Method for determining safety-critical regions along motorway-like roads

Abstract

The invention relates to a method for determining at least one safety-critical peak region (R1, R2, R3) on the basis of an indicator value (Z) for statistical characterization of event descriptions detected by at least one vehicle in a location-related manner, wherein those inflow- and outflow-free road sections of which the section length is greater than or equal to a predetermined minimum section length are identified on the basis of a road map or navigation map. Identified road sections which adjoin one another are connected to form in each case a section chain (10) with a starting point (11) and an end point (12). Event descriptions comprising a geoposition (P1, P2) and at least one parameter of a vehicle environment model are detected by vehicles and transferred to an analysis system. The analysis system in each case assigns a nearest road network location (O1, O2) of a section chain (10) to a geoposition (P1, P2) of an event description and describes it by a longitudinal distance (L, L1, L2) in relation to the starting point (11) of the section chain (10). For at least one section chain (10), a profile of a density value (p) of event descriptions that is related to the longitudinal distance (L, L1, L2) from the starting point (11) is determined by means of a kernel density estimator of a predetermined bandwidth. The profile of the density value (p) is transformed, using local, bandwidth-related, Z standardization, into a Z profile (Z1, Z2) of a Z value (Z) which is mean-free in relation to bandwidth. In each case a region of road network locations (O1, O2) along the motorway-like road (100) which are connected along a section chain (10) between a peak beginning (Lmin) and a peak end (Lmax) and at which the Z value (Z) exceeds a predetermined threshold value (θ) is identified as a safety-critical peak region (R1, R2, R3).

Description

[0001] Mercedes-Benz Group AG

[0002] Procedure for determining safety-critical regions along highway-like roads

[0003] The invention relates to a method for determining at least one safety-critical relevant peak region along a motorway-like road using an indicator value for the statistical characterization of location-related event descriptions recorded by at least one vehicle according to the preamble of claim 1.

[0004] Document DE 102022 105 919 A1 describes a system and a method for the early detection of structural hazards in road traffic using a digital road network map. For this purpose, accident data, user input data, and sensor data are assigned to segments of the map according to their georeferencing by a computer system. Based on one aspect, the computer system performs an evaluation, including at least a frequency determination and / or a data comparison.For early detection purposes, a georeferenced segment to which no accident data has been assigned is identified as a potential hazard location if sensor data or user input data has been assigned with a predetermined frequency, and / or if a data comparison, in particular a comparison with an artificial intelligence (AI)-based model, of user input data and / or sensor data shows that a predefined degree of feature similarity or feature correlation with critical patterns exists. This enables early detection of hazard locations. Furthermore, it is proposed to determine a hazard score for identified hazard locations.

[0005] Document US 2023 / 0245560 A1 describes a system for determining and ranking location-based hazards based on events recorded by vehicles and / or an accident database. The system includes an interface and a processor. The interface is configured to receive accident data and / or event data. The processor is configured to determine groups of incidents, assigning to each group a set of accidents from the accident database and / or a set of events from the event data, grouped according to their geographical proximity. The processor is further configured to determine a set of traffic estimates that are related to a group of incidents and, based on the set of accidents, events, and traffic estimates, to rank the groups of incidents.

[0006] Document DE 102019215 587 A1 describes a method for calculating relative positive or negative accelerations (RPA, RNA) using crowdsourcing, comprising the following steps: a) generating a digital map in which different route segments are connected by nodes; b) calculating a relative positive or negative acceleration (RPA, RNA) of a vehicle traveling in a route segment for that route segment (A, B, C,..., M), wherein the relative positive or negative acceleration (RPA, RNA) in the vehicle is calculated and sent to a backend, or wherein the vehicle sends data containing current geocoordinates, current vehicle speeds and / or current vehicle accelerations to a backend, which calculates the relative positive or negative acceleration (RPA, RNA) of the vehicle for that section of the route; and c) calculating a mean or distribution of the calculated relative positive or negative accelerations (RPA, RNA) of several vehicles for that section of the route in the backend.

[0007] Document DE 102019000630 A1 discloses a method for assigning a common geoposition to a plurality of events caused by a common root cause and recorded by at least one vehicle, wherein a separate geoposition is recorded for each individual event. The recorded events are indexed in a lexicographically sortable manner. Geopositions to which a plurality of recorded events are assigned are called hotspots. A method is proposed for determining maximum hotspots by grouping spatially adjacent and sufficiently close hotspots. Hotspots to which a set of events is assigned that is a proper subset of the set of events assigned to another hotspot are eliminated. A common geoposition is assigned to the hotspots thus merged into a maximum hotspot.Document WO 2021191051 A1 discloses a method for identifying potential hazards in road traffic using connected vehicles. An event indicating a potential hazard is recorded and transmitted, along with its geolocation, to a central computer unit. The transmitted event is entered into a digital map as a hotspot if a large number of similar events with the same geolocation are recorded. Contextual information is added to the transmitted event. The hotspot is analyzed to identify potential hazards. A current hotspot is compared with confirmed hotspots. The hotspots are visualized on a platform. The geolocation of a hotspot containing specific traffic-critical events is transmitted to vehicles located near that hotspot.

[0008] A hotspot, as known from the prior art, can thus be understood as a localized cluster of events that share the same or a very similar geoposition. Such a cluster (or hotspot) is characterized by an absolute frequency and a maximum radius. The absolute frequency indicates how often (within a predetermined time period) an event was recorded within the maximum radius around the hotspot.

[0009] This well-known form of describing a hotspot has the disadvantage that traffic-related events typically do not occur and are recorded in a circular area around a geoposition, but rather along traffic routes.

[0010] Furthermore, the analysis of such hotspots is made more difficult by the fact that often not the absolute, but the relative frequency of an event (related to the number of vehicle passages at the geoposition) is relevant for an evaluation, because it estimates the probability of the event occurring when passing through this geoposition.

[0011] Since the number of vehicle passages for the geolocation of a hotspot is typically unknown or only imprecisely known for technical reasons and due to data protection restrictions, the relative frequency of an event cannot be determined, or only roughly. Furthermore, for certain types of roads and traffic layouts, especially for highway-like roads designed for high-speed traffic with multiple parallel lanes in each direction, events occur essentially along the entire length of such a road or traffic layout, meaning that clusters or hotspots cannot be located and isolated solely based on the relative frequency of events.

[0012] Therefore, there is a need for an improved method for the detection, classification and analysis of hotspots, which are characterized by the conspicuousness of an indicator value, with which event descriptions recorded at such hotspots are characterized.

[0013] The publication by E. Schubert, A. Zimek, H.-P. Kriegei: Generalized Outlier Detection with Flexible Kernel Density Estimates. Proceedings of the 14th SIAM International Conference on Data Mining (SDM), Philadelphia, PA, 2014, pp. 542–550, https: / / doi.org / 10.1137 / 1.9781611973440.63 describes a method for analyzing the relationship between the estimation of a density function and the identification of outliers based on such a density function.

[0014] The invention is based on the objective of providing an improved method for determining an indicator value for the statistical characterization of location-related event descriptions recorded by at least one vehicle.

[0015] The problem is solved according to the invention by a method having the features of claim 1.

[0016] Advantageous embodiments of the invention are the subject of the dependent claims.

[0017] To detect safety-critical hotspots, parameters of the vehicle's surroundings are continuously recorded in at least one vehicle. For example, a dynamic vehicle environment can be recorded by using a stereo camera and / or radar and / or lidar system, or similar measurement methods, to measure distances (relative to the vehicle's own position), speeds, and trajectories of surrounding vehicles, people, and / or other road users. Additionally or alternatively, a static vehicle environment can be recorded using the vehicle's sensors, such as the arrangement and properties of lane markings and / or traffic signs and / or traffic control devices. For example, lane markings can be recorded and measured with regard to their spacing, marking types, lane widths, curvatures, and similar parameters.

[0018] In parallel with the recording of such vehicle environment parameters, the vehicle position is continuously determined, i.e. the geoposition recorded by means of a global navigation satellite system (GNSS) or with a similar geoposition determination system.

[0019] The objects and their properties detected in the static and / or dynamic vehicle environment by means of various sensors are fused into an environment model, which can also include map data from a road or navigation map and optionally additional dynamic, geoposition-related live data, such as the current traffic volume.

[0020] This complex environmental model defines triggers that describe safety-critical events relevant to vehicle operation (especially those related to traffic). Such events can include warnings triggered by vehicle sensor parameters. They can also include interventions by an assistance system that monitors the vehicle's surroundings. In particular, these safety-critical events can also include manual driver interventions, especially abrupt braking and / or steering maneuvers, such as emergency braking.

[0021] When such a trigger is activated, an event description is generated and sent from the respective vehicle to the backend of an analysis system. The event description includes the specific geolocation at which the trigger was activated, as well as the sensor data and parameters recorded by the vehicle's sensors that are relevant for describing the event.

[0022] The analysis system can, for example, be implemented on a server or in a cloud and provide a set of predefined services and / or functions as a backend in the form of an application programmer interface (API).

[0023] The analysis system generates an analysis map from event descriptions transmitted to the backend by one or more vehicles, based on a road or navigation map or similar map data. In such map data, transportation routes, especially roads, are assigned geopositions, which are referred to below as road network locations. In other words, while geopositions can generally be assigned any topographic features, road network locations denote those geopositions that are accessible via a road.

[0024] First, the longest possible continuous road segments, uninterrupted by intersections, are identified. Specifically, road segments whose length meets or exceeds a predetermined minimum can be selected for the analysis map. Typically, these road segments will be highways or expressways with few intersections. Road segments shorter than the predetermined minimum length are filtered out and not included in the analysis map.

[0025] Adjacent road segments are then linked (in the manner of a linked list) to form segment chains; in other words, an ordered sequence of road segments is transformed into a topologically one-dimensional (i.e., only linearly extended) segment chain, which has a single starting point and a single endpoint along its linear extent. Each segment chain is uniquely assigned a chain identifier and a starting point (with respect to the analysis map). Thus, every location along a segment chain can be uniquely determined by the chain identifier and a distance from the starting point of the segment chain.

[0026] In a method for determining at least one safety-critical peak region of an indicator value for the statistical characterization of event descriptions recorded by at least one vehicle along a highway-like road, sections of road without inflows or outflows are identified based on a road or navigation map or similar mapping material in which traffic routes are mapped with reference to geopositions. Along such road sections, the number of passing vehicles can be assumed to be approximately the same everywhere. In this way, longer traffic routes are divided into road sections that are connected to at least one other road section at intersections, junctions, exits, entrances, and similar junctions.

[0027] In the following, the term "motorway-like road" shall be understood to mean expressways designed for high-speed traffic, in particular expressways with multiple lanes in each direction.

[0028] Among the road sections identified in this way, those are identified whose section length is equal to or greater than a predetermined minimum section length.

[0029] This makes it possible to select road segments exclusively, or at least essentially only, along specific types of transport routes for which a substantially uniform vehicle throughput (along the respective route) can be assumed to be a good approximation. In particular, the segment length can be chosen so that the selected road segments are classified as motorways or expressways.

[0030] In a subsequent step, among the identified road segments, those are determined that are connected along a traffic route via a crossroads, junction, exit, entrance or similar intersection and where, consequently, the geoposition of one end of a first identified road segment coincides with the geoposition of one end of a second identified road segment.

[0031] A chain of such road segments, linked together like a chained list, is called a segment chain. A segment chain extends linearly from a starting point to an endpoint and therefore has no branches along which a traffic flow could distribute itself. Consequently, the assumption of a uniform traffic flow throughout such segment chains is a particularly good approximation.

[0032] At least one event description is recorded for at least one vehicle and transmitted to an analysis system. Such an event description includes at least one parameter of a vehicle environment model, which encompasses the vehicle's static and / or dynamic environment as well as its operating and / or control parameters. Furthermore, such an event description includes the geolocation at which these parameters were recorded.

[0033] Event descriptions are captured and transmitted particularly when the vehicle environment model describes a condition or change of condition that is particularly relevant for assessing traffic flow and / or vehicle condition.

[0034] The analysis system evaluates incoming and / or stored event descriptions. For each geoposition of an event description, which typically does not correspond exactly to a road network location due to tolerances and inaccuracies in data acquisition by a geolocation system, the analysis system assigns the nearest road network location within a segment chain, typically by (perpendicular) projection onto the nearest segment chain.

[0035] This nearest road network location is described by a unique identifier of the segment chain and by a longitudinal distance that specifies the distance (in predetermined length units chosen identically for all segment chains) from the starting point of the segment chain.

[0036] Furthermore, the analysis system uses a kernel density estimator to determine, for at least one segment chain, the density of event descriptions relative to the longitudinal distance from the starting point. For this purpose, densities (density values) are determined at incremental, preferably equidistant, longitudinal distances from the starting point of the segment chain. These densities estimate the frequency of event descriptions within an increment of the longitudinal distance. By using a kernel density estimator, the variance of these determined values ​​can be reduced by considering the frequencies of event descriptions from neighboring increments of the longitudinal distance when determining the density within an increment (in the manner of a weighted, moving average).

[0037] The density is determined at a certain location, defined by the longitudinal distance from the starting point of the segment chain, by projecting the longitudinal distances assigned to the event descriptions onto this segment chain.

[0038] The core density is estimated using a core density estimator with a predetermined bandwidth. In this process, a sequence of Dirac pulses arranged at the assigned longitudinal distances is convolved with a non-negative, preferably symmetric, core weighting function.

[0039] Methods for selecting the kernel weighting function, in particular for choosing the bandwidth of such a kernel weighting function, are known from the prior art, for example from the publication Wand, MP; Jones, MC (1995). Kernel Smoothing. London: Chapman & Hall / CRC. ISBN 978-0-412-55270-0.

[0040] Each density of event descriptions determined along a segment chain is then transformed into a Z-value using local Z-standardization.

[0041] Local Z-standardization, as used here and in the following, refers to the transformation of a density value p. e understood, which was determined for one or for a set of road network locations, taking into account a set of density values ​​p^ Pz, ...pj ...p, w , which within a predetermined environment, referred to as bandwidth, along the respective segment chain around the road network location of the density value f were determined. For this set of density values, a (local) mean and a (local) standard deviation are calculated. determined. The local Z-standardization transforms the density value p. f into a Z-value Z, according to the regulation

[0042] In other words, local Z-standardization transforms the density value into a Z-value that, relative to the environment known as the bandwidth, represents the deviation of the density from the mean relative to the standard deviation.

[0043] The higher the local Z-value, the more conspicuous the point under consideration is compared to its surroundings, measured by the variability of the surroundings, regardless of the absolute density. A higher Z-threshold allows the result set to be reduced to more conspicuous locations. This makes it possible to characterize and identify road network locations as conspicuous based on the Z-value, especially those road network locations where the probability of an event description being captured / reported is significantly higher than at road network locations in a predetermined environment along the segment chain. An advantage of this method over the prior art is that it enables a reliable characterization of such conspicuousness independent of the recording of the absolute (total) number of vehicles passing a road network location.

[0044] This advantage is achieved, firstly, by selecting road network locations with comparatively homogeneous vehicle density through the formation of segment chains according to the invention. Secondly, this advantage is achieved by transforming density values ​​into an indicator via local Z-standardization, which is largely insensitive to fluctuations in the dispersion of density values ​​along a segment chain.

[0045] At least one safety-critical peak region is identified as an area where, along a chain of segments between a peak start and a peak end, a Z-value exceeding a predetermined threshold has been assigned to connected road network locations. This allows for the highly reliable identification of areas with safety-critical traffic situations, such as accident hotspots.

[0046] Compared to conventional methods for identifying safety-critical locations based on clustering event descriptions, the proposed method has the advantage that safety-critical peak regions can also be identified when a largely homogeneous distribution of events is recorded along the segment chain, events that are solely attributable to normal traffic flow (for example, normal lane changes or braking maneuvers along roads with multiple lanes in each direction), and when a location-specific cluster of safety-critical events is only slightly apparent compared to this largely homogeneous distribution. The proposed method is therefore particularly well-suited for identifying and locating safety-critical areas along highway-like roads. Furthermore, segment chains associated with highway-like roads exhibit particularly good homogeneity of traffic flow.In other words, due to the comparatively low inflows and outflows, density values ​​recorded along such segment chains are particularly comparable; in particular, increased density values ​​of event descriptions reliably indicate an exceptional traffic situation. This enables a particularly sensitive detection of safety-critical areas.

[0047] Preferably, the predetermined threshold is set to one. Based on a statistic of the density values ​​that approximately follows a normal distribution, those areas are identified as peak regions that exhibit a higher density of event descriptions than approximately 85 percent of all road network locations in a segment chain. This allows for a particularly reliable determination of peak regions with conspicuous traffic situations.

[0048] In one embodiment of the method, braking maneuvers and / or warnings are recorded as safety-critical events. A warning is detected based on at least one safety-relevant parameter of the vehicle environment model. Braking maneuvers can, for example, be recorded as emergency braking and / or hazard braking, depending on the deceleration initiated (automatically by a driver assistance function or by manual intervention of the driver). Optionally, other safety-critical events can also be recorded, such as particularly abrupt steering movements indicating an evasive maneuver. By recording such safety-critical events, safety-critical areas can be monitored with a particularly high degree of reliability (i.e., with high sensitivity and / or high specificity).

[0049] In a further development of this embodiment, a common density value (related to each location on the road network) is determined from all the recorded safety-critical events, with braking maneuvers being weighted more heavily than warnings in the common density value. Preferably, braking maneuvers are weighted three to seven times more heavily than warnings. Particularly preferably, braking maneuvers are weighted five times more heavily than warnings. This enables a particularly specific (i.e., with a low number of false positives) detection of safety-critical peak regions.In one embodiment of the method, the predetermined minimum length for road segments to be used as the basis for forming segment chains is selected such that at least one segment chain is formed exclusively or predominantly from road segments that are assigned to a motorway or expressway.

[0050] In one embodiment of the method, the vehicle environment model merges a static and a dynamic vehicle environment. The "static vehicle environment" here refers to the vehicle environment that is essentially unchanged over time, particularly unaffected by the activities of other road users, such as the arrangement and properties of lane markings and / or traffic signs and / or traffic control devices. A static vehicle environment can be captured, for example, using surround-view cameras, distance sensors, radar, and lidar sensors.

[0051] The term "dynamic vehicle environment" here refers to the vehicle environment that changes over time, in particular the vehicle environment altered by the movement of other road users. For example, a dynamic vehicle environment can be captured by using a stereo camera and / or a radar and / or lidar device, or similar measurement methods, to measure distances (relative to the vehicle's own position), speeds, and trajectories of surrounding vehicles, people, and / or other road users.

[0052] In one embodiment of the method, a peak region is assigned an identifier for the respective segment chain (which is uniquely assigned, for example, as a number or string), a longitudinal distance of the peak start relative to the starting point of the respective segment chain, a longitudinal distance of the peak end relative to the endpoint of the respective segment chain, a maximum value of all Z-values ​​within the peak region, and a description of the peak region's geometry. A description of the peak region's geometry can, for example, be specified as a set of geopositions that are assigned to all or some (for example, selected at equidistant intervals) road network locations within the peak region. Using such characteristics, the evaluation of peak regions is particularly easy and reliable.In one embodiment of the method, the vehicle environment model comprises at least one object and / or parameter that has been detected by vehicle sensors. Based on the vehicle environment model, a trigger is determined that identifies and / or marks an event worthy of consideration, for example, an event relevant to controlling and driving the vehicle. Based on such a trigger, an event description is generated and transmitted to the analysis system.

[0053] Such a trigger allows for the targeted creation of event descriptions that are interesting and relevant for subsequent analysis.

[0054] Preferably, a trigger marks a warning issued by a driver assistance system and / or an intervention by such a system. This allows driving situations in which the driver's behavior potentially requires correction to be captured in event descriptions. Determining a statistical indicator for such driving situations in relation to road network locations (i.e., specific locations within road segments) enables a particularly meaningful assessment of potentially dangerous locations and the identification of corresponding peak regions, such as potential accident hotspots. This allows for the identification of insights that contribute to improving road safety.

[0055] In one embodiment of the method, event classes are defined and recorded in a list of event classes. Event classes can be (by way of example only) a braking maneuver initiated by a driver assistance system, a vehicle parameter indicating reduced vehicle stability, or a potential collision with an object in the vehicle's surroundings detected by a driver assistance system.

[0056] Each event description is assigned at least one such event class. To determine the density of event descriptions, only those (filtered) event descriptions whose event classes meet a specific filter criterion are used. For example, only event descriptions from a specific event class or from a subset of event classes are considered. This allows for a particularly specific evaluation of density values ​​and / or a particularly specific determination of peak regions. In one embodiment, an event description includes a direction of travel that refers to a segment chain pointing from its starting point to its endpoint.To determine the density of event descriptions, only those descriptions are considered where the direction of travel is the same, meaning where the vehicle (along the segment sequence) moved in the same direction. This makes it possible to identify direction-related anomalies with particular precision (for example, differentiating between downhill and uphill travel on a road segment with a steep gradient, or differentiating between other direction-dependent hazards such as obstructed views, ramps, exits, or curve radii).

[0057] In one embodiment of the method, Z-values ​​determined for road network locations along at least one segment chain are visualized on a map. This visualization can be achieved, for example, through color coding, with different colors assigned to different Z-value ranges. Additionally or alternatively, Z-values ​​can be visualized by markings applied perpendicular to the respective segment chain, the length of which is determined based on the Z-value for that location. Preferably, the length of such a marking is proportional to the Z-value.

[0058] This allows for a particularly clear presentation of gradual differences in anomalies, especially hazards, along a chain of sections, without the need for arbitrary threshold setting and segmentation of peak regions.

[0059] In one embodiment, a road map is provided on which the location of at least one safety-critical peak region is marked for each motorway-like road. Such road maps enable a particularly clear representation and assessment of road traffic safety across large motorway-like road networks.

[0060] Exemplary embodiments of the invention are explained in more detail below with reference to the drawings. These show:

[0061] Fig. 1 schematically shows an analysis map with a section chain,

[0062] Fig. 2 schematically shows the determination of a Z-standardized Z-value for a density of event descriptions determined by means of a kernel density estimation.

[0063] Fig. 3 schematically shows the direction-dependent representation of Z-values ​​along a chain of sections, as well as

[0064] Fig. 4 schematically shows the direction-dependent representation of Z-values ​​along a motorway-like road.

[0065] Corresponding parts are marked with the same reference symbols in all figures.

[0066] Figure 1 schematically shows an analysis map with a segment chain 10. The segment chain 10 was formed from the linking of intersection-free road segments (not shown in detail) whose length reaches or exceeds a predetermined minimum length. Each of these road segments (not specifically labeled in Figure 1) comprises at least one, typically several, road network locations 01, 02, which can, in principle, be arbitrarily close together (limited only by the resolution of the underlying map data). For clarity, only a first and second road network location 01, 02 are shown as examples in Figure 1.

[0067] The segment chain 10 can be obtained, for example, by analyzing a road or navigation map. Its position and extent are shown in Figure 1 in relation to a geocoordinate system, which is roughly simplified and schematically has a longitudinal coordinate (geographic longitude), represented here as the x-direction x, and a latitudinal coordinate (geographic latitude), represented here as the y-coordinate y.

[0068] The section chain 10 has a starting point 11 at a first end and an endpoint 12 at the opposite second end and is identified by an identifier, which in this case is only chosen as an example of the string “#2”.

[0069] At least the starting point 11 and the endpoint 12 are each to be assigned to both a (not specified in detail here) road network location 01, 02 and a (also not specified in detail here) geoposition P1, P2.

[0070] Each point along the segment chain 10 can be uniquely determined by a longitudinal distance L from the starting point 11 (specified in arbitrarily selectable, but identical distance units for all segment chains 10) and by the identifier of the segment chain 10.

[0071] Each point in a segment chain 10 can thus be assigned a road network location 01, 02 on a road segment of the underlying road or navigation map according to its distance from the starting point 11. For example, the sequence of the composite road segments of the segment chain 10 can be traced, starting from the starting point 11, until the longitudinal distance L corresponding to the respective point from the starting point 11 is reached. Thus, by referencing the geoposition P1 of the starting point 11, coordinates in the geocoordinate system can also be assigned to each point in a segment chain 10.

[0072] To illustrate this, Figure 1 shows two geopositions P1 and P2, which were determined by one or more vehicles using a geolocation system and transmitted with an event description. For example, a braking maneuver could have been triggered at each of these geopositions P1 and P2 by a driver assistance system. Due to inaccuracies in the geolocation system, the geopositions P1 and P2 reported by different vehicles typically deviate from the course of the segment chain 10 (that is, the set of road network locations 01 and 02 of this segment chain 10 in the geocoordinate system).

[0073] Each geoposition P1, P2 is assigned a nearest road network location 01, 02 along the segment chain 10. Typically, the nearest location can be uniquely determined by perpendicular projection onto the segment chain 10. If there are multiple nearest road network locations 01, 02 at the same distance, one of them can be randomly selected.

[0074] Figure 1 shows an example of a first geoposition P1, described by a first pair of geocoordinates x1 (along the x-direction x) and y1 (along the y-direction y). A first road network location 01 is assigned to the first geoposition P1 on the segment chain 10, which is located at a first longitudinal distance L1 from the starting point 11. For example, the first longitudinal distance L1 is assumed to be 5.2 length units. The first geoposition P1 can then be described by the identifier "#2" of the segment chain 10 and by the longitudinal distance L1 "5.2 length units" from the starting point 11.

[0075] An event description reported for the first geoposition P1 is assigned in the analysis the first road network location 01 and the lateral distance d1 of the first geoposition P1 from the nearest first road network location 01, which in this case is "1.0 length units". For example, all the information on the identifier of the assigned segment chain 10, as well as the location and lateral distance of the nearest road network location 01, can be summarized in a string "#2, 5.2, +1", which is assigned to the first geoposition P1. In general, this assignment will not be unique; that is, there may be event descriptions at other (different) geopositions to which the same string is assigned.

[0076] Similarly, the second geoposition P2 is described by a second pair of geocoordinates x2 (along the x-direction x) and y2 (along the y-direction y). The second geoposition P2 is assigned the second road network location 02 as the nearest point of the segment chain 10. Furthermore, the second geoposition P2 is assigned a descriptive string "#2, 4.1 , +1" corresponding to the second longitudinal distance L2 of the nearest road network location 02 from the starting point 11 (here "4.1 length units") and the lateral transverse distance d2 between the road network location 02 and the second geoposition P2 (here "1.0 length units").

[0077] With this assignment rule, explained here only as an example, each geoposition P1, P2 can be assigned a point in a generally curvilinear coordinate system, which is given by the set of segment chains 10. Thus, determining the location, with respect to the nearest segment chain 10, can be reduced to specifying the longitudinal distance L relative to its starting point 11.

[0078] By performing all further evaluations per section chain 10, the complexity of the recorded data can thus be reduced from two dimensions (coordinates along the x-direction x and the y-direction y) to only one dimension (longitudinal distance L from the starting point 11 of the respective section chain 10).

[0079] A density distribution is calculated from the event descriptions recorded along a chain of segments 10 by applying established methods known as kernel density estimation (KDE) according to the length intervals L of the road network locations 01, 02 assigned to the respective geopositions P1, P2. For this purpose, a density p of event descriptions is determined for a chain of distances 10 at regular intervals, for example, starting from the starting point 11 at length intervals L of one meter each, which is shown schematically in the upper part of Figure 2.

[0080] The density p is determined by subjecting the longitudinal distances L, L1, L2, which are assigned to the event descriptions by projection onto the nearest section chain 10, to a kernel density estimation using a kernel density estimator with a predetermined bandwidth. Here, a sequence of Dirac pulses arranged at the assigned longitudinal distances L, L1, L2 is convolved with a non-negative, preferably symmetric kernel weighting function.

[0081] A density p can be determined with respect to one or more classes of event descriptions. For example, one class of event descriptions (also called an event class) can be defined as braking interventions by a driver assistance system (e.g., an anti-lock braking system), and another event class as an abruptly reduced vehicle speed.

[0082] The sum of the density values ​​p determined in this way over the segment chain 10 is therefore equal to the number of data points used (that is: the total number of event descriptions of the respective event class(es) recorded over the segment chain 10). This allows for a comparison of different road segments, a comparison of density values ​​p determined based on different data sources, and a weighting of such density values ​​p from different data sources.

[0083] Typically, the density p of event descriptions at certain locations on the road network (01, 02) is considerably higher than the statistical average. Such particularly conspicuous locations are referred to below as peak regions R1, R2, R3. A peak region R1, R2, R3 can be caused, for example, by unclear traffic management, particular weather conditions (such as a tendency towards black ice or fog formation, or ruts that increase the risk of aquaplaning in rain), or by structural characteristics of the road layout (narrowing of the roadway, changes in surface, road damage, or similar).

[0084] To identify peak regions R1, R2, R3, a classification of road network locations 01, 02, or of areas encompassing several such road network locations 01, 02, is carried out. Such a classification is also referred to as scoring.

[0085] The invention is based on the finding that scoring based on a constant threshold e, i.e. the classification of a road network location 01, 02 as a peak region R1, R2, R3 when the density p of event descriptions recorded there (for example, the number of braking events recorded within a radius of one meter around this road network location 01, 02) exceeds the predetermined constant threshold ö (for example, 100 braking events), does not work satisfactorily.

[0086] In particular, such scoring requires that the number of vehicle crossings along the respective road segment is known. However, for the technical and / or organizational-legal reasons already explained, this requirement is hardly achievable in practice. With methods known from the prior art, this leads to heavily trafficked road network locations 01, 02 being marked as conspicuous, even though there is no increased risk there of, for example, intervention by a driver assistance system.

[0087] To circumvent this problem, the invention assumes that the vehicle density along a chain of segments 10, particularly between road network locations 01 and 02 in close proximity, fluctuates only slightly. This assumption is particularly plausible for chain of segments 10 designed as highways or expressways, since for such traffic routes, a change in the number of vehicles passing through is only possible at on- and off-ramps. In particular, this makes scoring values ​​S between adjacent on- and off-ramps readily comparable. It is proposed to determine a scoring value S, referred to as a Z-value or Z-score Z, for each data point. For this purpose, a symmetrical influence area around a road network location, referred to as the bandwidth, is used.

[0088] Location 01, 02 the deviation from a mean value of the density p is determined and related to a standard deviation o- of the density p also determined in this symmetrical influence area.

[0089] To illustrate this procedure, let us assume that along a road section with a number of passing vehicles assumed to be approximately constant regardless of location, at a first to Nth

[0090] Road network location 01, 02 each have a density value ..,pi ..,p lv - determined from event descriptions, each reported within a predetermined range around the road network location 01, 02. From this, an empirical mean (sample mean) is calculated. and an empirical standard deviation βr^ is calculated. In the calculation of unc | ff(Only event descriptions from those road network locations 01, 02 that lie within the predetermined bandwidth around the road network location 01, 02 for which a

[0091] Scoring is to be carried out. In other words: and are expressed as a moving average or moving standard deviation over a range encompassing the bandwidth around the respective density value. certainly.

[0092] It is possible that the moving average p w and / or the sliding

[0093] Standard deviation The densities can be determined by weighted averaging, where densities p measured further away from road network location 01, 02 are weighted less than densities p measured closer to it. For example, a weighting of the adjacent densities p can be applied that decreases linearly with the distance from the respective road network location 01, 02.

[0094] The course of the mean p and the standard deviation a along the chain of sections 10 is shown in the upper part of Figure 2, in addition to the course of the density p. Based on the mean and the standard deviation o'* is assigned to each of the

[0095] Density values ​​p e a normalized density value or Z-value assigned to Z:

[0096] In other words, the respective density value p of event descriptions is subjected to statistical Z-standardization, which is limited to its range of influence (its bandwidth). Within this bandwidth, the respective determined Z-value Z is mean-free and can have a standard deviation u of 1. This makes Z-values ​​Z easily comparable even when they are determined for road sections with different conditions (especially different traffic densities).

[0097] Figure 2 shows the course of the Z-value Z along a segment chain 10 as a function of the segment length L in the middle area.

[0098] Peak regions R1, R2, R3 can be easily identified as statistically significant areas within a segment chain 10 using the Z-value Z. For example, a peak region R1, R2, R3 can be identified as a contiguous area along the segment chain 10 where the Z-value Z exceeds a predetermined threshold of 0.

[0099] As an example, a threshold value 9 = 1 is shown in the middle section of Figure 2. In other words, exceeding this threshold value 0 = 1 indicates that the density value p has deviated from the (moving) mean p by more than the (moving) standard deviation a. However, other threshold values ​​Q can also be used.

[0100] Based on such a determined threshold exceedance, the trend of the Z-value Z can be transformed into a binary scoring value S: for ; > fl otherwise '

[0101] The lower part of Figure 2 shows, as an example (for e = 1), the course of the scoring value S along the section chain 10. Based on the scoring value S, a first to third peak region R1, R2, R3 is determined.

[0102] A peak region R1, R2, R3 is assigned according to the invention:

[0103] - the identifier of the respective section chain 10 (in this case, for example, the string “#2”),

[0104] - the peak start Lmin as the point of the peak region R1, R2, R3 closest to the starting point 11 as well as

[0105] - the peak end Lmax as the point furthest from the starting point 11 in the peak region R1, R2, R3.

[0106] For improved clarity, the peak start Lmin and the peak end Lmax in Figure 2 are only marked for the second peak region R2.

[0107] Furthermore, a peak region R1, R2, R3 can be assigned an explicit length and / or a maximum value and / or a geometric description. The length specifies the distance between the peak start (Lmin) and the peak end (Lmax). The maximum value specifies the highest density p and / or the highest Z-value (Z) in the peak region R1, R2, R3. The geometric description specifies the arrangement and location of the peak region R1, R2, R3 in the geocoordinate system, for example, according to the longitude and latitude of the center point of the peak region R1, R2, R3.

[0108] Figure 3 schematically shows, in a simplified geocoordinate system, the course of a chain of road sections 10 with several road segments connected via intersections, junctions, exits, on-ramps, and similar junctions. On both sides of the chain of road sections 10, on the right-hand side in a direction of travel 11, I2, a Z-curve Z1, Z2 of Z-values ​​Z is indicated by markings M perpendicular to the chain of road sections 10.

[0109] The lengths of the vertical markers M assigned to the first Z-curve Z1 indicate location-dependent Z-values ​​Z, which were recorded corresponding to a first direction of travel 11. Accordingly, the lengths of the vertical markers M assigned to the second Z-curve Z2 indicate Z-values ​​Z, which were recorded by vehicles moving in the opposite second direction of travel 12.

[0110] In addition to the determination of peak regions R1, R2, R3 based on exceeding a threshold value & as explained in Figure 2, a representation according to Figure 3 allows for a particularly clear presentation of route sections with an increased occurrence of event descriptions.

[0111] Figure 4, analogous to Figure 3, shows the course of a chain of sections 10, which is designed as a motorway-like road 100. In this context, a motorway-like road 100 is understood to be a road designed for high-speed traffic, in particular a road with a physical or other separation of traffic in different directions 11, I2. Typically, such motorway-like roads 100 have several parallel lanes in each direction 11, I2, which are not specified in detail in Figure 4.

[0112] Along such motorway-like roads 100, events, such as braking maneuvers or warnings derived from the environmental sensors of vehicles, are recorded at in principle any location on the road network 01, O2, because, for example, overtaking vehicles on an overtaking lane have to brake because of a slower vehicle in front.

[0113] The Z-profiles Z1 and Z2 extending along a highway-like road 100 therefore exhibit different Z-values ​​Z, but these differ less significantly from each other than on roads not designed for high-speed traffic. In particular, it is possible that along a longer stretch of highway-like road 100, no "gaps" are detected in the Z-profiles Z1 and Z2 where the Z-values ​​Z are equal to or very close to zero. This makes the detection of hotspots using simple clustering methods more difficult.

[0114] In particular, the location of hotspots using simple clustering methods is not very reliable if events uncorrelated to the route are recorded with a comparatively high homogeneous (location-independent, the same) frequency and this homogeneous frequency profile is superimposed only by a comparatively small clustering of events that correlates with the route.

[0115] According to the invention, it is therefore proposed to record event descriptions from at least one vehicle, preferably from a fleet of vehicles, and to combine them in an analysis system. From the entirety of the recorded event descriptions, an analysis map designed as a crossing map is created based on the respective assigned geoposition P1, P2, in which the frequencies of events of certain (selectable in the manner of a filter) event classes are assigned to road network locations 01, 02.

[0116] Events from different event classes can be combined into a single density value p, although different event classes can be weighted differently in this common density value p. For example, braking events can be weighted five times more heavily than warnings detected by environmental sensors.

[0117] The density p, which is thus potentially determined jointly for several event classes, is subjected to a kernel density estimation with a kernel density estimator with a predetermined bandwidth in the manner already described and is then transformed into a Z-value Z by means of a local Z-standardization.

[0118] A peak region R1, R2, R3 is then identified as an area in which, along a highway-like road, 100 interconnected road network locations 01, 02 between a peak start Lmin and a peak end Lmax have been assigned a Z-value Z that exceeds a predetermined threshold B. This allows areas with conspicuous traffic situations, such as accident hotspots, to be identified with particularly high reliability.

[0119] A road map marking such peak regions R1, R2, and R3 can then be made available to users. In particular, such road maps can be used to clearly illustrate and assess road safety on large networks of highway-like roads.

Claims

Mercedes-Benz Group AG Patent claims 1. Method for determining at least one safety-critical relevant peak region (R1, R2, R3) using an indicator value (Z) for the statistical characterization of location-specific event descriptions recorded by at least one vehicle along a motorway-like road (100), characterized in that, based on a road or navigation map, those road sections of a motorway-like road (100) without inflows or outflows are identified whose section length is equal to or greater than a predetermined minimum section length. - adjacent identified road sections are connected to form a chain of sections (10) assigned to one of the motorway-like roads (100), with a starting point (11) and an end point (12), - Event descriptions associated with safety-critical events, including a geoposition (P1, P2) and at least one parameter of a vehicle environment model, are recorded and transmitted to an analysis system. - from the analysis system of a geoposition (P1, P2) of an event description, a nearest road network location (01, O2) of a section chain (10) is assigned and described by a longitudinal distance (L, L1 , L2) relative to the starting point (11) of the section chain (10), - for at least one segment chain (10) a progression of a density value (p) of event descriptions related to the longitudinal distance (L, L1 , L2) to the starting point (11) is determined using a kernel density estimator with a predetermined bandwidth, - the course of the density value (p) is transformed with a local Z-standardization based on the bandwidth into a Z-curve (Z1, Z2) of a Z-value (Z) that is mean-free with respect to the bandwidth and - each area of ​​road network locations (01, 02) along a chain of sections (10) between a peak start (Lmin) and a peak end (Lmax) along the motorway-like road (100), where the Z-value (Z) exceeds a predetermined threshold (0), preferably the value 1, is identified as a safety-critical peak region (R1 , R2, R3).

2. Method according to claim 1, characterized in that braking maneuvers and / or warnings are recorded as safety-critical events, wherein a warning is recorded based on a safety-critical relevant parameter of the vehicle environment model.

3. Method according to claim 2, characterized in that Braking maneuvers are weighted more heavily than warnings when determining the density value (p), preferably three to seven times more heavily than warnings.

4. Method according to one of the preceding claims, characterized in that a peak region (R1 , R2, R3) is each assigned an identifier of the respective section chain (10), a longitudinal distance (L, L1 , L2) of the peak start (Lmin) to the starting point (11) of the respective section chain (10) as well as a longitudinal distance (L, L1 , L2) of the peak end (Lmax) to the end point (12) of the respective section chain (10), a maximum value of all Z-values ​​(Z) within the peak region (R1, R2, R3) as well as a description of the geometry of the peak region (R1, R2, R3).

5. Method according to one of the preceding claims, characterized in that the vehicle environment model comprises at least one object and / or parameter detected by means of vehicle sensors and an event description is created based on a trigger and transmitted to the analysis system, where the trigger marks an event identified as worthy of consideration based on the vehicle environment model.

6. Method according to claim 5, characterized in that the trigger marks a warning and / or an intervention of a driver assistance system.

7. Method according to one of the preceding claims, characterized in that at least one event description is assigned an event class selected from a list of event classes and only event descriptions of one or a plurality of event classes are considered for determining the density (p).

8. Method according to one of the preceding claims, characterized in that an event description includes a direction of travel (11, I2) relating to the direction of travel of a segment chain (10) pointing from its starting point (11) to its end point (12) and only event descriptions of the same direction of travel (11, I2) are taken into account for the determination of the density (p).

9. Method according to one of the preceding claims characterized in that Z-values ​​(Z) determined at road network locations (01, 02) along at least one segment chain (10) are visualized by markings (M) running perpendicular to the segment chain (10), the length of which is selected proportionally to the respective Z-value (Z).

10. Method according to one of the preceding claims characterized in that a road map is provided on which at least one safety-critical peak region (R1 , R2, R3) is shown in its location for at least one motorway-like road (100).

Images