EP3014598A1

EP3014598A1 - Method for processing measurement data of a vehicle in order to determine the start of a search for a parking space

Info

Publication number: EP3014598A1
Application number: EP14731561.8A
Authority: EP
Inventors: Heidrun Belzner; Peter PEDRON
Original assignee: Bayerische Motoren Werke AG
Current assignee: Bayerische Motoren Werke AG
Priority date: 2013-06-26
Filing date: 2014-06-04
Publication date: 2016-05-04
Anticipated expiration: 2034-06-04
Also published as: JP2016522526A; US10115309B2; JP6247754B2; EP3014598B1; CN105359200A; WO2014206699A1; US20160210860A1; CN105359200B; DE102013212235A1

Abstract

The invention relates to a method for processing measurement data of a vehicle in order to determine the start of a search for a parking space, comprising the following steps: a) detecting a number (N) of drive data vectors (x_i), wherein each drive data vector (x) comprises information about a speed (v;), position data (p_j), and a point in time (t _i) of the detection of the speed (v) and of the position data (p_i); b) determining an attribute vector (m;) at each point in time (t_i) of the detection of a drive data vector (x_i), wherein the information of the current and prior drive data vectors (x) is processed, wherein the attribute vector (m_i) comprises at least speed information and path information as attribute components; c) classifying each attribute vector (m_i), wherein a first traffic category (c _z) representing a drive of the vehicle or a second traffic category (cp) representing parking search traffic is associated with each of the attribute vectors (m_i), and wherein a probability (p(P | m_i)) is determined, which indicates the probability with which the feature vector is to be associated with the first or the second traffic category (c_z, c_p); d) segmenting over the temporal course of the traffic categories (c_z, c_p) of the attribute vectors (m_i), wherein the drive from the start to the last detection of a drive data vector is divided into two segments in accordance with the determined traffic categories (c_z, c_p) of the attribute vectors (m_i) and the transition from one segment to the other segment represents the start of the search for a parking space.

Description

Method for processing measurement data of a vehicle for determining the beginning of a search for a parking space

The invention relates to a method for processing measurement data of a vehicle for determining the beginning of a parking space search.

Parking information on free parking spaces are used, for example, by parking guidance systems and / or navigation devices for navigating a parking space-seeking vehicle. Modern inner-city systems work according to a simple principle. If the number of parking spaces and the inflow and outflow of the vehicles are known, this can easily determine the availability of free parking. Appropriate signage of the access roads and dynamic updating of the parking place information enable vehicles to navigate to free parking spaces. In principle, this results in restrictions in that the parking areas must be clearly defined and the entry and exit of the vehicles must always be precisely controlled. For this purpose, structural measures, such as barriers or other access control systems are required.

Due to this limitation, navigation is only possible for a small number of free parking spaces. With the necessary structural measures usually only parking garages or fenced parking areas can be integrated into a parking guidance system. However, the much larger number of parking spaces at the roadside or not bounded parking lots is not considered, since the parking situation in public space is largely unknown. Only a few municipalities or traffic management centers provide information for special areas.

In order to search for free parking spaces, an identification of parking spaces along respective streets is desired, especially in inner cities and densely populated areas. From DE 10 2009 028 024 A1 it is known to use parking-exploring vehicles, such as public transport vehicles, such as regular buses or taxis, which have at least one sensor for parking lot recognition. The sensors can be based on optical and / or non-optical sensors. Furthermore, community-based applications are known in which the users of vehicles, for example, enter information into an app when they leave a parking space. This information is then provided to other users of the service. The disadvantage of this is that the information on available parking is only as good as they are provided by the users.

In both described alternatives, there is the problem that the information about the presence of a single parking lot is very fast-moving, i. In areas with a lot of parking search traffic, in which a parking lot information would be helpful, a free parking space is usually occupied in no time.

The Applicant also describes under the application number 10 2012 201 472.1 a method for providing parking information on free parking spaces, in which a knowledge database with historical data is generated from ascertained information about available, free parking spaces. The historical data includes statistical data on free parking spaces for given streets and / or given times or periods. From the historical data and current information, which are determined at a given time for one or more selected streets of vehicles in circulation, a probability distribution of expected free parking spaces for the selected street or streets is determined. The probability distribution represents parking information on free parking spaces in the selected street or streets. The accuracy of the probability distribution depends inter alia on the knowledge of a so-called parking rate λ _ρ . The parking rate will be according to the form! _λ ρ (t) = (1-P _n) λ (ί), where λ (ί) represents a request rate, which for a parking segment, ie, a region considered in which a parking operation is desired, the number of requests for a Parking space per time (ie time unit) indicates. P _n indicates the probability of a free parking space.

The more accurately the parking rate λ _{ρ is} known, the more accurate the probability for a free parking space can be determined.

It is an object of the present invention, based on this method, to provide the applicant with a method which can automatically determine the beginning of a parking space search in order to improve the precision of the determination of the parking rate. This object is achieved by a method according to the features of patent claim 1 and a computer program product according to the features of claim 16. Advantageous embodiments are specified in the dependent claims.

The invention provides a method for processing measurement data of a vehicle for determining the beginning of a search for a parking space. The method described below may be onboard, i. in the vehicle looking for the parking space or offboard, i. be performed by a central computer to which the driving data are transmitted. Further, the proposed method offers the possibility of making the calculations online, i. in real-time while driving, or offline, i. to carry out downstream after the trip.

In a first step, a number of travel data vectors are acquired, wherein each travel data vector comprises information about a speed, position data and the time of the detection of the speed and the position data. The detection of the number of travel data vectors takes place in a predetermined time interval (also referred to below as the sampling rate) in the seconds range, e.g. every second or every five or ten seconds. The trip data vectors thus follow a fixed chronological order. The position data may be represented by GPS (Global Positioning System) data. The position data can be determined by a GPS module of the vehicle. The speed can optionally be determined by the speed sensor of the vehicle or from the position data and detection times of two consecutive measurements.

In a next step, the determination of a feature vector at each time of the detection of a trip data vector, wherein the information of the current and temporally past travel data vectors are processed, wherein the feature vector comprises as feature components at least one speed information and a path information. As a result, the course of the journey of the vehicle is taken into account. In this step, the values of the features are recalculated for each newly acquired trip data vector and combined in a feature vector. To each (Measurement or detection) time is thus a Merkmaisvektor calculated using current and previous trip data vectors are used.

In a next step, the classification of each feature vector vector is performed, wherein each of the feature vectors is assigned to one of two traffic categories. The first traffic category denotes the train traffic, wherein the driver does not search for a parking space, while the second traffic category designates the parking search traffic, wherein the driver searches for a parking space. When determining the traffic category, a probability is calculated which indicates with which probability the feature vector is to be assigned to the first or the second traffic category. In this step, the generated feature vectors are considered individually and classified with respect to two traffic classes, namely a destination traffic represented by the first traffic category and a parking search traffic represented by the second traffic category. At the end of this step, there is a probability for each feature vector which indicates with which probability a feature vector belongs to the parking search traffic and to the destination traffic.

Finally, a segmentation takes place over the time profile of the determined traffic categories of the feature vectors, a subdivision of the journey from the start to the last detection of a travel data vector corresponding to the particular traffic categories of the feature vectors into two segments and the transition from one segment to the other segment the beginning of Parking search represents. The task of the segmentation is to determine, based on the analysis of the time profile of the classification of feature vectors, the travel data vector which marks the start of the search for a parking space. The result of the segmentation is a subdivision of the journey into two segments, according to the traffic categories, which forms the basis for calculating the desired information on the intensity and location of the parking search traffic.

If the beginning of a parking space search is known, it can be used to calculate the probability of an available parking space in the area more accurately. For this purpose, for example, the method described in the introduction can be used in DE 10 20 2 201 472.1 of the Applicant. Furthermore, the knowledge of the beginning of a parking Planners can also be used by city planners to estimate parking situations in individual streets or neighborhoods.

In order to keep the amount of data to be processed as low as possible, it may be expedient to carry out a prefiltering of the travel data vectors. Thus, travel data vectors may be disregarded in determining the beginning of the parking space search if the information about the speed of the trip data vector is greater than a first threshold or less than a second threshold. As a result, e.g. disregard extra-urban journeys and stances of the vehicle. The first threshold may e.g. between 50 km / h and 100 km / h and is in particular 80 km / h. The second threshold may e.g. between 2 km / h and 8 km / h and is in particular 4 km / h.

In another embodiment, for the determination of a respective feature vector, the travel data vectors within a feature window representing a predetermined distance are processed, wherein the feature window propagates the travel data vectors from the current position or measurement to the first position or measurement on the traveled distance includes as the predetermined distance ago. The number of travel data vectors in a feature window can thus vary as a function of sampling rate and speed. For example, if the size of the feature window is 1 km, then with a higher average speed in the last kilometer, fewer trip data vectors will be included in the feature window than at a lower speed if a constant scan rate is assumed.

In a further refinement, the feature vector as feature corn components comprises, in addition to the velocity information and the path information, one or more of the following feature components:

an information about a circularity of the distance traveled. The

Circularity takes into account a typical pattern of behavior in the case of parking-seeking vehicles, the route of which often describes a circular choice of route guidance (for example, by foot traffic). The reference quantity here is the distance of the current position to a center of gravity of the previously detected waypoints, which result from the position data of the respective travel data vectors. an information about a PCA circularity of the distance traveled. Here, the so-called PCA (Principal Component Analysis) is used as a tool to determine the circularity of the routing. If the PCA is applied to the two-dimensional position vectors of a feature window, one obtains, in addition to the two main components which describe the mutually orthogonal axes with the highest variance of the individual waypoints, a relative value for the proportion of the total variance of the axes.

an information about a change of direction. Parking lot-seeking vehicles turn off frequently. Based on the current and a previous position, it is possible to calculate the direction of travel for each trip data vector in the form of an angle (0 ° to 359 °, corresponding to the cardinal directions). To calculate a meaningful value for the change in the course of the journey, the arithmetic mean can be formed over all changes in direction. This is preferably done with normalized values.

information about a target inefficiency. This feature calculates the inefficiency of the route based on the destination of the trip. During the journey, the destination can not be determined on the basis of the trip data, therefore this feature can only be formed after the end of the journey, after all the trip data vectors are known. The destination position assumed is the position of the last trip data vector, which represents the location of the found parking space. This feature component can thus be used only in a method that is performed offline after the completed drive.

According to one embodiment, the speed information may be an arithmetic mean and / or the median of the mean speeds of the travel data vectors taken into account for the determination of a respective feature vector.

According to an embodiment, the route information may be a route inefficiency, which indicates how inefficient the traveled route is by the ratio of the actually traveled route with respect to the shortest route between the positions of two travel data vectors. The inefficiency of the route guidance is a feature that indicates how inefficient the driver-selected, driven route is with respect to the approach to the vehicle. This takes into account the characteristics of the classifiers (traffic classes), since vehicles belonging to the destination traffic try to reach the desired destination As fast as possible and efficient way closer, while parking-looking vehicles have already reached their destination and in the search for a parking lot orbiting.

Here it can be provided to process as path information that path inefficiency for a feature vector which is the maximum for the processed amount of travel data vectors.

In a further embodiment, the feature vectors are normalized to classify each feature vector. Different feature components (in short: Merkmaie) have different value ranges. So that feature components with numerically higher value range do not dominate over feature component with numerically smaller value range and in order to make the feature value comparable, the features are normalized. This has the effect that both features with a large range of values and features with a small range of values reflect the same range of values.

For the calculation of normalized feature components, a z normalization known to the person skilled in the art can be used in which the mean value and the standard deviation are determined for each feature component and the feature components are transformed therewith.

Subsequently, it is expedient to reduce the feature components by a vector projection, in particular by using a principal component analysis (PCA). Principal component analysis is an unsupervised feature reduction technique. It pursues the goal of finding those principal axes in a feature space on which the feature vectors mapped on them reach a maximum variance.

The calculation of the probability of the classifier can then be done with the set of Bayes, which is known to those skilled in the art, e.g. from [1] or [2] is known.

In another embodiment, the beginning of the parking space search is defined by a positive transition of the first traffic category to the second traffic category, wherein the travel data vector, which is assigned to the second traffic category, represents the beginning of the parking space search. In the reverse case, a transfer of the second The first category of traffic is referred to as a negative transition. In the ideal case, a positive transition occurs on a journey at most once. However, the reality shows that several positive transitions can occur during a journey. The beginning of the parking space search can then be determined with the following alternatives:

In a first alternative, the last positive transition of the first classifier to the second classifier is selected as the start of the parking search as long as the classification result of the subsequent travel data vectors constantly comprises the second classifier. After a negative transition, the travel data vector marking the beginning of the parking space search is discarded so that from this point on there is no longer any detected search start.

In a second alternative, the temporally referenced positive transition of the first classifier to the second classifier is selected as the start of the parking space search, as long as the classification result of the subsequent trip data vectors for a given journey path constantly comprises the second traffic category. This segmentation alternative extends the first alternative with a distance criterion. In this case, a determined journey data vector is not immediately forgotten after a negative transition, but maintained for a certain distance after the negative transition. If a further positive transition is found within this route, it is ignored and the previously determined travel data vector is retained. If no positive transition is found, the previous trip data vector marking the beginning of the parking space search is forgotten at the end of the route after the negative transition.

In a third alternative, the beginning of the parking space search is determined on the basis of an integral of the course of the probability over the distance traveled. In this third alternative, not only the hard decision as to whether a feature vector is parking search traffic or not is utilized to determine the start of search, but also the security with which the decision was made. If, if no search start exists, a positive transition is detected with a new trip data vector, the integral of the course of the so-called a posteriori probabilistic chain over the distance traveled is continuously calculated. If the result of the integral calculation is negative, the previously determined travel data vector is discarded. The invention further provides a computer program product which can be loaded directly into the internal memory of a digital computer or computer system, eg a computer of a vehicle or a central computer, and comprises software code sections with which the steps according to one of the preceding claims are executed Product on the computer or computer system is running.

The invention will be explained in more detail with reference to an Ausführungsbeispieis in the drawing. FIG. 1 shows a schematic representation of travel data vectors collected temporally successively; FIG.

FIG. 2 is a schematic representation of a flowchart of the method according to the invention; FIG.

FIG. 3 shows a schematic illustration of display window windows applied to the detected trip data vectors; FIG.

4 shows a representation for a path inefficiency;

Fig. 5 and 6 is an illustration for explaining a circularity;

FIGS. 7 and 8 are diagrams for explaining a PCA circularity;

Fig. 9 is an illustration for explaining an average change in direction;

10 shows a schematic representation of a smoothing of the processed data taken in the context of the method;

11 shows a table with a training matrix; Figures 12, 13 and 14 show a histogram, the kiassendichte and resulting decision limits for determining a probability for the classification of the feature vectors. and

Figs. 15, 16, and 17 show various alternatives for performing a segmentation.

The proposed method, described in detail below, makes it possible to determine the parking search amount of a car trip in order to determine information about a car park search that has been made, e.g. the time that has effectively elapsed to find a parking space, or the distance traveled while searching for a parking space, or the location or area in which a parking space was searched. The method makes it possible in particular to determine the beginning of the search for a parking space for a car.

The method may be performed by a computing unit in the vehicle (i.e., onboard) or by a central off-vehicle computing unit (i.e., offboard). Starting point of the method are so-called travel data vectors (i = 1... N) of a drive of the vehicle. The trip data vectors x, e. determined by the vehicle at predetermined measuring times and processed sequentially in several steps. If the method is carried out offboard, the journey data vectors are preferably transmitted in real time via a communication interface to the central computer.

A trip is represented by a payable set of N trip data vectors [, x ₂ ; x _N ], where

This is exemplified in Fig. 1 dargesteiit. A trip data vector x, (Eq. 3.1) consists in each case of information on speed v, and GPS position p, at the time ti of the detection of the travel data vector. The trip data vectors follow a fixed time order since t, <tj + i. The GPS postions i can be detected by a navigation system installed in or incorporated in the vehicle. The speed is eg detected by a sensor of the vehicle and is typically available in the vehicle in a computing unit or on a data bus.

It is assumed that a driver of the vehicle makes the decision to search for a parking space only once during a journey. Occasionally, a driver may begin to search in one area and, after a certain period of time, stop the search and continue in another area. In this case, the last decision to search for a parking space is taken as a true search start. In addition, it is assumed that each trip ends in a public parking lot at the roadside.

According to these assumptions, each trip has exactly one true time T _par k, from which a parking space is sought. If a parking space is found directly after the decision to search for a parking space, T _park = T _enC i _e . On the basis of this time, a journey into two segments can be subdivided according to the type of driving type Cj (so-called traffic categories or traffic classes): The first segment since the start of the journey always represents the so-called "destination traffic" ZV, while the second segment represents the so-called Parking search "PSV corresponds. Ais target traffic ZV one TEII driving is referred to, in which the driver of a start point T _star t the way to the region goes, where a parking lot is being sought. During the destination traffic ZV the driver does not look for a parking space.

The assignment of the travel data vectors to the respective traffic class is carried out by a class label c ,. For the trip data vectors of a trip:

Ci = 0; for i = 1; ...; i _park- i -> destination traffic

Cj = 1; for i = i _par k; N - Park search traffic (3.2) i _park is the first index from which the trip data vector Xi belongs to the parking search traffic and thus represents the beginning of the parking _search . The true time T _pai1t and the corresponding location of the start of a journey can be determined by ty », * and Pi_p ₃ rk can be approximated in Xu »*.

Are i _park and the _trip data _vectors x _jpark ; x _N known, search duration T _park and search _distance S _pafk can be approximated as follows: Tpark - i ^j V tipark 3) where δ (.,.) represents the distance between two GPS positions on the earth's surface in meters. Alternatively, a distance function may be used which calculates the shortest path between two points with respect to a correct navigation map.

The location of the search for a parking space can either be indicated directly by the GPS positions of the search route, by map matching of the positions on roads, or indirectly by an indication of the so-called search center of gravity and average search radius of the search route. The basis for calculating these values is _p3rk . The determination of i _park , and thus the beginning of the parking space search, is the goal of the method described in more detail below.

2 shows the procedure of the method according to the invention in a flow chart.

At the beginning, non-relevant trip data vectors are sorted out in the course of an optional prefiltering (step S1). Subsequently feature vectors are generated on the basis of known travel data vectors (step S2) and optionally smoothed (step S3). The classification (step S4) calculates for each feature vector a probability for the class affiliation to the traffic class parking search traffic. The subsequent segmentation (step S5) analyzes the classification history and determines the final class labels by the determined search start of the journey. At the end of the journey, the determined results are optionally pi-sensitized (step S6).

These steps are described in detail below.

Those trip data vectors x _{i (} which play no role in the determination of the search start on the basis of specified criteria are identified and rejected in step S1 of the prefiltering.) In this context, discarding means that the relevant trip data vector j is not sent to the next step for further processing noticeably malsextraktion, is passed on. These include, for example, driving outside of city traffic and stance phases.

A roadside parking lot is typically searched for in city traffic. The upper limit for the permitted speed is 80 km / h on inner-city roads. Since at a higher speed it is no longer possible to proceed from parking search traffic, a travel data vector x _a with eg v,> 80 km / h is sorted out. This limit can also be set lower or higher.

While a vehicle is stationary (for example at a traffic light or in a traffic jam), there are no changes in speed or position changes. The recorded journey data vectors x _s during stance phases contain, with the exception of the tally stamp, the same information. Since information about stance phases is not relevant for any of the subsequent steps, trip data vectors Xj with eg Vi <4 km / h are sorted out. The choice of this welding value is due to the fact that in this way parking operations are not recorded, the speed is typically between 0-4 km / h.

In the following step S2, the feature extraction takes place. For the identification of parking lot-seeking vehicles, a low average speed, frequent turns and Blockumfahrten can be used. In order to obtain information about these characteristics of the journey, a single trip data vector with its information about the current speed and position is not sufficient, but its course must be taken into account.

The progression of the signal values of individual trip data forms the basis of the extraction of features presented in this section. In this case, the values of the features are recalculated for each newly entering trip data vector and combined in a feature vector m. At each time t _t , a feature corn vector is calculated with the following feature components: .Mittler (k ^ i mdigkdi

Krcisfönnigkdt

III

It is sufficient if the average speed and the path inefficiency are taken into account as feature components (also referred to below as features). By taking into account further feature components, the accuracy of the determination of the beginning of the parking space search can still be improved, the accuracy increasing only to a small extent. The current and previous trip data vector are used to calculate the various characteristics. The travel data vectors to be taken into account for calculating features are determined on the basis of a feature window MF, which is shown in greater detail in FIG.

The size l _{f of} the feature window MF is based on the distance traveled, since most of the constructed feature analyzes the course of the route. If the feature window was based on the past time, the length of the route section in a feature window F _{S would} vary depending on the speed, and a minimum length of the route section would not be guaranteed. However, this is necessary in order to be able to compare the calculated characteristics in the course of the journey.

The feature window MF- _a includes the travel data vectors from the current position x, to the first position, which lies farther back than 1 _f on the traveled distance. The number of travel data vectors in a feature window can thus vary as a function of sampling rate and speed. For example, if the size of the beacon window is 1 km, then at a higher average speed in the last kilometer, fewer trip data vectors are included in the feature window than at a lower one, assuming a constant sample rate. Feature vectors m, can only be calculated from a distance covered by l _f since the beginning of the journey in order to ensure the verglatsbarkeit between the calculated Merkmaisvektoren m. For the rest of the description, the travel data vectors within a feature window anchored at Xj will be x _f1 ; χ _β ^, where x _{f is} the oldest and x ^ the most recent trip data vector. Similarly, Xj =

XFM.

The characteristics calculated from travel data (feature components) are described in more detail below.

Average speed

For the calculation of the average velocity v, the arithmetic mean over all velocity values in the feature window is not formed, but the median. The reason for this is its robustness against outliers.

Vi = median {v _t "v / ₂ ,. ·., Vf _M }

By virtue of the method step of pre-filtering (step S1) of the travel data vectors, this value represents the mean velocity in driving phases.

Off-inefficiency

The inefficiency of the route guidance η is a feature which indicates how inefficient the driver-selected, driven route is with respect to the approach to the destination. The idea for this is derived from the characteristics of the traffic classes, since vehicles belonging to the destination traffic try to get closer to the desired destination in the quickest and most efficient way, whereas parking-seeking vehicles have already reached their destination and in the search for one Park the car park.

If a route through the waypoints pi p ₂ ; p «3 given the starting position p _r and the end position p _K , two distance measures can be calculated, which form the basis for the calculation of this feature. This is illustrated by way of illustration in FIG. 4. s _d is the shortest distance between pi and p, the airline being used in this description. s _z represents the distance covered between ρ ·, and p _K. This corresponds to the length of the selected route from to P _K. It is s _z > s _d . The relationship of the two routes to each other indicates whether the selected route represents a direct route to the end position (efficient) or a detour (inefficient). A value for the inefficiency of the Wegführung can through

Te ([pi, jf |) = 1 -,; // € [0, 1], k € [l i ~ 1]

(3.7). The traveled distance s _z is approximated by the sum of all partial distances between the individual waypoints. The index k indicates soft waypoint in the set [pi; ...; p "J should serve as a starting point for calculating the inefficiency. A value of η «-► 0 indicates an efficient route, while η _κ -> 1 means an inefficient route.

The waypoints of a feature window [p _f ; P _f a; · ■; Pm] available. The goal in the calculation of the feature is to determine the highest inefficiency between the current position p _m and all remaining positions in the feature window:

In this way, circles and reversals, which are contained in the course of several consecutive feature windows, similarly affect the feature value.

circularity

Since a typical behavior pattern in the case of parking-seeking vehicles describes a circular choice of the route guidance (for example by block drives), this feature is intended to provide the circularity κ of the route within the feature area. ters. The reference quantity here is the distance s _{m of} the current position p _M to the center of gravity of the waypoints p _f . If s _m ~ l _f / 2, then we assume a straight line (Fig. 5). The smaller this distance becomes, the more circular the routing becomes (FIG. 6).

The center of gravity of the route is calculated by the arithmetic mean over the individual components of the positions in the characteristics window:

M

(3

The value for the circularity is calculated by:

(3.10)

Here, the distance between the center of gravity and the current position is normalized by the effective size of the feature window to obtain a value between 0 and 1. In order to be able to assume a straightforward route guidance for K -> 0 and a circular route guidance for κ-► 1, the normalized term is additionally subtracted from 1.

PCA Kreisförmicjkeit

Another way to determine the circularity of the Wegführung used the PCA (Principal Component Analysis, which is described for example in [1]) as an aid. Applying the PCA to the two-dimensional position vectors of a feature window gives, in addition to the two major components describing the mutually orthogonal axes with the highest variance of the individual waypoints, a relative value for the proportion of the total variance of the axes described by and λ ₂ . λι corresponds to the relative variance component of the axis with the highest variance, which is why λι £ λ ₂ applies. If the examined route runs on a straight line, the total variance of the waypoints is distributed only on the axis described by the first main component (FIG. 7). On the axis of the second main component accounts for only a small proportion of the total variance. Describe the waypoints a completely circular routing, the proportion of the second main component of the total variance increases, so that λ-s = λ ₂ (Fig. 8).

To calculate the PCA circularity p, the PCA is applied to the position information within a feature window. Subsequently, the quotient is formed from the resulting scales and λ ₂ :

Due to the restriction λι £ λ ₂ , the value of p moves between 0 and 1, where p _f - 0 means a straight line and p _f -> 1 means a circular path.

change of direction

Parking-looking vehicles often turn off. On the basis of the current and a previous position, it is possible to calculate for each trip data vector x, the direction of travel Φί in the form of an angle (0 ° to 359 °, corresponding to the cardinal directions). With the help of Φ can be a value for the change of direction _Δ φ normalized are calculated on the distance _d s between two waypoints:

(3.12) with

A _f i = min {| <Pi - i_i |, 360 ° - \ ι - <_ij} In order to calculate a meaningful value for the change of direction during the journey, the arithmetic mean Δ ^. over all standardized direction changes

^ ο formed Λ i ^m feature window:

Fig. 9 shows directions φ- _ιΑ and at the corresponding positions p _M and their difference.

Target inefficiency

This feature calculates the inefficiency of the routing relative to the destination of the ride. During the journey, the destination can not be determined on the basis of the journey data, therefore this feature can not be formed until after the end of the journey (ie offline), after all travel data vectors are known. The destination position assumed is the position p _{N of} the last trip data vector, which represents the location of the found parking space.

The inefficiency of the routing relative to the destination ζ is calculated for each trip data vector as follows (see equation 3.7):

Since it may happen, for example, in the case of courier journeys that the start and end positions of a journey lie in the vicinity, a maximum target inefficiency is already present at the beginning of the journey. This can be circumvented by taking the waypoint furthest away from the target p _{/ rf} = argmax _p . {^ (., ρ ^)} ^ is determined from the drawing position and the values of the feature for i <i _d are set equal to zero 0:

In the optional smoothing step (step S3), the feature vectors of a journey are smoothed. The purpose of smoothing is to combine feature vectors on a particular link to a smooth feature vector. In this way, not individual waypoints are processed, but sections. The generation of smoothed feature vectors is done by combining several feature vectors m, which are within a smoothing window GMF. The smoothing window GMF is advanced with respect to the traveled distance and may overlap. This is shown in FIG. 11.

The length of a respective smoothing window GMF is determined by! _gf determined. It is anchored to the first feature vector m _{g1 of} a route section. The feature vector at the end of the route segment m _gR is the last _{traffic vector vector that} is less than Ig _t away from m _g1 with respect to the distance traveled. The number of feature vectors rrti within a smoothing window GMF can vary analogously to the number of travel data vectors Xj within a feature window MF. In order to overlap the smoothing windows MGF, a new smoothing window MGF can be anchored within the current smoothing window MGF after exceeding a certain distance l _gr . So that no more than two smoothing windows overlap simultaneously in order to limit the complexity of this step, the same applies

Each smoothed feature vector generated by the feature vectors in a smoothing window MGF characterizes a link segment of length _lgr> which starts at the position of the first enclosing feature vector and ends at the position of the first feature vector of the next smoothing window. The corresponding stretch of the last smoothed feature vector of a ride may be shorter or longer. In this way, it is ensured that all sections of a journey represented by smoothed feature vectors are disjoint, whereby the smoothing windows themselves do not necessarily have to be disjoint. As a result, in the calculation of a smoothed feature vector m ₈ , which characterizes a specific route section, feature vectors of the following route section can also be included. To prevent this, i _gf = l _gr can be selected.

From the R feature vectors in smoothing windows [m _g1 ; ...; m _gR ], the individual components of the smoothed feature vector are calculated as follows: m _g = [median {v _{gi i} ..., v _gR }, max A _wt "., ih _w }], ιϊι = πι \ ϊ7

The mean velocity corresponds to the median of the mean velocity of all feature vectors, while the maximum of all other features is determined.

If the label c for belonging to a traffic class is additionally known for the feature vectors in the smoothing window, the median of all lables is determined for the corresponding label in the smoothed value. This corresponds to a majority decision, with the same number of votes for parking search traffic is decided. In the special case l _gf = l _gr = 0, the smoothing has no effect: The smoothed feature vectors then correspond to the original feature vectors.

In the step of classification (step S4), the generated feature vectors are considered individually and classified with respect to the traffic classes destination traffic Z and parking search traffic P. At the end of this step, there is a probability p (P | mj) for each feature vector m, which indicates with which probability a feature vector belongs to the parking search traffic.

To calculate this probability, the feature vector vectors are first normalized, reduced, and then classified. For all these sub-steps, training data in the form of feature vectors is required in order to be able to learn the parameters for the individual sub-steps. This method uses supervised learning methods. Therefore, the class affiliation of the individual feature vectors, in Form of true labeis c, to be known. This is made possible by the use of test drives recorded for learning purposes, where the traffic class is known at all times.

The training data are in the form of an N × K matrix T, each row representing a feature, and each column representing a feature vector is shown in FIG. Figure 11. 11 shows a training matrix T. The times of the matrix represent the different features, while the columns represent their occurrences in a feature vector. According to the class membership of the feature vectors, the training data in T can be divided into two matrices T _z and T _P.

Different characteristics have different value ranges. So that features with a numerically higher value range do not dominate over features with a numerically smaller value range and in order to make the characteristic values more comparable, the features are normalized. This has the effect that both features with a large range of values and features with a small value range reflect the same range of values.

Normalized feature values are calculated using z-normalization known to those skilled in the art. Here, on the basis of training data in T for each Merkmai m is the mean _μ η _n and the standard deviation _σ η determined.

To calculate the entries m of the normalized training matrix f, each entry of the training matrix is transformed using the calculated parameters: (3.20)

One column of the resulting thus contains a normalized feature vector m.

The background of the feature reduction is the reduction of the feature components in a feature vector with minimal information loss. The number of features in m is reduced from N to 1 .s D <N. Thus, a vector projection □ ^N → ^D □ ^{D is} performed. The reduced feature vector m is calculated using the A / x D transformation matrix W: lii = Will (3 21)

The preferred technique for feature reduction is Principal Component Analysis (PCA), which involves reduction of N - »D. The PCA is an unsupervised method of feature reduction. It pursues the goal of finding those principal axes in the feature space on which the feature vectors depicted on them reach a maximum variance.

The basis for calculating the transformation matrix is the A / x N covariance matrix Σ of the training matrix T, consisting of the entries ο%.

Subsequently, the eigenvectors and eigenvalues of the covariance matrix are calculated, as described eg in [3]. The eigenvectors w ₍ represent axes in the feature space, while the eigenvalues λ, indicate the relative proportion of the total variance of the feature vectors projected on the resulting eigenvectors.) W _! Corresponds to the eigenvector with the largest eigenvalue A _{1 (} while w _{N is} the eigenvector with the smallest Eigenvalue A _N. If the eigenvectors are known, an arbitrary 1 s D <N can now be selected, which determines the dimension of the transformed features. indicated. The D lines of the transformation matrix are then combined with the first D eigenvectors [wt; ...; w _D ] filled.

The transformation of a feature vector m into the reduced feature space is applied by ιΐι = W (m- μ) (3.24) where μ represents the mean value vector with the mean values of the individual features [U; μ _Ν ] represents. If the feature vectors are already averaged by a previous normalization (μ = 0), the transformation can also be performed using the rule in Equation 3.21.

The classification assigns a probability to each (reduced) feature vector. Based on this probability, it is possible to make a statement about the class membership c of a Merkmaisvektor. Here, c _z denotes the membership of the class "destination traffic", while c _{P represents} the affiliation to "parking search traffic".

To calculate the likelihood of a feature vector to belong to the traffic class parking search traffic class, also referred to as a posteriori probability, the known set of Bayes is applied, e.g. in [1] or [2].

(3.27) (m | c) is the class-specific density function, which gives the probability for a

Feature vector indicates to belong to class c. p (c) is called the a-priori probability and represents the probability with which class c occurs. Finally, p (m) indicates the probability of the occurrence of a feature vector without distinguishing between classes. It can be calculated by summing up all class-specific probabilities multiplied by the

Probability of occurrence of the corresponding class, to be calculated.

The density functions or probabilities required for calculating the a posteriori probability can be estimated from the training data in T, or T _z and T _P :

In order to be able to estimate the class-specific density functions, it is assumed that the individual components of the feature vectors are within the different

Classes are normally distributed. On the basis of this assumption, the values for i? (M | c) are to be calculated on the basis of the density function of the normal distribution, which is defined by the parameters mean μ and covariance matrix Σ.

μ is calculated according to the mean value in the normalization step and Σ according to the covariance matrix of the PCA. The training data divided into classes in T _z and T _P serve as the data basis for the calculation of the parameters of the class-specific density functions. Consequently, ώ | ί _ζ ) = Κ ( _ζ , Σ _ζ ) and (ώ | _ζ ) = (μ _/) , Σ _ρ ).

To estimate the a-priori probability of the different classes, the number of feature vectors in the training data is used. In this case, N indicates the number of feature vectors in T, and N ₂ and N _P the number of feature vector vectors in the class-specific training matrices T _z and T _P. p {cz) - p {cp) = obf p (c) = 1

< ^? € { ₍ ^ ρ} ^ 3 29)

With the aid of the a-posteriori probabilistic chain, a statement can now be made about the classification of a Merkmai vector, since p (c ₂ | m) = 1 - p {c _p | ih). The used

Kiassifikator is a maximum a-postori Klassifikaior. This means that a feature vector is classified on the basis of the greatest a posteriori probability:

CM AP = max / p (c ImAH} =) < ^CP > I Ϊ

(3.30)

The result of the classification applies beyond the feature vector also for the underlying trip data vector,

The course of the decision function in feature space is marked by the set M of those points in the feature space which are on the decision boundary:

The course of the decision function, which comes about through the parametric classification presented here, is quadratic due to the choice of different covahance matrices.

The position of the decision boundary is affected by the a priori probabilities: the smaller the a priori probability of a class, the further the decision boundary shifts in the direction of the corresponding class, by the adaptation of the number of feature vectors of each class Thus, the result of the classification can be influenced. Figures 12 to 14 illustrate the construction of the decision boundary with the aid of the class-specific training data in the one-dimensional feature space. Those graphic elements constructed using the feature vectors in T ₂ are designated 10, 12, 14, while the elements labeled 11, 13, 15 have been constructed from the feature vectors in T _P. The decision limit is indicated in FIG. 14 by GR.

The task of the segmentation is to determine, based on the analysis of the temporal course of the classification of feature vectors, that travel data vector which marks the start of the parking space search. The result of the segmentation is a subdivision of the journey into two segments, according to the traffic classes, which forms the basis for calculating the desired information on the intensity and location of the parking search traffic.

Considering the time course of the classification result C _M AP, it is assumed that a transition of the classification result of c _z - ^■ c _{P represents} the beginning of the parking space search. Such a transition is called a positive transition, while the opposite case c _P -> c _{z is} called a negative transition.

In the Idealfail, a positive transition occurs on a journey at most once. In reality (see Figures 15 to 17), however, shows that several positive transitions can occur during a trip,

If there is no positive transition during the entire journey, the last trip data vector is accepted as the start of the parking space search. This ensures that a value> 0 for parking search distance and parking search duration can be calculated.

In the following, three methods will be described with reference to FIGS. 15 to 17, which determine at most one travel data vector x ₊ with a positive transition of the classification result as the start of the parking space search. A driving data vector x_ represents a driving data vector with a negative transition of the classification result. The so-called simple segmentation method (FIG. 15) decides for the (temporally) last positive transition, as long as the classification result of the subsequent travel data vectors remains constant c _p . After a negative transition, x _{+ is} discarded so that from this point on there is no longer any recorded search start. This means that at no time does this method capture park search with c = c _z .

The segmentation with distance criterion (FIG. 16) extends the simple segmentation method with a distance criterion. In this case, a determined travel data vector for x _{+ is} not immediately forgotten after a negative transition, but maintained for a certain distance l _s after the negative transition. If another positive transition is found within this route, it will be ignored and retained x *. If no positive transition is found, then x _{+ is} forgotten at the end of the route after the negative transition. If l _s = 0, this segmentation method achieves the same result as simple segmentation.

In addition to the information for classifying a feature vector c p, the segmentation with integral criterion shown in FIG. 17 also uses the a posteriori

Probability p {c _p | m). Thus, not only the hard decision as to whether a feature vector is parking search traffic or not is utilized to determine the start of search, but also the security with which the decision was made.

If, if no search start exists, a positive transition is detected with a new travel data vector x ₊ , until the next negative transition x found is found. continuously the integral l _{+ of} the course of the a posteriori probability over the distance s from x ₊ to x. calculated.

In this case, 0.5 is subtracted from the a posteriori probability in order to obtain positive values for c = c _P and negative values for c = c _z . This decision boundary is marked EGR in FIGS. 15 to 17. Is integrated over a route with only negative course of this modified value of the a posteriori probability receives thus, a negative term. Moreover, this subtraction term ensures that route courses represent the same absolute integral value with the same certainty of the classification but different classification result.

For the following travel data vectors classified with c = c _z , the negative integral value I is continuously calculated until either I.> 1+ or a new positive transition is found. If I.> l _+> , the current search beginning is forgotten, while in another positive transition, the positive integrage! with l ₊ = i ₊ + I. recalculated.

This means that a sufficiently strong train traffic behavior, which follows a stretch of road with traffic search behavior, can revise the search start. On the other hand, the integral criterion ensures that a low on-track behavior over a longer distance can not cancel the current search start.

Since the a posteriori probability sequence does not follow an analytically calculable function, and moreover, there is no continuous value curve, the integral in Equation 3.32 must be given for a segment represented by the positions [p ^ p ₂ ; p _N j and the associated a posteriori probabilities [p _ap i; p _aP 2; _8Ρ ] can be numerically approximated:

The segmentation step provides a result regarding the start of the search for a parking space. The result is not necessarily true, because it is based on the result of the classification. Again, the classification is based on a probabilistic model, which was first used with the help of training data.

In the optional Piausibiiisierungsschritt (step S6), the result of the segmentation is assessed and discarded if necessary. This means that this step allows for the possibility of leaving journeys unobserved regarding the search for a parking space. The criteria for withholding the segmentation result are eg an implausible long park search route. It seems implausible after the assumed journey that almost one entire journey serves the search for a parking space. Since it may be that the search for a parking space takes longer due to possible obstacles, this criterion is measured by the distance traveled in the destination traffic and in the parking search traffic. It is believed that the result implausibe! if the distance s _p traveled to the destination since the start of the search _{for a} parking space is greater than half of the distance traveled since the start of the journey until the start of search s _z :

(3.34)

Sources

[1] E. Aipaydin, Introduction to Machine Learning (Adaptive Computation and Machine Learning), The MIT Press, 2004.

[2] C.M. Bishop, Pattern Recognitton and Machine Learning (Information Science and Statistics), Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2006.

[3] G. Fischer, Linear Algebra, Vieweg Studies: Maths Basic Course, Vieweg,

2005

LIST OF REFERENCE NUMBERS

Xi trip data vector (i = 1 ... N)

c _f class label / traffic lane

i no. of a measurement

N number

mi feature vector

c _z first classifier

c _P second classifier

p probability

Tpark time, from the parking is sought

ZV destination traffic

PSV parking search traffic

MFj feature window

Ii size of the feature window

GMF smoothing window

l _gf length of the smoothing window

EGR decision limit

GR decision limit

Claims

claims

A method of processing vehicle measurement data to determine the beginning of a parking space search, comprising the steps of:

a) detecting a number (N) of travel data vectors (Xj), wherein each travel data vector (Xj) information about a speed (vi), position data (p) and a time (t,) of the detection of the speed (v _ä ) and the Position data (Pi) comprises;

b) Determining a erkmafsvektors (mi) at each time (t,) of the detection of a trip data vector (x,), wherein the information of the current and past time travel data ektoren (x,) are processed, wherein the feature vector (m _ä ) as Feature components at least one speed information and a path information includes;

c) classifying each feature vector (m,), wherein each of the feature vectors (nrii) of a first traffic category (c _z ) representing a drive of the vehicle, or a second traffic category (c _P ) representing a parking seeker traffic, and being a probability

(p {P jm,)) is determined, which indicates with which probability the characteristic corn vector of the first or the second traffic category (c _z , c _P ) is to be assigned;

d) segmentation over the time course of the determined traffic categories (c _z , c _P ) of the feature vectors (m,), wherein a subdivision of the journey from the start to the last detection of a driving data vector corresponding to the particular traffic categories (c _z , c _P ) of the feature vectors (m,) takes place in two segments and the transition from one segment to the other segment represents the beginning of the parking space search.

2. The method of claim 1, wherein in the travel data vectors (x,) in the determination of the beginning of the parking space search are disregarded when the information about the speed (v,) of the travel data vector (x,) greater than a first Schweilwert or less than a second threshold.

3. The method of claim 1 or 2, wherein for the determination of a respective feature vector (m,) the travel data vectors (x,) are processed within a feature window (l _f ) representing a predetermined distance, wherein the Merkmaisfenster (l _f ) includes the travel data vectors (x,) from the current measurement to the first measurement that is more than the predetermined distance on the covered distance.

4. Method according to one of the preceding claims, in which the feature vector (m,) as feature corn components comprises, in addition to the speed information and the route information, one or more of the following feature components:

an information about a circularity of the distance covered,

information about PCA circularity of the distance covered,

- information about a change of direction,

- information about a target inefficiency.

5. Method according to one of the preceding claims, in which the speed information is an arithmetic mean and / or the median of the average speeds of the travel data vectors (Xj) taken into account for the determination of a respective feature vector (rrt |).

6. The method of claim 1, wherein the distance information is a path inefficiency indicative of the ratio of the actually traveled distance with respect to the shortest distance between the positions of two travel data vectors (x,), how inefficient the driven one Track is.

7. The method according to any one of the preceding claims, is processed as the path information that Weg- tnefflzienz for a feature vector (m,), which is the maximum for the processed amount of travel data vectors (χ,).

8. The method according to any one of the preceding claims, wherein for the classification of each feature vector (irij) the Merkmaisvektoren (m,) are normalized.

9. The method of claim 8, wherein for the calculation of normalized feature components a z-Normaläsierung is used in which for each feature component of the mean value and the standard deviation are determined and transformed with these the Merkmaiskomponenten.

10. The method of claim 9, wherein the feature components are reduced by a vector projection, in particular by applying a principal component analysis. Method according to one of the preceding claims, wherein the calculation of the probability of the classifier takes place with the set of Bayes.

12. The method according to any one of the preceding claims, wherein the beginning of the parking space search by a positive transition of the first traffic category (c _z ) to the second traffic category (c _P ) is defined, wherein the trip data vector (x,), the second traffic category ( c _P ), represents the beginning of the parking space search.

13. The method of claim 12, is selected as the beginning of the parking space search in terms of time last positive transition of the first traffic category (c ₂ ) to the second traffic category (c _P ), as long as the classification result of the subsequent trip data vectors (x,) constant the second traffic category (c _P ).

14. The method of claim 12, is selected as the beginning of the parking space search in terms of time last positive transition of the first traffic category (cz) to the second traffic category (c _P ) as long as the Kiassifikationsergebnis the subsequent trip data vectors (x _f ) for a given Ride (l ₃ ) constantly the second traffic category (c _P ) includes.

15. The method of claim 12, wherein the beginning of the parking space search is determined based on an integral of the course of the probability over the distance traveled.

16. Computer program product directly into the internal memory of a digital

Computers can be loaded and includes software code sections, with which the steps are carried out according to one of the preceding claims, when the product is running on a computer.