TW201913528A - Method and processing device for driving risk assessment based on multiple kernel learning - Google Patents
Method and processing device for driving risk assessment based on multiple kernel learning Download PDFInfo
- Publication number
- TW201913528A TW201913528A TW106129882A TW106129882A TW201913528A TW 201913528 A TW201913528 A TW 201913528A TW 106129882 A TW106129882 A TW 106129882A TW 106129882 A TW106129882 A TW 106129882A TW 201913528 A TW201913528 A TW 201913528A
- Authority
- TW
- Taiwan
- Prior art keywords
- driving
- risk assessment
- risk
- driving risk
- classifier
- Prior art date
Links
Landscapes
- Traffic Control Systems (AREA)
Abstract
Description
本發明是有關於一種駕駛風險評估方法及其裝置，且特別是有關於一種基於多核學習的駕駛風險評估方法及其裝置。The invention relates to a driving risk assessment method and a device thereof, and in particular to a driving risk assessment method based on multi-core learning and a device thereof.
隨著資訊技術的演進以及車聯網多元應用的發展，基於大量駕駛資料所分析出的UBI車險（usage based insurance，UBI） 將逐漸地取代傳統的車險，以使得駕駛風險的估算結果更貼近真實的用車狀態。因此，如何有效地利用大量且多樣性的駕駛資料以有效並且精確地評估駕駛風險等級已成為重要的議題之一。With the development of information technology and the development of multiple applications of the Internet of Vehicles, UBI auto insurance (UBI) based on a large amount of driving data will gradually replace the traditional auto insurance, so that the estimation of driving risk is closer to the real one. Car status. Therefore, how to effectively utilize a large and diverse driving data to effectively and accurately evaluate driving risk levels has become one of the important topics.
有鑑於此，本發明提供一種基於多核學習的駕駛風險評估方法及其裝置，其可利用駕駛風險評估模型針對駕駛的過程所感測到大量且多樣性的駕駛資料有效並且精確地評估各個駕駛者的駕駛風險等級。In view of this, the present invention provides a multi-core learning-based driving risk assessment method and apparatus thereof, which can utilize the driving risk assessment model to sense a large and diverse driving data for a driving process and effectively and accurately evaluate each driver's Driving risk level.
在本發明的一實施例中，上述的駕駛風險評估方法適用於連接於多個感測器的處理裝置，並且包括下列步驟。首先，自感測器接收所感測到的駕駛資料，其中感測器至少安裝於運輸載具以及運輸載具的駕駛人。接著，將駕駛資料輸入至駕駛風險評估模型，其中駕駛風險評估模型包括多核產生模組、核學習模組以及核最佳化模組，其中多核產生模組利用對應於多個核的多個弱分類器，針對各筆駕駛資料進行安全等級標記以及風險群組分類，以判斷各筆駕駛資料分別對應於各個弱分類器的機率密度函數值，據以取得各筆駕駛資料相對於各個弱分類器的主導駕駛風險標記，並且利用強分類器根據所有主導駕駛風險標記輸出駕駛風險評估結果，其中核學習模組於駕駛風險評估模型的訓練階段時利用自適應性增強（AdaBoost）架構訓練弱分類器以及強分類器，其中核最佳化模組利用動盪粒子群最佳化（TPSO）演算法訓練各個弱分類器的參數。之後，取得由駕駛風險評估模型所輸出的駕駛風險評估結果。In an embodiment of the invention, the above-described driving risk assessment method is applicable to a processing device connected to a plurality of sensors, and includes the following steps. First, the self-sensing device receives the sensed driving information, wherein the sensor is installed at least on the transport vehicle and the driver of the transport vehicle. Then, driving data is input to a driving risk assessment model, wherein the driving risk assessment model includes a multi-core generation module, a nuclear learning module, and a nuclear optimization module, wherein the multi-core generation module utilizes a plurality of weak corresponding to the plurality of cores The classifier performs safety level marking and risk group classification for each driving data to determine that each driving data corresponds to a probability density function value of each weak classifier, thereby obtaining each driving data relative to each weak classifier Leading the driving risk marker, and using a strong classifier to output driving risk assessment results based on all leading driving risk markers, wherein the nuclear learning module trains the weak classifier using the adaptive enhancement (AdaBoost) architecture during the training phase of the driving risk assessment model And a strong classifier, wherein the kernel optimization module uses the turbulent particle swarm optimization (TPSO) algorithm to train the parameters of each weak classifier. After that, the driving risk assessment result output by the driving risk assessment model is obtained.
在本發明的一實施例中，上述的處理裝置連接於多個感測器，並且包括記憶體以及處理器，其中感測器至少安裝於運輸載具以及穿戴於運輸載具的駕駛人。記憶體用以儲存資料。處理器耦接記憶體，用以自感測器接收所感測到的駕駛資料，將駕駛資料輸入至駕駛風險評估模型，以及取得由駕駛風險評估模型所輸出的駕駛風險評估結果，其中駕駛風險評估模型包括多核產生模組、核學習模組以及核最佳化模組，其中多核產生模組利用對應於多個核的多個弱分類器，針對各筆駕駛資料進行安全等級標記以及風險群組分類，以判斷各筆駕駛資料分別對應於各個弱分類器的機率密度函數值，據以取得各筆駕駛資料相對於各個弱分類器的主導駕駛風險標記，並且利用強分類器根據所有主導駕駛風險標記輸出駕駛風險評估結果，其中核學習模組於駕駛風險評估模型的訓練階段時利用自適應性增強（AdaBoost）架構訓練弱分類器以及強分類器，其中核最佳化模組利用動盪粒子群最佳化（TPSO）演算法訓練各個弱分類器的參數。In an embodiment of the invention, the processing device is coupled to a plurality of sensors and includes a memory and a processor, wherein the sensor is mounted on at least the transport carrier and a driver worn on the transport vehicle. Memory is used to store data. The processor is coupled to the memory for receiving the sensed driving data from the sensor, inputting the driving data to the driving risk assessment model, and obtaining the driving risk assessment result output by the driving risk assessment model, wherein the driving risk assessment The model includes a multi-core generation module, a nuclear learning module, and a nuclear optimization module, wherein the multi-core generation module uses a plurality of weak classifiers corresponding to the plurality of cores to perform security level marking and risk groups for each driving data. Classification, to determine that each driving data corresponds to a probability density function value of each weak classifier, respectively, to obtain a leading driving risk flag of each driving data with respect to each weak classifier, and use a strong classifier according to all leading driving risks The flag outputs the driving risk assessment result, wherein the nuclear learning module uses the adaptive enhancement (AdaBoost) architecture to train the weak classifier and the strong classifier in the training phase of the driving risk assessment model, wherein the nuclear optimization module utilizes the turbulent particle group The Optimum (TPSO) algorithm trains the parameters of each weak classifier.
為讓本發明的上述特徵和優點能更明顯易懂，下文特舉實施例，並配合所附圖式作詳細說明如下。The above described features and advantages of the invention will be apparent from the following description.
圖1是根據本發明一實施例所繪示的處理裝置的方塊圖，但此僅是為了方便說明，並不用以限制本發明。首先圖1先介紹處理裝置之所有構件以及配置關係，詳細功能將配合圖2一併揭露。1 is a block diagram of a processing apparatus according to an embodiment of the invention, but is for convenience of description and is not intended to limit the present invention. First, all components and configuration relationships of the processing device will be described first in FIG. 1. The detailed functions will be disclosed in conjunction with FIG.
請參照圖1，處理裝置100至少包括記憶體110以及處理器120，其中處理器120耦接於記憶體110。在本實施例中，處理裝置100可以是個人電腦、筆記型電腦、應用程式伺服器、雲端伺服器、資料庫伺服器、工作站等具有運算能力的電腦系統，其可透過通訊模組（未繪示）經由網路取得運輸載具的駕駛資料。在另一實施例中，處理裝置100可以是智慧型手機、平板電腦、車用電腦（例如安裝於擋風玻璃上）等位於運輸載具的電腦系統，以即時地針對運輸載具的駕駛風險進行評估。在此的運輸載具可以是汽車、公車、貨櫃車、電動車、機車等。Referring to FIG. 1 , the processing device 100 includes at least a memory 110 and a processor 120 , wherein the processor 120 is coupled to the memory 110 . In this embodiment, the processing device 100 may be a computer system with computing power, such as a personal computer, a notebook computer, an application server, a cloud server, a database server, a workstation, etc., which can communicate through a communication module (not drawn Show) obtaining driving information of the transport vehicle via the network. In another embodiment, the processing device 100 may be a computer system of a transportation vehicle such as a smart phone, a tablet computer, or a vehicle computer (for example, mounted on a windshield) to instantly drive the driving risk of the transportation vehicle. to evaluate. The transport vehicles here may be automobiles, buses, container trucks, electric vehicles, locomotives, and the like.
處理裝置100的記憶體110用以儲存資料，其可以例如是任意型式的固定式或可移動式隨機存取記憶體（random access memory，RAM）、唯讀記憶體（read-only memory，ROM）、快閃記憶體（flash memory）、硬碟或其他類似裝置、積體電路及其組合。The memory 110 of the processing device 100 is configured to store data, which may be, for example, any type of fixed or removable random access memory (RAM), read-only memory (ROM). , flash memory, hard disk or other similar device, integrated circuit and combinations thereof.
處理裝置100的處理器120用以執行所提出的駕駛風險評估方法，其可以例如是中央處理單元（central processing unit，CPU），或是其他可程式化之一般用途或特殊用途的微處理器（microprocessor）、數位訊號處理器（digital signal processor，DSP）、可程式化控制器、特殊應用積體電路（application specific integrated circuits，ASIC）、可程式化邏輯裝置（programmable logic device，PLD）或其他類似裝置、晶片、積體電路及其組合。The processor 120 of the processing device 100 is configured to execute the proposed driving risk assessment method, which may be, for example, a central processing unit (CPU), or other programmable general purpose or special purpose microprocessor ( Microprocessor, digital signal processor (DSP), programmable controller, application specific integrated circuit (ASIC), programmable logic device (PLD) or the like Devices, wafers, integrated circuits, and combinations thereof.
圖2是根據本發明之一實施例所繪示的駕駛風險評估方法的流程圖，而圖2的方法流程可以圖1的處理裝置100的各元件實現。2 is a flow chart of a driving risk assessment method according to an embodiment of the present invention, and the method flow of FIG. 2 may be implemented by various components of the processing apparatus 100 of FIG.
請同時參照圖1以及圖2，首先，處理裝置100的處理器120將自感測器接收所感測到的駕駛資料（步驟S202）。在此的感測器可以例如是車上診斷系統（on-board diagnostics，OBD）、內部量測單元（internal measurement unit，IMU）、相機等安裝於運輸載具內部或是外部以進行監控的感測裝置，駕駛者身上的穿戴式裝置、電子貼片，或者是其它來源用以偵測運輸載具周圍環境的感測器。在此的感測資料則可以例如是運輸載具的情況（車道坡度、車道偏離、車距）、駕駛者的情況（呼吸率、注意力、心率、情緒）、車道的情況（速限、道路類型、交通量、氣象）。此外，在本實施例中，處理器120可以例如是透過通訊模組（未繪示）定時地接收來自運輸載具的感測資料以及來自其它來源與運輸載具相關的開放資料（open data），以針對運輸載具進行即時且最新的風險評估。Referring to FIG. 1 and FIG. 2 simultaneously, first, the processor 120 of the processing device 100 receives the sensed driving data from the sensor (step S202). The sensor here may be, for example, an on-board diagnostics (OBD), an internal measurement unit (IMU), a camera, etc., mounted inside or outside the transport vehicle for monitoring. A measuring device, a wearable device on the driver's body, an electronic patch, or other source for detecting the environment surrounding the carrier. The sensing data here can be, for example, the case of transporting the vehicle (lane gradient, lane departure, distance), the driver's condition (respiratory rate, attention, heart rate, mood), the condition of the lane (speed limit, road) Type, traffic volume, weather). In addition, in this embodiment, the processor 120 may receive the sensing data from the transportation vehicle and the open data related to the transportation vehicle from other sources periodically, for example, through a communication module (not shown). To provide an immediate and up-to-date risk assessment for transport vehicles.
接著，處理器120將駕駛資料輸入至駕駛風險評估模型（步驟S204），以取得由駕駛風險評估模型所輸出的駕駛風險評估結果（步驟S206）。詳細來說，駕駛風險評估模型是基於多核學習所提出的一個架構，以取得各個不同駕駛者的駕駛風險等級。基於大量且多樣性的駕駛資料，以下將考慮兩個要素來進行駕駛風險評估。首先，為了解決大量的駕駛資料導致分析上的困難，在此將採用由多核針對大量的駕駛資料所產生的數個機率模型來取得統計資訊以及用以勘探所有駕駛資料的代表資訊。接著，為了解決多樣性的駕駛資料導致難以歸納出特徵，在此將採用自適應性增強（AdaBoost）架構來整合前述機率模型以取得各個駕駛者更為可靠且實用的風險評估。基此，駕駛風險評估架構將包括三個模組：多核產生模組、核學習模組以及核最佳化模組，以下將分敘說明。Next, the processor 120 inputs the driving data to the driving risk assessment model (step S204) to obtain the driving risk assessment result output by the driving risk assessment model (step S206). In detail, the driving risk assessment model is based on a framework proposed by multi-core learning to obtain driving risk levels for different drivers. Based on a large and diverse range of driving data, two factors will be considered below for driving risk assessment. First of all, in order to solve the analysis difficulties caused by a large amount of driving data, several probability models generated by multi-core for a large amount of driving data are used to obtain statistical information and representative information for exploring all driving data. Then, in order to solve the diversity of driving data, it is difficult to generalize the features. Here, the adaptive enhancement (AdaBoost) architecture will be used to integrate the aforementioned probability models to obtain a more reliable and practical risk assessment for each driver. Based on this, the driving risk assessment framework will consist of three modules: a multi-core generation module, a nuclear learning module, and a nuclear optimization module, which will be described below.
在多核產生模組中，將利用多核來產生多個機率模型，其中每個機率模型將視為弱學習者（weak learner）來針對各筆駕駛資料進行駕駛風險等級的標記，以有效地分析大量資料。In the multi-core generation module, multiple cores will be used to generate multiple probability models, each of which will be regarded as a weak learner to mark the driving risk level for each driving data, so as to effectively analyze a large number. data.
具體來說，機率模型是基於特定的核以及駕駛資料集來從核所產生的駕駛風險標記集中選擇主要的駕駛風險標記。以圖3根據本發明一實施例所繪示的弱學習者的建構方法的功能流程圖為例，在弱學習者的建構300中，為包括駕駛資料的駕駛資料集，其中為駕駛資料序列302中第個時刻的駕駛資料。假設存在一個「車速」核，則在經過安全等級的分配304後，速度安全等級陣列可以表示為，其中為根據「車速」核所得到第個時刻的駕駛資料的安全等級。接著，藉由標記為特定等級的駕駛資料來進行駕駛風險的分類306，則將可產生駕駛風險群組。在此，駕駛資料可以是標記為以下三種等級：高風險、中間風險以及低風險，因此、、將分別代表高風險、中間風險以及低風險的駕駛風險群組。Specifically, the probability model is based on a specific core. Driving data set Come from the core The resulting driving risk markers focus on selecting the main driving risk markers. FIG. 3 is a functional flow chart of a weak learner construction method according to an embodiment of the present invention. In the construction of a weak learner 300, Including driving information Driving data set, where For the driving data sequence 302 Driving information for a moment. Assuming that there is a "vehicle speed" core, after the security level assignment 304, the speed security level array can be expressed as ,among them In order to obtain the first according to the "speed" The safety level of driving information at a time. Next, by classifying 306 the driving risk by marking the driving data of a particular level, a driving risk group will be generated. Here, the driving information can be marked as the following three levels: high risk, intermediate risk and low risk, so , , It will represent high risk, intermediate risk and low risk driving risk groups.
此外，以下表1是根據本發明一實施例所繪示各個安全等級範圍以及各個風險群組的組成。低風險群組包含安全等級標記為1到4的駕駛資料，中間風險群組包含安全等級標記為3到4的駕駛資料，高風險群組包含安全等級標記為3到6的駕駛資料。
接著，請回到圖3，在進行完駕駛風險的分類306後，將根據速度安全等級陣列以及駕駛風險群組、、來判斷各個駕駛風險的機率密度函數308。具體來說，圖4A是根據本發明一實施例所繪示車速的分布示意圖，而圖4B是根據本發明一實施例所繪示機率密度分布示意圖。Next, please return to Figure 3, after completing the classification 306 of driving risks, according to the speed security level array Driving risk group , , To determine the probability density function 308 for each driving risk. Specifically, FIG. 4A is a schematic diagram showing a distribution of vehicle speed according to an embodiment of the invention, and FIG. 4B is a schematic diagram showing a probability density distribution according to an embodiment of the invention.
請參照圖4A，第一輪RD1的車速具有高風險標記，第二輪RD1的車速具有中間風險標記，第三輪RD3的車速具有低風險標記。為了取得各個群組的駕駛資料的機率，上述分布是用以定義三個機率密度函數（probability density function），其可以方程式(1)來表示：其中為可使得高斯曲線更為平緩或是陡峭的平滑參數，其可以例如是設定為0.2。Referring to FIG. 4A, the vehicle speed of the first wheel RD1 has a high risk flag, the vehicle speed of the second wheel RD1 has an intermediate risk flag, and the vehicle speed of the third wheel RD3 has a low risk flag. In order to obtain driving information for each group The probability that the above distribution is used to define three probability density functions, which can be expressed by equation (1): among them A smoothing parameter that can make the Gaussian curve more gradual or steep, which can be set, for example, to 0.2.
請再參照圖4B，曲線C1為對應於高風險的機率密度分布，曲線C2為對應於中間風險的機率密度分布，曲線C3為對應於低風險的機率密度分布。Referring again to FIG. 4B, curve C1 is a probability density distribution corresponding to high risk. Curve C2 is the probability density distribution corresponding to the intermediate risk Curve C3 is the probability density distribution corresponding to low risk .
請再回到圖3，在取得機率密度函數306後，可根據、以及，利用方程式(2)來判斷駕駛資料的駕駛風險標記： Please return to Figure 3, after obtaining the probability density function 306, , as well as , using equation (2) to judge driving data Driving risk marker :
最後，將自駕駛風險標記集中來進行主導駕駛風險標記的選擇310，以取得駕駛資料集中的主導駕駛風險標記。在本實施例中，將採用表2所示的20個核來標記各筆駕駛資料的駕駛風險。
基於核的多樣性，在此可根據方程式(1)來將各個核具有的資訊轉換成機率模型，以做為弱學習者。接著，將根據強分類器合併所有的弱學習者（亦即，所有核所產生的機率模型的聯集），以在取得各個駕駛者更為可靠且實用的風險評估。也就是說，在取得弱學習者後，將利用AdaBoost架構來線性地組合自不同核所產生的弱學習者來針對多樣化的架駛資料進行駕駛風險等級的標記。以下將以圖5根據本發明一實施例所繪示的多核學習架構的示意圖為例，說明核學習模組以AdaBoost架構將藉由持續地分配權重至訓練資料，以迭代地選擇判別弱學習者（discriminant weak learners），建立強分類器的詳細步驟。Based on the diversity of the kernel, the information possessed by each core can be converted into a probability model according to equation (1) as a weak learner. Next, all weak learners (ie, a union of probability models generated by all cores) will be merged according to a strong classifier to obtain a more reliable and practical risk assessment for each driver. That is to say, after obtaining the weak learner, the AdaBoost architecture will be used to linearly combine the weak learners generated by the different cores to mark the driving risk level for the diverse driving data. FIG. 5 is a schematic diagram of a multi-core learning architecture according to an embodiment of the present invention. The core learning module uses the AdaBoost architecture to continuously assign weights to training data to iteratively select and identify weak learners. (discriminant weak learners), the detailed steps to build a strong classifier.
假設給定訓練駕駛資料集其及所對應的風險標記，其中。為了建構多核學習架構以取得自多樣性的駕駛資料取得所需的特徵，在進行每次迭代時，將會尋找兩種最佳權重，即各個訓練樣本的權重以及各個弱分類器的權重。在此，用以訓練的駕駛資料510可以是來自智慧型手環501、心率監控器502、相機503、車上診斷系統504、內部量測單元505、開放資料506。最終，將根據弱分類器與權重進行加權計算Σ，以建構出強分類器。Assume a given training driving data set And the corresponding risk markers ,among them . In order to construct a multi-core learning architecture to obtain the features required for the diversity of driving data, each iteration will look for two optimal weights, ie the weights of the individual training samples. And each weak classifier the weight of . Here, the driving data 510 for training may be from the smart bracelet 501, the heart rate monitor 502, the camera 503, the onboard diagnostic system 504, the internal measuring unit 505, and the open material 506. Finally, the weight classifier will be weighted according to the weak classifier and weight to construct a strong classifier. .
具體而言，在此多核學習架構中，訓練樣本的權重將初始化為所有樣本的平均值，如方程式(3)所示：其中為訊練駕駛資料集中訓練樣本的數量。此外，第次迭代中的分類錯誤率（error rate of misclassification）可以表示為方程式(4)：其中為訓練樣本在第次迭代中的權重。必須說明的是，當標記成1時，代表的條件為是，而當標記成0時，代表分類器在第次迭代時已達到訓練樣本正確的分類。Specifically, in this multi-core learning architecture, training samples The weight of the weight will be initialized to the average of all samples, as shown in equation (3): among them Train driving data set The number of training samples in the middle. In addition, the first The error rate of misclassification in the second iteration can be expressed as equation (4): among them Training sample In the first The weight in the iteration. It must be stated that when When marked as 1, it represents The condition is yes, and when When marked as 0, it represents the classifier In the first Training samples have been reached at the next iteration Correct classification.
接著，分類器在第次迭代時的權重可以表示成方程式(5)：其中為風險等級的數量，而本實施例中是設定為3。在第次迭代時，訓練樣本的權重可以是根據方程式(6)來分配： Next, the classifier In the first Weight at the next iteration Can be expressed as equation (5): among them It is the number of risk levels, and is set to 3 in this embodiment. In the first Training samples at the next iteration the weight of It can be assigned according to equation (6):
標記成1時，即代表分類器正確地分類訓練樣本，否則，標記成0。在調整訓練樣本的權重之後，將會正規化（normalize）至0到1之間，如方程式(7)所示： When marked as 1, it represents the classifier Correctly classify training samples ,otherwise, Marked as 0. Adjusting training samples the weight of after that, Will normalize to between 0 and 1, as shown in equation (7):
最後，具有多核的AdaBoost架構可以是經由上述兩個最佳化權重而做為強分類器，其可表示成方程式(8)：其中標記成1時，即代表弱分類器在第次迭代時將訓練樣本分類至類別。假設駕駛風險標記包含低風險、中間風險以及高風險（即，）。在計算各個風險標記的分類權重後，具有最高分類權重的風險標記將被視為最終分類結果。因此，多核學習架構可確保可從多樣性的駕駛資料中最佳化地整合出所需的特徵，並且線性的強分類器可有效地針對大量且多樣性的駕駛資料有效地進行駕駛風險等級的標記。Finally, the AdaBoost architecture with multiple cores can be used as a strong classifier via the above two optimization weights. , which can be expressed as equation (8): among them When marked as 1, it means weak classifier In the first Train samples at the next iteration Classification to category . Assume that the driving risk marker contains low risk, intermediate risk and high risk (ie, ). After calculating the classification weights for each risk marker, the risk marker with the highest classification weight will be considered the final classification result. Therefore, the multi-core learning architecture ensures that the required features can be optimally integrated from a diverse range of driving data, and a linear strong classifier can effectively drive driving risk levels for a large and diverse range of driving data. mark.
必須說明的是，基於AdaBoost架構中的弱學習者的能力對於整體的分類效能極為重要，為了提升各個弱學習者的效能，在此可以2013年2月由Chou等人於IEEE所發表的論文「Turbulent-PSO-Based Fuzzy Image Filter with No-Reference Measures for High-Density Impulse Noise」中所提出的動盪粒子群最佳化演算法（turbulent particle swarm optimization，TPSO）來針對各個核進行最佳化處理。It must be noted that the ability of weak learners based on the AdaBoost architecture is extremely important for the overall classification performance. In order to improve the effectiveness of each weak learner, the paper published by Chou et al. in IEEE in February 2013 The turbulent particle swarm optimization (TPSO) proposed in Turbulent-PSO-Based Fuzzy Image Filter with No-Reference Measures for High-Density Impulse Noise is optimized for each core.
在核最佳化模組中，將利用TPSO演算法來取得各個核的最佳參數（即，各個駕駛風險群組的等級範圍）。TPSO演算法事實上為PSO演算法的延申，其可根據已知的品質衡量標準（quality metric）來迭代地修正候選解。一開始，群體包括個粒子（每個粒子的標號為，其中），各個粒子具有個元素（其中）。在第次迭代時，第個粒子的第個元素的速度可以方程式(9)表示：並且位置可以方程式(10)表示： In the core optimization module, the TPSO algorithm will be used to obtain the optimal parameters of each core (ie, the level range of each driving risk group). The TPSO algorithm is in fact an extension of the PSO algorithm, which iteratively corrects candidate solutions based on known quality metrics. In the beginning, the group included Particles (each particle is labeled ,among them ), each particle has Elements (where ). In the first On the iteration, the first Number of particles Speed of elements It can be expressed by equation (9): And location It can be expressed by equation (10):
為了解決一般PSO演算法常發生的非成熟收斂（premature convergence），本實施例所採用的TPSO演算法可驅動惰性粒子以避免達到局部最佳化。基此，可以方程式(11)來約束速度：其中以及分別為最大速度閥值以及最小速度閥值。In order to solve the premature convergence that often occurs in general PSO algorithms, the TPSO algorithm used in this embodiment can drive inert particles to avoid local optimization. Based on this, equation (11) can be used to constrain the speed. : among them as well as The maximum speed threshold and the minimum speed threshold are respectively.
綜上所述，本發明所提出基於多核學習的駕駛風險評估方法及其處理裝置，其可利用駕駛風險評估模型針對駕駛的過程所感測到大量且多樣性的駕駛資料精確地評估各個駕駛者的駕駛風險等級，以有效地應用於UBI車險技術服務。In summary, the present invention proposes a multi-core learning-based driving risk assessment method and a processing apparatus thereof, which can utilize the driving risk assessment model to accurately estimate the drivers of a large amount and diversity of driving data for the driving process. Driving risk levels for effective application in UBI car insurance technology services.
雖然本發明已以實施例揭露如上，然其並非用以限定本發明，任何所屬技術領域中具有通常知識者，在不脫離本發明的精神和範圍內，當可作些許的更動與潤飾，故本發明的保護範圍當視後附的申請專利範圍所界定者為準。Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and any one of ordinary skill in the art can make some changes and refinements without departing from the spirit and scope of the present invention. The scope of the invention is defined by the scope of the appended claims.
100‧‧‧處理裝置100‧‧‧Processing device
110‧‧‧記憶體110‧‧‧ memory
120‧‧‧處理器120‧‧‧ processor
S202～S206‧‧‧步驟流程S202～S206‧‧‧Step procedure
D‧‧‧駕駛資料集D‧‧‧Driving data set
300‧‧‧弱學習者的建構300‧‧‧Construction of weak learners
302～310‧‧‧步驟流程302~310‧‧‧Step process
‧‧‧駕駛資料 ‧‧ Driving information
‧‧‧安全等級 ‧‧‧Security Level
、、‧‧‧駕駛風險群組 , , ‧‧‧Driving risk group
‧‧‧駕駛風險標記集 ‧‧‧ Driving Risk Marker
R‧‧‧主導駕駛風險標記 R‧‧‧Leading driving risk marker
RD1‧‧‧第一輪的車速RD1‧‧‧The first round of speed
RD2‧‧‧第二輪的車速RD2‧‧‧The second round of speed
RD3‧‧‧第三輪的車速RD3‧‧‧The third round of speed
C1、C2、C3‧‧‧曲線C1, C2, C3‧‧‧ curves
501‧‧‧智慧型手環501‧‧‧Smart bracelet
502‧‧‧心率監控器502‧‧‧ heart rate monitor
503‧‧‧相機503‧‧‧ camera
504‧‧‧車上診斷系統504‧‧‧Onboard diagnostic system
505‧‧‧內部量測單元505‧‧‧Internal measurement unit
506‧‧‧開放資料506‧‧‧Open materials
510‧‧‧駕駛資料510‧‧ Driving information
‧‧‧弱分類器 ‧‧‧Weak classifier
‧‧‧權重 ‧‧‧Weights
Σ‧‧‧加權計算 Σ‧‧‧weighted calculation
‧‧‧強分類器 ‧‧‧strong classifier
圖1是根據本發明一實施例所繪示的處理裝置的方塊圖。 圖2是根據本發明之一實施例所繪示的駕駛風險評估方法的流程圖。 圖3是根據本發明一實施例所繪示的弱學習者的建構方法的功能流程圖。 圖4A是根據本發明一實施例所繪示車速的分布示意圖。 圖4B是根據本發明一實施例所繪示機率密度分布示意圖。 圖5是根據本發明一實施例所繪示的多核學習架構的示意圖。FIG. 1 is a block diagram of a processing apparatus according to an embodiment of the invention. 2 is a flow chart of a driving risk assessment method according to an embodiment of the invention. FIG. 3 is a functional flowchart of a method for constructing a weak learner according to an embodiment of the invention. FIG. 4A is a schematic diagram showing the distribution of vehicle speed according to an embodiment of the invention. FIG. 4B is a schematic diagram showing probability density distribution according to an embodiment of the invention. FIG. 5 is a schematic diagram of a multi-core learning architecture according to an embodiment of the invention.
Claims (10)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW106129882A TWI646490B (en) | 2017-09-01 | 2017-09-01 | Method and processing device for driving risk assessment based on multiple kernal learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| TW106129882A TWI646490B (en) | 2017-09-01 | 2017-09-01 | Method and processing device for driving risk assessment based on multiple kernal learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| TWI646490B TWI646490B (en) | 2019-01-01 |
| TW201913528A true TW201913528A (en) | 2019-04-01 |
Family
ID=65804021
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| TW106129882A TWI646490B (en) | 2017-09-01 | 2017-09-01 | Method and processing device for driving risk assessment based on multiple kernal learning |
Country Status (1)
| Country | Link |
|---|---|
| TW (1) | TWI646490B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020202450A1 (en) * | 2019-04-01 | 2020-10-08 | ヤマハ発動機株式会社 | Personality analysis method, personality analysis device, information processing method using personality data, and information processing device using personality data |
| TWI758915B (en) * | 2020-10-22 | 2022-03-21 | 中華電信股份有限公司 | Driving risk analysis apparatus and method thereof |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20060053038A1 (en) * | 2004-09-08 | 2006-03-09 | Warren Gregory S | Calculation of driver score based on vehicle operation |
| US8140359B2 (en) * | 2008-09-11 | 2012-03-20 | F3M3 Companies, Inc, | System and method for determining an objective driver score |
| US20130166326A1 (en) * | 2011-12-21 | 2013-06-27 | Scope Technologies Holdings Limited | System and method for characterizing driver performance and use in determining insurance coverage |
| US20140278574A1 (en) * | 2013-03-14 | 2014-09-18 | Ernest W. BARBER | System and method for developing a driver safety rating |
| US20140379385A1 (en) * | 2013-06-24 | 2014-12-25 | Elwha, Llc | System and method for monitoring driving to determine an insurance property |
| TWM541410U (en) * | 2017-01-06 | 2017-05-11 | 創星物聯科技股份有限公司 | System for detecting severity of dangerous driving behavior |
-
2017
- 2017-09-01 TW TW106129882A patent/TWI646490B/en not_active IP Right Cessation
Also Published As
| Publication number | Publication date |
|---|---|
| TWI646490B (en) | 2019-01-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20210312567A1 (en) | Automobile Monitoring Systems and Methods for Loss Reserving and Financial Reporting | |
| CN109947086B (en) | Mechanical fault migration diagnosis method and system based on counterstudy | |
| US20210188290A1 (en) | Driving model training method, driver identification method, apparatuses, device and medium | |
| Malik et al. | Driving pattern profiling and classification using deep learning | |
| CN112270252A (en) | Multi-vehicle target identification method for improving YOLOv2 model | |
| WO2018121690A1 (en) | Object attribute detection method and device, neural network training method and device, and regional detection method and device | |
| EP3029606A2 (en) | Method and apparatus for image classification with joint feature adaptation and classifier learning | |
| JP6498107B2 (en) | Classification apparatus, method, and program | |
| Chen et al. | Driving safety risk prediction using cost-sensitive with nonnegativity-constrained autoencoders based on imbalanced naturalistic driving data | |
| CN111079640B (en) | Vehicle type identification method and system based on automatic amplification sample | |
| CN113378890B (en) | Lightweight pedestrian vehicle detection method based on improved YOLO v4 | |
| CN111950488B (en) | Improved Faster-RCNN remote sensing image target detection method | |
| WO2021049005A1 (en) | Information processing device and electronic apparatus equipped with same | |
| CN105760833A (en) | Face feature recognition method | |
| Khandelwal et al. | Segmentation-grounded scene graph generation | |
| TWI646490B (en) | Method and processing device for driving risk assessment based on multiple kernal learning | |
| US20210124982A1 (en) | Method for rating a state of a three-dimensional test object, and corresponding rating system | |
| CN113139594A (en) | Airborne image unmanned aerial vehicle target self-adaptive detection method | |
| US9177215B2 (en) | Sparse representation for dynamic sensor networks | |
| CN110852358A (en) | Vehicle type distinguishing method based on deep learning | |
| CN114154557A (en) | Cancer tissue classification method, apparatus, electronic device, and storage medium | |
| CN114067294B (en) | Text feature fusion-based fine-grained vehicle identification system and method | |
| CN105654174B (en) | System and method for prediction | |
| CN115439654A (en) | Method and system for finely dividing weakly supervised farmland plots under dynamic constraint | |
| CN107169532B (en) | Internet of vehicles fuel consumption data quality evaluation method based on wavelet analysis and semi-supervised learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| MM4A | Annulment or lapse of patent due to non-payment of fees |