EP4638225A1 - Driving profiling system - Google Patents

Abstract

The invention relates to a driving profiling system (100) comprising a data storage (104) containing vehicle movement data and processing circuitry (102) configured to perform the following steps: retrieving (302) the vehicle movement data from the data storage, extracting (304) trajectories from the vehicle movement data, each trajectories being associated to a driver or a vehicle, determining (306) linear segments of the trajectories, determining (308) clusters from the linear segments, determining (310) a dominant driving feature for each cluster; determining (312) values for at least the dominant driving feature for each cluster from the linear segments within the respective clusters, determining (314) driver behavior on a driving feature by analyzing distributions in each cluster.

Description

202201113 1 Description Driving profiling system Technical Field The invention relates to a driving profiling system for determining a driver behavior, a method for determining a driver behavior, computer program and a computer readable medium. Background Driver profiling is often applied to differentiate drivers’ behaviors, e.g., conservative/precautious, normal, and aggressive. Nonetheless, most existing driver profiling models are developed and tested using large-scale datasets comprising hundreds of drivers and multiple driving trajectories of each driver. Such models are not applicable at the cold-start stage wherein only a few drivers and a couple of driving trajectories of each driver are available. Methods used for driver profiling can be broadly categorized into two types: event-based and cluster-based. The event-based driver profiling methods detect aggressive events made by drivers and then profile the drivers based on how many such detected events are associated with them. The cluster-based methods group hundreds of drivers into several clusters and profile drivers by the characteristics of the clusters. Generally speaking, the cluster-based methods have a high-standard requirement on the quantity of the training data, such as hundreds of driving trajectories. However, a large number of driving trajectories are not publicly available for research purposes. The available data, however, usually involve several driving trajectories of a few, for example, 6 or 10, drivers. Summary of the invention There may be a desire to improve driver profiling. 202201113 2 The problem is solved by the subject-matter of the independent claims. Embodiments are provided by the dependent claims, the following description and the accompanying figures. The described embodiments similarly pertain to the driving profiling system for determining a driver behavior, the method for determining a driver behavior, the computer program and the computer readable medium. Synergetic effects may arise from different combinations of the embodiments although they might not be described in detail. Further on, it shall be noted that all embodiments of the present invention concerning a method might be carried out with the order of the steps as described, nevertheless this has not to be the only and essential order of the steps of the method. The herein presented methods can be carried out with another order of the disclosed steps without departing from the respective method embodiment, unless explicitly mentioned to the contrary hereinafter. Technical terms are used by their common sense. If a specific meaning is conveyed to certain terms, definitions of terms will be given in the following in the context of which the terms are used. According to a first aspect, a driving profiling system comprising a data storage containing vehicle movement data and processing circuitry configured to perform the following steps is provided: In a first step, the vehicle movement data are retrieved from the data storage. In a second step, trajectories are extracted from the vehicle movement data. In a third step, linear segments of the trajectory are determined. In a fourth step, clusters are determined from the linear segments. In a fifth step, a dominant driving feature is determined for each cluster. In a sixth step, values for at least the dominant driving feature for each cluster are determined from the linear segments within the respective clusters. Finally, a driver behavior on a driving feature is determined by analyzing distributions of the values of the dominant driving feature in each cluster. 202201113 3 In order to classify a profile of a driving style with respect to different categories, a driving profiling system is provided. “Driving” is related to driving a vehicle or movement of a vehicle such as a car, a bus, a transport vehicle, etc. The expressions “driver” or “vehicle” are related to a driving session, that is, the “driver behavior” during a driver session can be seen equivalently to characteristics of a trajectory recorded during the driving session. Clearly, a vehicle can be driven by several drivers or a driver may behave differently in various driver sessions. Therefore, technically, the trajectory relates to a driving session, i.e., from a start point to a destination, and the “driver” or “vehicle” shall be interpreted as such a single driving session or being related to such a driving session where vehicle movement data is captured to obtain a single trajectory. Further, the term “driving behavior” is used here equivalently for “driving style profile” or “driver profile”. The driver behavior corresponds to the movement of a vehicle. The profile represented as a trajectory characterizes a style, how a vehicle has been guided, independent on whether it has been guided by a human driver or an autonomous or assisted driving system, so that a “driver” may be interpreted abstractly. The system first has to collect data related to the vehicle movement, for example speed, acceleration, deflection of a steering wheel and/or along a path section on which a vehicle had moved. Further data may be collected. For example, camera data or environment data may be collected to filter data according to exceptional situations, where, for example, emergency braking had been carried out, which may be filtered out or where the driving has been performed on a snowy street. Thus, the vehicle movement data are samples, sensed by a speed meter, accelerometer, gyrometer, etc. and are collected and stored in a data storage. The vehicle movement data do not necessarily contain position data, however data such as speed and acceleration along the path, and may be recorded at, for example, pre-determined distance intervals or equal time intervals. The term “trajectory” is used here for samples ordered, for example, as a time series, in a vector or tensor comprising such vehicle movement data. Each entry of the vector represents a state describing the vehicle movement at a point on the path. E.g., the trajectory comprises a first speed at a first point, a second speed at a second point, etc. The 202201113 4 entries may be ordered chronologically. The state may include further parameters such as acceleration etc. Furthermore, the parameters such as speed and acceleration may be signed values and may be two-dimensional, or three-dimensional, or comprise direction information to cover the driving behavior when driving along a curve. The trajectory may be finite and comprise a first entry and a last entry. In the first step, the processing circuitry retrieves movement data from the storage. The data may be obtained in near real-time, but preferably, the data has been recorded. The movement data is represented as trajectory in form of states at distinct points. In the second step, the trajectory is partitioned in fine-grained linear segments, each having a start point, an end point and at least one intermediate point. “Linear” is to be seen as “pseudo”-linear as the states at the points need not to be exactly on a linear line and the data may be noisy. Nevertheless, small segments are extracted in the trajectory that may be approximately linear. Further, as described later, the slight non-linear segment is fitted to a linear segment. Several procedures may be applied to obtain the final linear segment, which is the input for the next steps. The linearity discussed herein relates to these states. Therefore, when speaking of “linearity of segments” or “linearity of points” in the segment, the linearity of the states at these points is meant. Since position, speed, and acceleration (positive and negative) are related to each other, only one of these parameters defining a state may be contained in the trajectory and the further parameters may be derived from that parameter. However, the trajectory may also contain more than one parameter. In this case, a separate evaluation may be performed for each parameter, or a merge may be performed at any of the steps described herein. Therefore, having more than one parameter may allow for verifying intermediate results and/or may provide a higher robustness and reliability due to the enlarged amount of data based on different measurements. In the fourth step, the processing circuitry determines clusters from the linear segments. The linear segments, or more exactly, the features of the linear 202201113 5 segments, herein also referred to as segment features, are input to a clustering algorithm, which generates clusters and determines a central point for each cluster. Such segment features may be, for example, a slope of the segment, an initial speed, herein also referred to as intercept, a mean value of the speed within the segment, and a standard deviation of the speed within the segment. Once the clusters are identified based on these segment features, a statistical value such as a mean value of the segment features over all segments in a cluster is determined for each cluster. Relative or absolute threshold values may be defined and applied to this mean value to determine a dominant feature of a cluster. Several clusters may have the same dominant driving feature. The clusters differ nevertheless by the further segment features. For example, dominant features are speed or acceleration. In case of speed, one cluster may contain segments with low speed, and a further cluster may contain segments with high speed. In case of acceleration, one cluster may contain segments with positive acceleration, and a further cluster may contain segments with negative acceleration. The final evaluation of the driving behavior may then be determined separately for each cluster but based on the same segment feature(s). For example, the segment feature relevant for the dominant feature “speed” is “mean velocity”. Then, in one cluster the mean velocity for driving at high speed, for example on a motorway, is analyzed, and in a further cluster the mean velocity for driving at low speed, for example in a city, is analyzed. The cluster represents therefore a driving characteristic such as acceleration, driving at high speed, driving at low speed, deceleration, etc. The values of each segment within a cluster, for example, slope, intercept, mean velocity and standard deviation of the velocity over a segment are assumed to be normal distributed. A normal test may be conducted to test if the assumption holds. The normal distribution is used for determining the driver behavior on a driving feature by analyzing distribution in each cluster. Criteria are defined how to classify a behavior of an individual driver on that cluster based on the distribution. For example, these classes may be mapped to the 1-, 2- and 3- sigma regions of the normal distribution. 202201113 6 The driver behavior on a dominant feature is thus classified by determining in which of these regions the driver’s mean value falls. The following embodiments are derived from the explanations above. According to an embodiment, the vehicle movement data contains a time series of vehicle speed data and / or vehicle acceleration data. The data may contain speed and/or acceleration samples captured at pre-determined and equal time intervals. Alternatively, the samples may be captured in equidistant intervals. Alternatively, the samples may be captured arbitrarily or statistically distributed, e.g., randomly. The speed and/or acceleration may be interpolated and/or extrapolated with respect to time points, a position and/or a distance. In any case, the samples or the data used for further processing is ordered with respect to time. The vehicle speed and the vehicle acceleration features are the most characteristic parameters according to which a driver behavior may be derived. Therefore, initially the trajectories are divided into segments of constant speed or constant positive or negative acceleration, which are clustered using the segment features and analyzed. This may also be applicable in case a curve is passed. In a “simple” case, the one-dimensional amount of, for example, the speed or acceleration may be sufficient. However, a more detailed analysis may be obtained if more-dimensional vectors of these parameters are used. According to an embodiment, determining linear segments of a trajectory, that is, of each of the trajectories, includes the following sub-steps: Using an initial set of trajectory points, determining a line from the initial set of trajectory points, extending the initial set of trajectory points with a further trajectory point, and checking whether the further trajectory point and/or the extended set of trajectory points fulfills at least one linearity criterion. If yes, it is determined that the segment comprises the further trajectory point. If no, it is determined that the segment does not comprise the further trajectory point and the further trajectory 202201113 7 point then is used as a trajectory point in an initial set of trajectory points of a next trajectory segment. The determination of a line is performed by a linear fit. If a point is added, the determination of the line is repeated including this point so that an adaptive linear fitting is performed. A “further point” is for example the next point in the vector or list of trajectory points of the trajectory, that is, the point after the set of initial points in the trajectory. “Further point” may also be denominated as neighboring point. In this way, first small linear segments, i.e., segments that are considered being nearly linear are determined. The initial set may be, for example, three points. The extension of the number of points is performed as long as they are located on a line, or near a line, and a linearity criterion is fulfilled. Algorithms using mean square error or summed square error determination may be applied. The linearity criterion may be checked for each new point individually, or for all points, that is, for the initial points and the new point. For that, the line or the linear fit may be re-calculated. A skilled person will understand that the initial set of trajectory points may include any points in a trajectory segment. E.g., a segment comprising a pre-defined number of trajectory points may be investigated where any at least points are used for defining a line, and the further points or all points in the segment may be checked for lying in the accepted linear region. Further, instead of linear segments, curved segments with a maximum curvature may be regarded, etc. According to an embodiment, determining segments of the trajectory includes merging linear segments for obtaining merged segments. The segmenting in the way as described above might not be optimal. The segmentation process might stop the segmentation process for a single linear segment too early. This may occur for example at the beginning of a segment, when only a few points are available, and, for example due to noise, the line is not yet stable, or due to other reasons. Therefore, two neighboring linear segments are investigated. If, according to a linearity criterion, the linearity of a combined segment 202201113 8 including both neighboring linear segments is better than a combined linearity each of neighboring linear segment taken alone, then the neighboring linear segments are merged to one segment. The metrics characterizing the linearity and the line representing the linear segment are determined. This step may be repeated with the next neighboring element such that the merged segment may include several original or “first” linear segments. According to an embodiment, the linearity criterion for merging is an Akaide information criterion (AIC) and/or a Bayes information criterion (BIC). For calculating the Akaide information criterion and/or Bayes information criterion algorithms, least square error determination or mean square error determination may be applied. The AIC may be calculated over separated linear segments and over the merged linear segment and compared with each other. The merging is performed only when the AIC of the merged linear segment is lower than the AIC over the separated linear segments. According to an embodiment, determining linear segments of the trajectory includes revising an end point of a linear segment by comparing the information criterion for a linear segment including the end point with the information criterion for the linear segment not including the end point. This step may be performed for the linear segment, independent of whether they have been merged or not. Preferably, it is performed after the merging procedure. A check is performed, whether a neighboring point of an end point in a linear segment fits better to its current segment or to the neighboring segment. If in the latter case the linearity is improved, the point is assigned to the neighboring segment, where it is the new start point. This procedure is performed recursively and/or iteratively. The described method above may be used, which compares the combined linearity of the single segments in both constellations. 202201113 9 According to an embodiment, determining clusters from the linear segments is performed using K-means clustering, wherein by K-means clustering, dominated driving features are determined. K-means clustering is a known and proven method, where the within-cluster sum of squares (WCSS) is minimized: According to an embodiment, before determining clusters, the number of clusters is determined. The determination of the number of clusters may be performed using the elbow check, where for a plurality of cluster numbers the K-means algorithm is applied. The higher the number of clusters, the worse becomes the distinction or definition of dominant features. The elbow method determines where higher numbers of clusters do not lead to clearly distinguishable features any more. The method is explained in more detail further below. According to an embodiment, a distribution over the mean values of one or more cluster features is determined, the distribution is divided into classes, a mean value of one or more segment features over all linear segments belonging to a trajectory in a cluster is calculated, and a driver behavior is determined by determining in which class the mean value of all linear segments belonging to a driver in a cluster falls, wherein the one or more segment features relate to the dominant feature of the cluster The distribution is a statistical distribution such as a normal distribution. The classification may be performed according to the 3-sigma rule of the normal distribution for each dominant feature. The dominant feature, for example speed, may relate to one cluster feature such as the average speed in the cluster. However, the dominant feature may also or alternatively relate to the cluster feature “standard deviation of the speeds”. If the dominant feature relates to one cluster feature, also the term “corresponds to” instead of “relates to” may be used. 202201113 10 According to an embodiment, the distribution is tested on being a normal distribution. A Shapiro-Wilk test may be applied for testing the normality of these feature values in the case of a small number of drivers. According to an embodiment, if the distribution is a normal distribution, classes are determined correspond to the 1-, 2- and 3- sigma regions of the normal distribution, and the driving behavior is classified corresponding to the classes. According to an aspect, a method for determining a driver behavior in a system comprising a data storage containing vehicle movement data and processing circuitry is provided, wherein the processing circuitry performs the following steps: retrieving the vehicle movement data from the data storage; extracting trajectories from the vehicle movement data, each trajectories being associated to a driver or a vehicle; determining linear segments of the trajectories; determining clusters from the linear segments; determining a dominant driving feature for each cluster; determining values for at least the dominant driving feature for each cluster from the linear segments within the respective clusters; determining driver behavior on a driving feature by analyzing distributions in each cluster. According to an aspect, a computer program is provided, which when being executed by the processing circuitry of a system as described herein, instructs the driver profiling system to perform the method as described herein. The computer program may be part of a computer program, but it can also be an entire program by itself. For example, the computer program element may be used to update an already existing computer program to get to the present invention. According to an aspect, a computer readable medium on which a computer program according to the previous claim is stored is provided. 202201113 11 The controller may comprise circuits without programmable logics or may be a micro controller or comprise a micro controller, a field programmable gate array (FPGA), an ASIC, a Complex Programmable Logic Devices (CPLD), or any other programmable logic devices known to person skilled in the art. The computer readable medium may be seen as a storage medium, such as for example, a USB stick, a CD, a DVD, a data storage device, a hard disk, or any other medium on which a program element as described above can be stored. Therefore, the cold-start problem, is solved by a novel driver profiling model comprising the following procedures: adaptive linear segment fitting, clustering, normality test, and 3-sigma rule check, to profile a small number of drivers, each with only a limited number of driving trajectories. Driving trajectories are segmented into fine-grained segments and further clustering of these segments is performed instead of the raw driving trajectories. As such, the rich features of the driving trajectories for driver behavior differentiation can be exploited, lifting the need of having a large-scale dataset. That is, the driving trajectory is appropriately partitioned into segments, and subsequently, the segments are clustered and the drivers’ behavioral patterns in each cluster are differentiated by applying the 3-sigma rule after the normality test. These and other features, aspects and advantages of the present invention will become better understood with reference to the accompanying figures and the following description. Short Description of the Figures Fig.1 shows a block diagram of a driver profiling system according to an embodiment. Fig.2 shows a block diagram of a driver profiling system according to a further embodiment. 202201113 12 Fig.3 shows a flow diagram with steps of a method according to an embodiment. Fig.4 shows selected steps of the method shown in Fig.3 in more detail. Fig.5 shows a diagram of a trajectory with linear segments. Fig.6 shows a diagram with clusters. Fig.7 shows a table with an example of clusters. Detailed Description of Embodiments Corresponding parts are provided with the same reference symbols in all figures. Fig.1 shows a block diagram of a driver profiling system 100 comprising a data storage 104 containing vehicle movement data and processing circuitry 102. The processing circuitry may comprise, for example one or more microprocessors 112, 114, FPGA(s), or other logic capable for executing logic instructions and processing data. The data storage 104 may be a volatile storage such as a buffer and/or a non-volatile storage. External or internal sensors 106, 108 provide the vehicle movement data. The data is stored and received by the processor. The processor is responsible for the data management, especially for storing and retrieving the vehicle movement data. Such sensors may be cameras, radar devices, Lidar devices, navigation device, odometers, speed meters, gyrometers, or any other sensor providing vehicle movement data. The processing circuitry may further receive movement data of other vehicles via a communication device 116. The processing circuitry 102 comprises a clock 110 that provides time stamps to relate sensor data to an absolute or at least a relative time stamp. The processing circuitry may alternatively receive required time stamps by the sensors themselves or by a clock external to the processing circuitry. If several sensors provide vehicle movement data, the data is time-related to each other, e.g., using synchronization mechanisms or intra-/extrapolating mechanisms. The arrangement of Fig.1 allows for driver profiling on the fly. 202201113 13 The system may also be a post-processing system as shown in the block diagram of Fig.2. In this case, it only comprises the data storage 104 in which the above-described data is collected and stored in a similar way, and used by the processing circuitry 102. The movement data of a plurality of vehicles is collected beforehand and stored in the data storage 104. Fig.2 has to be interpreted in a logical way, where the blocks 120, 122 indicate arrangements for collecting data of different vehicles, which are stored in the data storage 104. The skilled person is aware of data storage and transfer using wired, wireless or portable storage mediums. The processing circuitry 102 is performing instructions that fulfill the functions of the method as described at hand of Figs.3 and 4. Fig.3 shows a flow diagram with the steps of a method for determining a driver behavior or driving profile. Fig.4 shows steps 306, 308 and 314 of Fig.3 in more detail. The driver behavior corresponds to the movement of a vehicle and is digitally represented in a recorded trajectory. A driver behavior may be characterized by changes or constancy or deviating from mean values of speed, acceleration, deceleration, etc. These characteristics may be applied in one, two or more dimensions. When applying to two dimensions, also the behavior in curves can be determined. The steps of the methods are the following: First of all, in step 302 vehicle movement data is retrieved from the data storage. The vehicle movement data is for example one- two- or three-dimensional data that had been collected as described above. In a second step, 304, a trajectory from the vehicle movement data is extracted by the processing circuitry. A driving trajectory of a vehicle is a discrete time series T = {(t₁,v₁), (t₂,v₂), … , (tN,vN)},where (ti,vi) denotes the vehicle’s state vi at timestamp ti and N denotes the length of the discretized driving trajectory. Physically, the speeds of a vehicle in its driving trajectory are continuously changing in time due to inertia. Based on this 202201113 14 physical fact, the motion of a vehicle can be roughly assumed as uniform motion, uniform acceleration motion, or uniform deceleration motion in a short period. Under this uniform assumption, a vehicle is always in one of the three motion states and switches among them. For example, a vehicle is under uniform acceleration motion in the first period, under uniform motion in the second period, under uniform deceleration motion in the third period, and so on. As a result, the speed states in a driving trajectory form many linear segments when they are chronologically connected with their neighboring states. Fig.5 shows an exemplary trajectory T that contains states v₁…v_N. Each state relates to a point on the trajectory so that coordinates P = {(1, v1) … (i…vi₎… (N, vN)} comprising states vi at time points 1…N are obtained. The coordinate points in P may be defined by time intervals or distances. The coordinate points may be equidistant in terms of time or distance. However, the skilled person knows how to handle non-equidistant points for, e.g., relating the intervals to a spatial distance, speed, acceleration, etc. In Fig.5, the t-axis may represent, for example, points of time and the v-axis may represent speed. Then, it can be seen that the vehicle accelerates between time points 1 and 9 constantly, and decelerates in the following constantly but with a higher amount. Point 1 corresponds to a trajectory state v₁ comprising a speed value in this example. In the next step, 306, the processing circuitry determines linear segments of the trajectory. Fig.4 shows this step in more detail and the diagram of Fig.5 illustrates the linear segmentation. Starting from, for example, three initial points, a line may be drawn through these points by linear fitting. Every following point that is on or near this line may be added to these initial points so that a linear or nearly linear segment is obtained. The process of adding points stops if a point is too far away from the line or if a criterion is not satisfied any more such that a threshold is exceeded. A criterion for this determination may be the distance from the line or a threshold according to a square error. Known algorithms as the Original Least Square (OLS) method are applied to obtain a linear fit and for assessing the linearity. OLS uses the Sum of Square Error (SSE) and Mean of Square Error (MSE) for determining a square error. The SSE is a loss function that determines the sum of the squared differences between actual and 202201113 15 estimated values, that is, for example, speed values at different points of time in this application. It can be written generally as: , where is the value fitted by The MSE divides the sum by the number of values or points in the present application. The MSE can be written as: A linear fit in the form of a vector y = ax + b, is calculated, e.g., by deriving the SSE and setting the derivation to zero such that the coefficients a and b are obtained. MSE, which is independent on the number of points, may be used for estimating the goodness. The threshold may be a pre-defined threshold or, preferably, a dynamic threshold. The dynamic threshold may be defined as the MSE of the points already added to the segment. The extension criterion can therefore be defined as: ; where denotes the MSE without the new point and denotes the MSE including the new point. That is, a check is performed whether the MSE with the new point under investigation is lower than the MSE without the new point. If yes, the point is added to the segment such that the segment is extended by that point. If no, the segment is “closed” and the new point is the first point of the next segment. Hence, a linear fit over all points is calculated using OLS. In this way, linear segments such as 502, 504 and 508 are determined. It may occur, that the point-wise determination whether a current point is added to the current segment is not optimal since the point and maybe neighboring points might be for example outliers or a disturbance in an otherwise linear point sequence, so that the size of 202201113 16 the linear segment is not optimal. Therefore, next, two neighboring linear segments are regarded for determining whether they can be merged. For that, two information criteria (ic) are proposed, which are the AIC and the BIC using the MSE: where is the maximum likelihood of a model, d is the number of model parameters, and n is the number of samples. The smaller the aic/bic, the better the model. In regression analysis, if all the candidate models assume normally distributed errors with a constant variance, then the criteria can be reformulated as: ^where ^ ^{denotes the MSE} Given segments of a trajectory produced by thefirst procedure, this merging procedure attempts to merge two adjacent segments into one uniformed segment according to either AIC or BIC. Specifically, given two adjacent segments, the former is denoted as and the latter is denoted as , where 0 ^ i < j < q ^ N. Without being merged, the former ^{segment is linearlyfitted by and its MSE is denoted as} Similarly, the latter segment is linearlyfitted by with its MSE ^{denoted as} ^^{i:q. There are two linear functions with four parameters} where s denotes separatefittings, With merging performed, the two segments, Ti:j and Ti:q, are merged into one uniformed segment denoted as Ti:q = . Then the 202201113 17 segment Ti:q is linearlyfitted by with the corresponding MSE ^{denoted as} ^^{i:q. Similarly, the information criterion of the mergedfitting is computed} as follows: where m denotes the mergedfitting. The AIC and/or the BIC are calculated over the separate MSEs, where each MSE is calculated for one segment each, and the AIC and/or the BIC is calculated over merged segments including the points of two neighboring linear segments. That is, for the merged segments, one MSE is calculated. The results are compared to each other. If the AIC and/or the BIC of the separated segments is lower than the AIC and/or the BIC of the merged segments, the segments are maintained separated, otherwise, the segments are merged to one segment, which then is linear fitted. The merged segment can then be investigated regarding a further merge with the next linear segment, etc. In a last partitioning step, the first and last points of a segment, also referred to as start points and end points, or as edge points, are investigated, because the current assignment to one of the neighboring segments might not be optimal. Therefore, such an edge point is temporarily moved to its neighbor segment, and the ics for both cases are compared to each other. The final assignment is performed according to the lowest ic. This step may be performed moving with edge point alone or more neighboring points. Further, when a point has been assigned to a neighbor segment, the resulting new edge points may be investigated. In the next step, step 308, clusters from the linear segments of all vehicles (or drivers) are determined. This step may be performed by first determining a number of clusters and subsequently performing a K-means algorithm. The objective of K-means is to minimize the within-cluster sum of squares (WCSS) 202201113 18 Given a dataset with n samples with d features, K-Means aims to divide with k clusters by finding k centroids. The k clusters are denoted by {C1, C2, · · · , Ck}, where ci is the centroid of Ci, i = 1, 2, · · · , k. The objective of K-Means is to minimize the within-cluster sum of squares (WCSS): where denotes the square of 2-norm of the vector x. The number of clusters k can be, for example, determined with the heuristic so-called “elbow”-method, where K-means is performed for a several numbers of clusters. The elbow method usually consists of two steps. Firstly, K-Means is conducted with various numbers of clusters, e.g., k ∈ [2, 30], and WCSS for each k ∈ [2, 30] is computed. In this step, WCSS is regarded as a function of the number of clusters, k. Secondly, the function of WCSS with respect to k is plotted as a polyline and the elbow of the polyline is chosen as the number of clusters of K-Means. The resulting clusters represent the vehicle movement features, and a driver profile can be derived. Fig.6 shows exemplarily a diagram with four visually recognizable clusters. The segment features of a cluster in a driving scenario may be speed (e.g. speed limit in a city or on a motorway), acceleration/deceleration (e.g. when a traffic light switches to green), acceleration/deceleration at a usual speed acceleration (e.g. braking when a traffic light switches to red), etc. In step 310, the segment features are analyzed and a significant segment feature is determined as dominant feature. For example, if the speed in a cluster is nearly constant, that is the acceleration is nearly zero, the speed may be the dominant feature. There may exist one, two, three or more clusters where the speed is the 202201113 19 dominant feature but in each cluster, the average speed over all segments in a cluster differ. That is, for all these clusters, the segment feature “mean”, i.e. average speed is the decisive feature used in the following steps for determining the driving behavior. Fig.7 shows a table with an example of clusters, segment features and dominant feature. From all linear segmented trajectories clusters, C1…C5 have been determined. The segment features are Slope (acceleration), Intercept (initial speed), Mean (average speed), and Std (standard deviation of the speed). For clusters C1…C3 the dominant feature “speed” has been determined. The reason for this is that the slope in these clusters is nearly zero, such that “Mean”, i.e., the speed is the significant feature. The three clusters differ at least in different mean speed, which is low speed LS, medium speed MS, and high speed HS as shown in the table. Similarly, for clusters C4 and C5, Slope is high (3.516) or low (-4,646) indicating an acceleration or a deceleration. Therefore, the dominant feature is acceleration represented by the segment feature “slope”. The clusters differ by the amount and/or sign of the slope and are denominated AC (acceleration) or DE (deceleration). In step 312, the values resulting from the means over the segments for a segment feature are determined. In principle, these values are already available; however they are prepared for the next step 314. The means over of the segments related to a cluster of each available trajectory are calculated. These means are assumed to be normal distributed. A normality test is performed to test the distribution. The normality test of samples is tested by the Shapiro-Wilk test. The null hypothesis of the Shapiro-Wilk test is that samples come from a normally distributed population. The test statistic is defined as follows: (2) 202201113 20 where z(i) is the i-th smallest number . The coefficient a_i is computed by: where consists of the expected values of the order statistics of i.i.d. random variables sampled from the standard normal distribution and V is the covariance matrix of those normal order statistics. Now, the mean of the corresponding segments of driver’s trajectory, that shall be profiled, may fall into the following regions of the normal distribution. If the mean is smaller than -σ (sigma), the label is conservative; If the mean is between -σ and σ, the label is normal; if the mean is greater than σ, the label is aggressive. In a normal distribution, the empirical rule (aka 3-sigma rule) is that 68%, 95%, and 99.7% of samples lie in the one, two, and three standard deviations of mean, respectively. In the empirical rule, an observation, X, lies within the one standard deviation of mean with the probability of Based on these, two corollaries can be drawn: and (6) Given a driving feature, intuitively, it is assumed that the larger its value, the more aggressive a driver is on this feature (If the smaller its value, the more aggressive, we can use the inverse of the value of this feature). Therefore, if 202201113 21 a driving feature of a driver lies in (5), (4), or (6), the driver is deemed to be conservative, normal, or aggressive on the feature, respectively. As an example, the system and the method may be applied to a usage-based Insurance (UBI). Traditional vehicle insurance policy around the world, charges the vehicle owner a lump-sum amount for their vehicle based on vehicle’s age, model, and make. This means that all the customers have to pay a similar amount for their vehicle insurance, irrespective of their driving behavior. Driver behavior profiling acts as an essential input in these UBI schemes by categorizing drivers into different categories based on their driving habits. In UBI schemes, drivers are provided with either a driving score (0–100) or a rating (good driver-poor driver) based on their driving behavior. Based on driving score/rating, the drivers with a higher risk of crashes are penalized by increasing the insurance premium amount, and drivers with a lower risk of crashes are given incentives by reducing their insurance premium amount. As a further example, the system and the method may be applied to a fleet management system. For a transportation firm, it is crucial to manage their fleet drivers in an efficient way, which will reduce the operational cost and consumption. One of the major objectives of the fleet management system is to determine the most appropriate driver for performance appraisal. Driver profiling acts as an important input for these fleet management systems, categorizing drivers into different categories based upon their driving behavior. Further, based on behavioral profile, drivers are provided with active feedback to promote safe driving. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from the study of the drawings, the disclosure, and the appended claims. In the claims, the word “comprising” does not exclude other elements or steps and the indefinite article “a” or “an” does not exclude a plurality. A single processor or other unit may fulfill the functions of several items or steps recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a 202201113 22 combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope of the claims. Reference Numerals 100 driver profiling system 102 processing circuitry 104 data storage 106 external sensors 108 internal sensors 110 clock 112 microprocessor 114 microprocessor 116 communication device 120 arrangement for collecting data 122 arrangement for collecting data 300 method for determining a driver behavior or driving profile 302…314 steps of method 300 502, 504, 508 linear segments

Claims

202201113 23 Patent Claims 1. Driving profiling system (100) for determining a driver behavior comprising a data storage (104) containing vehicle movement data, and processing circuitry (102) configured to perform the following steps: retrieving (302) the vehicle movement data from the data storage; extracting (304) trajectories from the vehicle movement data, each trajectories being associated to a driver or a vehicle; determining (306) linear segments of the trajectories; determining (308) clusters from the linear segments; determining (310) a dominant driving feature for each cluster; determining (312) values for at least the dominant driving feature for each cluster from the linear segments within the respective clusters ; determining (314) driver behavior on a driving feature by analyzing distributions in each cluster. 2. Driver profiling system (100) according to claim 1, wherein the vehicle movement data contains a time series of vehicle speed data and / or vehicle acceleration data. 3. Driver profiling system (100) according to claim 1 or 2, wherein the determining (306) linear segments of the trajectory includes: using an initial set of trajectory points; determining a line from the initial set of trajectory points; extending initial set of trajectory points with a further trajectory point; checking whether the further trajectory point and/or the extended set of trajectory points fulfills at least one linearity criterion; if yes, determining that the segment comprises the further trajectory point; if no, determining that the segment does not comprise the further trajectory point and using the further trajectory point as a trajectory point in an initial set of trajectory points of a next trajectory segment. 202201113 24 4. Driver profiling system (100) according to any one of the previous claims, wherein the determining (306) linear segments of the trajectory further includes merging linear segments (502, 504) for obtaining a merged segment (508). 5. Driver profiling system (100) according to any one of claims 3 or 4, wherein the linearity criterion for merging is an Akaide information criterion (AIC) and/or a Bayes information criterion (BIC). 6. Driver profiling system (100) according any one of claims 3 to 5, wherein the determining (306) linear segments of the trajectory further includes revising an end point of a linear segment by comparing information criterion for a linear segment including the end point with the information criterion for the linear segment not including the end point. 7. Driver profiling system (100) according to any one of the previous claims, wherein the determining (308) clusters from the linear segments is performed using K-means clustering. 8. Driver profiling system (100) according to any one of the previous claims, wherein before the determining clusters (308), the number of clusters is determined. 9. Driver profiling system (100) according to any one of the previous claims, wherein the processing circuitry is further configured for determining a distribution over the mean values of one or more segment features; dividing the distribution into classes; calculating a mean value of one or more segment features over all linear segments belonging to a trajectory; and determining the driver behavior by determining in which class the mean value of all linear segments belonging to a driver in a cluster falls, wherein the one or more segment features relate to the dominant feature of the cluster. 202201113 25 10. Driver profiling system (100) according to the previous claim, wherein the distribution is tested on being a normal distribution. 11. Driver profiling system (100) according to the previous claim, wherein, if the distribution is a normal distribution, classes are determined correspond to the 1-, 2- and 3- sigma regions of the normal distribution; and the driving behavior is classified corresponding to the classes. 12. Method for determining a driver behavior in a system comprising a data storage containing vehicle movement data and processing circuitry, comprising the steps: retrieving (302) the vehicle movement data from the data storage; extracting (304) trajectories from the vehicle movement data, each trajectories being associated to a driver or a vehicle; determining (306) linear segments of the trajectories; determining (308) clusters from the linear segments; determining (310) a dominant driving feature for each cluster; determining (312) values for at least the dominant driving feature for each cluster from the linear segments within the respective clusters; determining (314) driver behavior on a driving feature by analyzing distributions in each cluster. 13. Computer program, which when being executed by the processing circuitry of a system according to claim 1, instructs the driver profiling system to perform the method according to the previous claim. 14. Computer readable medium on which a computer program according to the previous claim is stored.