting training sets and validation sets.

The system of the present invention collects and regularly updates original data through the following methods. It uses the open data platform provided by the government or the road traffic data returned by users of self-made vehicle applications. Information and other data sources are used to continuously carry out statistics on road incidents in various regions. The self-made vehicle-mounted application stores the data returned by users in the cloud database for statistics, and integrates open data provided by the government. The crawler program included in the system of the present invention The component will automatically browse and access the government's open database on a regular basis. When the system detects that the government's open data has been updated, it will automatically download it to the traffic risk database 250, filter out the new data and integrate it with the old data. And integrated with own data stored in the traffic risk database 250.

In terms of self-owned data collection, after the system is set, it will automatically incorporate new data generated by each user operation into the model, and then retrain and update the model one by one, or automatically count the data provided by users of the self-made car app when it reaches a certain level. When the number is reached, data statistics and preprocessing are performed. Using an online training architecture, the system will start training the next new model and confirm whether the new model is accurate through model verification. First, update the models of some users to test the model. The stability and accuracy of the model allow the model proposed by the present invention to evolve over time and the amount of data increases through the above architecture, further improving the reliability and credibility of the model and making the model's prediction results closer to the real-world situation.

Data preprocessing includes the following steps: (1) Delete unnecessary fields in the original data, such as but not limited to the address, number of fatalities and injuries, and vehicle type; (2) Extract the specified time period, such as January 1, 2018 to December 2019 data between the 31st of the month; (3) extract the accident data according to the administrative area according to the longitude and latitude range of the accident; (4) correct the time format of the original data, for example, correct "year, month, day" and "hour, minute" to "/" and ":", and changed the Republic of China year to the AD year.

Figure 8 shows a schematic diagram of the input matrix and the matrix row and column structure of the traffic risk CNN-LSTM model used in the traffic risk CNN-LSTM model of the present invention; after the data pre-processing is completed, the spatial input matrix needs to be constructed. In this embodiment, it is given every 1 hour as time interval (time interval), create an hourly spatial input matrix. There are 24 hours in a day, so 24 matrices will be formed, and there will be 8,760 matrices in a year. Taking two years of data as an example, a total of 17,520 matrices need to be constructed.

Then the spatial pitch needs to be given to establish a spatial grid and convert the corresponding longitude and latitude differences. For example, but not limited to, it is better to give a pitch of 100 meters and cut the map into one grid every 100 meters. , then the latitude and longitude differences are 0.0008 and 0.001 respectively. Taking the Zhongli urban area from 20:00 on March 2, 2018 to 21:00 on March 2, 2018 as an example, a total of 2 traffic accidents occurred during this period, respectively at the latitude and longitude of (24.96523,121.2044), ( 24.93738,121.23228), so except for the row and column values of (1,36) and (29,1), which are 1, the input matrix for this period is all 0. The value of each element in the input matrix represents the current time of the region. Accident factors, such as but not limited to car accidents, weather, road width, road type, speed limit, number of lanes, etc.

The spacing can also be given as 200 meters or 50 meters. However, considering that too large a spacing will make the warning range too large, it will have no real meaning to the user. A small spacing will cause too much training time and hardware consumption. Therefore, it is necessary to balance the practicality of model prediction and For training cost, in this embodiment, it is better to choose the given spacing as 200 meters, 100 meters and 50 meters, respectively generating three spatial input matrices of 20×20, 40×40 and 80×80. Each spacing corresponds to a different Accuracy, memory usage, training time.

The system of the present invention allows users to dynamically adjust spacing or range cutting. Users can dynamically adjust the size and range of map grid cutting according to their own needs or according to different accuracy requirements. In actual use scenarios, larger grids The cutting size may cause the accuracy of the model to decrease, but a smaller cutting size will cause clutter in the mobile screen map. Users can set the size of the grid cutting by themselves.

The output matrix is defined as consisting of only 12 pairs of numbers. Each pair of numbers represents the corresponding row and column values of the predicted traffic accidents in the total number of time-sharing accidents matrix. If all road events in that period If the quantity is less than 12 pieces, it will be filled with 0 value, that is, [0,0]. In this embodiment, the data generated in the same time period of the previous week is used as the input matrix, that is, the data 168 hours ago, and the data in the time period of the prediction time is used as the basis for generating the output matrix.

After data preprocessing, a total of 17,520 spatial input matrices were generated, which were divided into two parts, the training dataset and the validation dataset, roughly according to a ratio of 8:2. The validation set can also be further divided into Testing dataset, the training set will be used to train the traffic risk CNN-LSTM model proposed by the present invention, and the verification set will be used to verify whether the CNN-LSTM model established after training is unbiased.

When the system is training the model, it will sort the processed data according to time, input the matrix y of the previous period of time as the feature, and the actual road incident results as markers, and train accordingly, for example: 7 days is used as the An interval (y=7days), use the data that occurred seven days ago to train the traffic risk CNN-LSTM model.

According to the above process, more traffic accident factors, such as but not limited to: weather, holidays, traffic flow, etc., are added to the spatial input matrix, and the traffic risk CNN-LSTM model is trained to learn to consider more traffic accident factors, and to have a longer The CNN-LSTM model is trained on the original data of traffic accident factors over time to improve the accuracy of the model in predicting accidents.

The traffic risk CNN-LSTM model proposed by the present invention is further combined with information and communication (ICT) technology, and uses Android or iOS mobile platform technology, platform as a service (PaaS) technology, and software as a service (SaaS) technology to achieve on-the-move In the form of a front-end program running on the device, such as an application or web browser, or in the form of a specific or closed front-end application running on the vehicle telematics device, it is provided to passers-by for operation and use, allowing passers-by to perform operations in advance Good contingency measures can indeed reduce the incidence of traffic accidents and create a safer and healthier urban traffic environment.

The system can predict the risk probabilities of various routes based on the current location of passers-by and the real-time road information system, and summarize them in the form of estimated time of arrival (ETA), so that passers-by can choose the route. The system will also When a road incident occurs or may occur in the future, it will be broadcast to passers-by, that is, system users. Users can respond early to prevent accidents, and can also trigger the cloud driving recorder to record for a longer period of time to protect the rights of users. .

After the model completes training, the model execution process will use the data from the previous period as input data. The model will predict the car accident risk prediction at the corresponding time in the area. For example, given y=7 (days), the prediction time is March 9 At 1 p.m. on March 2, the model will use the car accident data at 1 p.m. on March 2 as input data, and output a 2×12 matrix. The output matrix is the row and column values of the car accident event in the map cut above, and will return the wisdom. The mobile phone draws the corresponding color in the corresponding grid based on the returned row and column values. The user can see the risk of passing the road section on the mobile phone map.

When actually driving a vehicle, each driver has different risk-taking capabilities. When comparing the driving safety of the elderly and young people on the same navigation route, the risks between the two cannot be treated equally. The elderly due to their physical Deterioration of functional and mental status usually leads to lower risk-taking capacity.

Therefore, the system of the present invention allows users to customize risk maps based on their own situations and conditions. The system will conduct a personal risk tolerance assessment based on the user's personalized traffic information and customize the risk map for the user. The system evaluates the user's risk tolerance characteristics based on the following personal risk tolerance index formula:

R = Wa × A + Ws × S + Wt × T + Wc × C + Wh × H + M

Among them, R personal risk tolerance index (risk value), Wa is the user's age weight, A is the user's age, Ws is the user's gender weight, S is the user's gender, Wt is the user's vehicle type weight, T is the user's vehicle type , Wc is the road time weight, C is the road time, Wh is the climate condition weight, H is the climate conditions, etc., M is more personalized information, and users can customize more personalized information.

The system of the present invention will automatically calculate the risk assessment value and personal risk tolerance index of each area of the map for different attributes such as driving age, driving experience, gender, time, location and vehicle operation, and customize the risk assessment value for individual users. Risk map, and through the navigation planned route and combining the personal risk tolerance index and the risk map, calculate the risk assessment value of each route to recommend a safe or lower-risk route for passers-by.

When the system is actually operating, it evaluates the personal risk tolerance index and brings the calculated risk value into the calculation process of the risk map to provide a path suitable for the user's personal risk tolerance index and conducts analysis on the characteristics of the driver who caused the car accident. Statistics, grouping, and machine learning models are used to predict the risk value of drivers.

Figure 9 shows a schematic diagram of the recommended path given by the Google Map navigation function displayed in the front-end program operation interface of the present invention; the actual usage scenario of the present invention is as follows. The user wants to move from the starting point S to the end point E. The user is moving Start the front-end program 109 on the device 115, preferably a self-made application. The application of the present invention will first load the Google Map road map and obtain the recommended path given by the Google Map navigation function. In this embodiment, the Google Map navigation function A total of three paths A, B, and C are recommended, and the ETA of these three paths is estimated to be 20 minutes.

Figure 10 shows a schematic diagram of the risk map calculated by the traffic risk prediction machine learning model for the grid prediction range displayed in the front-end application operating interface of the present invention; Figure 11 shows a schematic diagram of the traffic risk prediction machine learning model displayed in the front-end application operating interface of the present invention superimposed on the Google Map recommended route and the risk map; then the system of the present invention starts to detect the three routes A, B, C together with the appropriate prediction range of the starting point S and the end point E, grids the prediction range, and executes the traffic risk prediction machine learning model of the present invention, calculates the road risk prediction, and generates a risk map and displays it on the mobile device 115 front-end application In the operation interface, as shown in Figure 10, it is then overlapped with the Google Map suggested route and displayed in the front-end application operation interface of the mobile device 115, as shown in Figure 11.

Figure 12 shows a schematic diagram of the road risk prediction results calculated by the traffic risk prediction machine learning model displayed in the front-end application operating interface of the present invention based on the Google Map recommended route and risk map and incorporating the personalized risk index; then the system of the present invention According to the personal risk tolerance index formula, the user's risk tolerance characteristics are evaluated, and the different risk values of each path A, B, and C are calculated respectively, and the risk value of each path is specifically converted into ETA and provided to the user for reference. .

In this example, the risk value of path A is calculated to be 16, which will affect the ETA + 2 points when converted into time, which means that the ETA of path A is adjusted to 22 points after risk prediction, and the risk value of path B is calculated to be 13 , when converted into time, will affect the ETA + 1 point, which means that the ETA of path B is adjusted to 21 points after risk prediction. The risk value of path C is 11 after calculation, and when converted into time, it will affect the ETA - 1 point, which means path C. After risk prediction, the ETA is adjusted to 19 points. Path C can be regarded as a personalized recommended path provided by the system of the present invention, but the user can freely select paths from paths A, B, and C and start navigation.

For example, the road risk assessment of route A comes from the fact that when encountering consecutive holidays or specific festivals, it is often encountered that the navigation estimate time cannot reflect the actual situation in a timely manner, resulting in As a result, drivers misestimate their departure or arrival time, or when they want to travel the next day, they are often unable to confirm when is the best departure time to avoid heavy traffic jams. Therefore, the system of the present invention collects real-time traffic information through machine learning algorithms. , make forward-looking traffic flow predictions (Proactive), and analyze the possible range of actual arrival time. It can also deduce the temporal prediction of traffic flow intervals, allowing users to decide when to go out to avoid traffic waves.

Figure 13 is a schematic diagram showing the road risk prediction results displayed in the web browser by the personalized road risk prediction system of the present invention using a web browser as a front-end program; the personalized road risk prediction system 10 of the present invention is preferably based on software. Built using SaaS and PaaS cloud technologies, the browser running on the user device 100 of the front-end presentation layer can be used as the front-end program 109 to generate various system interfaces on the browser, such as But it is not limited to the interfaces shown in Figures 9 to 12 that are provided for the user to operate. The user starts the browser on his/her user device 100 in an Internet-enabled environment and enters the correct address in the address bar. The Uniform Resource Locator (URL) can be used to log in to the system.

Figure 14 is a schematic diagram showing the road risk prediction results displayed in the vehicle information communication application program of the personalized road risk prediction system of the present invention using the vehicle information communication special application program as the front-end program; the personalized road risk prediction system 10 of the present invention Preferably, a dedicated application program executed on the vehicle information communication device 110 of the front-end presentation layer can be used as the front-end program 109, and various system interfaces are generated on the dedicated application program, such as but not limited to those disclosed in Figures 9 to 12 The interface is provided for users to operate in the car. In a network-enabled environment, the user can log in to the system by launching a dedicated application on his or her own vehicle telematics device 110 .

Figure 15 shows a flow chart of implementation steps of the personalized road risk prediction method; in summary, the personalized road risk prediction method 500 of the present invention preferably includes the following steps: at the back end Install the traffic risk prediction master computer program product on the system server (step 501); execute the front-end computer program product on the user device for the user to operate to drive the trained traffic risk prediction master computer program product included in the traffic risk prediction master computer program product. Traffic risk prediction machine learning model program component (step 503); after execution, the traffic risk prediction machine learning model program component implements the prediction of the risk map and selectively calculates the personalized risk index, and calculates at least one personalized recommended route accordingly (step 503) 505); and displaying the at least one personalized recommended path to the user through the user device (step 507).

In summary, through machine learning algorithms, from past historical data and combined with the user's current environment, the risk of the path provided by the user's navigation is predicted, and based on the user's relevant data, such as the vehicle used, the user's The system can also perform additional analysis on the collected information, such as exploring the causes of accident-prone locations, and then make improvements to target areas to enhance Safety for all passers-by.

The above embodiments of the present invention can be arbitrarily combined or replaced with each other, thereby deriving more embodiments, without departing from the scope of protection of the present invention. More embodiments of the present invention are further provided as follows:

Embodiment 1: A personalized road risk prediction system, which includes: a system server, which includes a traffic risk prediction main control computer program product; and a user device, which is communicated with the system server and provides a front-end computer program The product is for users to operate to drive the trained traffic risk prediction machine learning model program component included in the traffic risk prediction master computer program product. After execution, the traffic risk prediction machine learning model program component implements the prediction and selection of the risk map. Calculate a personalized risk index based on this and provide the user with at least A personalized recommended path.

Embodiment 2: The personalized road risk prediction system as described in Embodiment 1, wherein the traffic risk prediction machine learning model program component also includes an image space feature extraction model and a time series dependency learning model, wherein the image space feature extraction model The system is selected from the convolutional neural network model, regional convolutional neural network model, fast regional convolutional neural network model, deep convolutional neural network model, recurrent neural network model, convolutional recurrent neural network model, deep One of the recurrent neural network model, fully convolutional neural network model, multi-column convolutional neural network model, bidirectional neural network model and deep neural network model.

Embodiment 3: The personalized road risk prediction system as described in Embodiment 2, wherein the time series dependency learning model is selected from a long short-term memory model, an unsupervised learning model, a supervised learning model, a deep learning model, and an automatic learning model. One of the encoder model, neural network-like model, multi-layer perception model, deep neural network model and ensemble learning model.

Embodiment 4: The personalized road risk prediction system as described in Embodiment 1, wherein the user equipment is selected from the group consisting of in-vehicle telematics devices, mobile devices, desktop computers, notebook computers, smart phones and tablet devices. one.

Embodiment 5: The personalized road risk prediction system as described in Embodiment 1, wherein the front-end computer program product is selected from one of an application program, a micro-application program, a vehicle telematics dedicated application program, and a web browser.

Embodiment 6: The personalized road risk prediction system as described in Embodiment 1, wherein the personalized risk index is associated with user age weight, user age, user gender weight, user gender, user vehicle type weight, One of the user's vehicle type, road usage time weight, road usage time, climate condition weight, climate conditions and other personalized information.

Embodiment 7: A personalized road risk prediction method, which includes: installing a traffic risk prediction main control computer program product on a back-end system server; executing a front-end computer program product on a user device for the user to operate to drive the The traffic risk prediction master computer program product contains a trained traffic risk prediction machine learning model program component; after execution, the traffic risk prediction machine learning model program component implements the prediction of the risk map and selectively calculates the personalized risk index, and based on Calculate at least one personalized recommended path; and display the at least one personalized recommended path to the user through the user device.

Embodiment 8: The personalized road risk prediction method as described in Embodiment 7 further includes one of the following steps: causing the system server to access a traffic risk database; selecting a prediction range and gridding the prediction range; Calculate at least one recommended route; calculate the estimated arrival time of the at least one recommended route; calculate the risk value of the at least one recommended route based on the risk map and the personalized risk index, and display the risk to the user through the user device value; and adjust the estimated arrival time based on the risk value, and display the adjusted estimated arrival time to the user through the user device.

Embodiment 9: A personalized road risk prediction user equipment, which is communicated with a system server including a traffic risk prediction main control computer program product. The personalized road risk prediction user equipment includes: a display unit; and a process A server unit that executes a front-end computer program product for user operation to drive a trained traffic risk prediction machine learning model program component included in the traffic risk prediction master computer program product. The traffic risk prediction machine learning model program component is executed Then, the prediction of the risk map and the selective calculation of the personalized risk index are implemented, and at least a personalized recommended path is generated based on this, and displayed to the user through the display unit.

Embodiment 10: The personalized road risk prediction user equipment as described in Embodiment 9, wherein the display unit is selected from one of a display, an external display and a touch screen.

The various embodiments of the present invention can be arbitrarily combined or replaced with each other to derive more embodiments, but all do not deviate from the scope of protection of the present invention. The scope of protection of the present invention is defined by the patent scope of the present invention. The recorder shall prevail.

500: Personalized road risk prediction method of the present invention

501~507: Implementation steps

Claims (8)

A personalized road risk prediction system, which includes: a system server, which includes a traffic risk prediction master computer program product; and a user device, which is communicated with the system server and provides a front-end computer program The product is operated by a user to drive a trained traffic risk prediction machine learning model program component included in the traffic risk prediction master computer program product. The traffic risk prediction machine learning model program component is based on a two-dimensional space grid traffic The relevant event data is used to train an image space feature extraction model and a time series dependency fitting learning model. The image space feature extraction model is a CNN model and the time series dependency fitting learning model is an LSTM model. The CNN model is connected to the LSTM model through a transposition layer, a reshaping layer and a flattening layer. After execution, the traffic risk prediction main control computer program product generates one of a traffic risk spatial prediction and a traffic risk time series prediction. Predicting results, and calculating a personalized risk index and adjusting the prediction results according to the personalized risk index, and providing at least one personalized recommended path to the user accordingly, wherein the evaluation of the personalized risk index is based on a user's age, A user's gender, a user's car type, a driving time, a driving experience, and one of their combinations. The personalized road risk prediction system as described in claim 1, wherein the user device is selected from the group consisting of an in-vehicle telematics device, a mobile device, a desktop computer, a notebook computer, a smart phone and a tablet device one of them. The personalized road risk prediction system as described in request 1, wherein the front-end computer program product is selected from an application, a micro application, a vehicle information communication dedicated application and a web page One of the browsers. The personalized road risk prediction system as described in claim 1, wherein the personalized risk index is associated with a user age weight, a user gender weight, a user vehicle type weight, a road usage time weight and an other individual Information is one of them. A personalized road risk prediction method, which includes: training an image space feature extraction model based on two-dimensional space grid traffic-related event data and a time series dependency fitting learning model, wherein the image space feature extraction model is a CNN model And the time series dependency fitting learning model is an LSTM model. The CNN model is connected to the LSTM model through a transposition layer, a reshaping layer and a flattening layer; a back-end system server is installed that contains the traffic One of the risk prediction machine learning model program components is a traffic risk prediction master computer program product; executing a front-end computer program product on a user device for operation by a user to drive the execution of the traffic risk prediction master computer program product; After execution, the traffic risk prediction main control computer program product generates a prediction result of a traffic risk spatial prediction and a traffic risk time series prediction, and calculates a personalized risk index and adjusts the prediction result according to the personalized risk index. This calculates at least one personalized recommended path, wherein the evaluation of the personalized risk index is based on one of a user's age, a user's gender, a user's car type, a driving time, a driving experience, and a combination thereof; and by The user device displays the at least one personalized recommended path to the user. The personalized road risk prediction method described in claim 5 further includes one of the following steps: causing the system server to access a traffic risk database; selecting a prediction range and gridding the prediction range; calculating at least one Suggesting a route; calculating an estimated arrival time of the at least one suggested route; calculating a risk value of the at least one suggested route based on the prediction result and the personalized risk index, and displaying the risk value to the user through the user device ; and adjust the estimated arrival time based on the risk value, and display the adjusted estimated arrival time to the user through the user device. A kind of personalized road risk prediction user equipment, which is communicated with a system server including a traffic risk prediction main control computer program product. The personalized road risk prediction user equipment includes: a display unit; and a processor A unit that executes a front-end computer program product for operation by a user to drive a trained traffic risk prediction machine learning model program component included in the traffic risk prediction main control computer program product, and the traffic risk prediction machine learning model program component It is built based on the two-dimensional space grid traffic-related event data to train an image spatial feature extraction model and a time series dependency fitting learning model, in which the image spatial feature extraction model is a CNN model and the time series dependency fitting learning model. The combined learning model is an LSTM model. The CNN model is connected to the LSTM model through a transposition layer, a reshaping layer and a flattening layer. After execution, the traffic risk prediction main control computer program product generates a traffic risk spatial prediction and A prediction result of a traffic risk time series prediction, and calculating a personalized risk index and adjusting the prediction result according to the personalized risk index, and generating at least one personalized risk index accordingly. The recommended route is displayed to the user through the display unit, wherein the evaluation of the personalized risk index is based on a user's age, a user's gender, a user's vehicle type, a road usage time, a driving experience and a combination thereof one of them. The personalized road risk prediction user device as described in claim 7, wherein the display unit is selected from one of a display, an external display and a touch screen.

Images