Ai training method, training data generation process, and prediction method and system

Abstract

The invention relates to an AI training method (T), to a training data generation process (E), and to a prediction method (V) and system (1) for predicting the counting accuracy of a passenger counting system (10). During the training process of the AI, training data sets are provided (S1), each of which contains counting data (Z), ascertained using a passenger counting system (10), and condition data (U) which characterizes the conditions of the process of generating the counting data (Z). The counting data (Z) characterizes the number of passengers which have boarded a vehicle (100) and which have deboarded the vehicle (100). Additionally, the counting accuracy for each training data set is provided (S2) so that the training data sets with the respective corresponding counting accuracy can be presented (S5) to the AI.

Description

[0001] Description

[0002] AI training methods, training data generation, and prediction methods and systems

[0003] The present invention relates to a method for training an artificial intelligence, a method for generating training data sets and a method for predicting a counting accuracy of a passenger counting system and a corresponding system.

[0004] Passenger counting systems are known for automatically collecting passenger numbers in local and long-distance public transport, for example, in rail vehicles. Based on the collected passenger numbers, operators can, for example, optimize the use of their vehicles. If several such operators have joined together to form a transport association, it is also possible to allocate revenues among the operators based on passenger numbers.

[0005] When the determined passenger numbers are used in this way, great importance is usually placed on the reliability and precision of the corresponding passenger counting systems. Corresponding requirements are often laid down in contracts or standards, which must be verified, for example, during commissioning and / or ongoing operation of a passenger counting system.

[0006] It is an object of the present invention to improve the determination of a counting accuracy of a passenger counting system, in particular to make it easier, less complex and / or to enable it to be determined exclusively on the basis of data provided by the passenger counting system.

[0007] This object is achieved by a method for training an artificial intelligence, a method for generating training data sets and a method for predicting a counting accuracy of a passenger counting system and a corresponding system according to the independent claims.

[0008] Preferred embodiments are the subject of the dependent claims and the following description.

[0009] In a method, in particular a computer-implemented method, according to a first aspect of the invention, an artificial intelligence is trained in such a way that - with the aid of the artificial intelligence - a counting accuracy of a passenger counting system of a vehicle, in particular a rail vehicle, can be determined. Training data sets are provided which each contain counting data determined with the aid of a passenger counting system as well as circumstance data which characterize the circumstances of the generation of the counting data. The counting data characterize a number of passengers who have boarded and alighted from a vehicle. In addition, a counting accuracy is provided for each of the training data sets, so that the training data sets can be presented to the artificial intelligence with the respective corresponding counting accuracy.

[0010] A presentation of data within the meaning of the invention is preferably a "feeding" of an artificial intelligence with the data. For example, during the presentation, the data serves as input for an algorithm which, when processing the data, learns to recognize patterns in the data. During the presentation, for example, data is passed through a neural network or clustered by a k-means algorithm.

[0011] One aspect of the invention is based on the approach of predicting the counting accuracy of a passenger counting system of a vehicle based on counting data from this passenger counting system. For this purpose, a number of passengers boarding the vehicle, as counted by the passenger counting system, and a number of passengers disembarking from the vehicle, as counted by the passenger counting system, are linked to the circumstances under which the passengers are counted. An appropriately trained artificial intelligence can use this link to make robust predictions about the counting accuracy, even when data is sparse. In contrast to conventional methods, which require manual counting of the actual passenger volume and subsequent comparison with the counts of the automatic passenger counting system to determine the counting accuracy, the appropriately trained artificial intelligence does not require manual counting to estimate the counting accuracy.Rather, the trained artificial intelligence may be intrinsically capable of making a statement about the counting accuracy based on its counting when completing at least one journey chain.

[0012] A journey chain is preferably the journey of the vehicle along a specified route from a starting point, such as a departure station, to an end point, such as a destination station, with several stops in between where passengers can board and / or disembark. In other words, a journey chain is a chain of journeys along a route. The journey chain is, for example, the chain of journeys between the stops of a bus or train line.

[0013] Preferred embodiments of the invention and their further developments are described below. These embodiments can be combined with each other and with the aspects of the invention described below, unless expressly excluded.

[0014] In a preferred embodiment, the counting accuracy is determined based on a comparative count. A comparative count is preferably a manual or automated count of boarding and / or disembarking passengers, which takes place parallel to the (automated) counting by the passenger counting system. In this respect, a comparative count is preferably a count of passengers that is carried out independently of and in addition to the counting by the passenger counting system. By comparing the count by the passenger counting system with the comparative count, the counting accuracy of the passenger counting system can be determined objectively, reliably, and precisely.

[0015] This comparative count is preferably carried out over a specified minimum period. Considering a minimum period, for example, one or two weeks, allows for the collection of sufficient statistics. Thus, the counting accuracy can also be determined using counting data from the passenger counting system itself, for example, based on the difference between the number of boarding passengers collected over a period of one to two weeks and the number of alighting passengers collected during the same period.

[0016] Alternatively or additionally, sufficient statistics can also be collected by performing the comparison count for a predetermined minimum number of vehicle journey chains.

[0017] In another preferred embodiment, the counting data characterizes the number of passengers boarding and disembarking during a journey chain. In other words, each training data set can be based on (exactly) one journey chain. This allows sufficient statistics to be obtained for each training data set. One requirement for this may be that a comparative count—for example, a manual one—is available for each training data set, for example, the journey chain.

[0018] In a preferred embodiment, a difference between the number of passengers boarding and alighting determined using the passenger counting system is determined and linked to the circumstance data and / or the counting accuracy. For example, the difference between the number of passengers boarding and alighting, along with the circumstance data, is presented to the artificial intelligence as input, while the corresponding counting accuracy is presented as output. Considering the difference between the number of passengers boarding and alighting, in conjunction with the recording circumstances, has proven to be a robust parameter for AI-supported determination of counting accuracy.

[0019] In a further preferred embodiment, the circumstance data contains information on at least one of the following circumstances: i) stop, ii) location, iii) status of a passenger counting system sensor, iv) platform height, v) weather, vi) time of day, and / or vii) illuminance. It has been shown that these circumstances can have a particularly significant influence on the reliability of the passenger counting system and thus also on its counting accuracy.

[0020] For example, at a particular stop, particularly at a particular train station, the ground conditions may impair the detection of passengers by sensors positioned above the vehicle doors (and therefore also detecting the ground as a background), and thus the counting of passengers. If such ground conditions are limited locally, for example at a construction site, passenger detection may be impaired by only one of several sensors, i.e., at a particular location. Similarly, detection reliability may vary with different platform heights, since the passengers to be detected are then at different distances from the respective detecting sensor.

[0021] Passenger counting may also be affected by the operating status of one or more of the sensors and / or the lighting conditions, which may be caused by weather, time of day and / or illuminance, i.e. brightness.

[0022] Alternatively or additionally, the circumstance data may contain a value for the detection probability of a passenger counting system sensor, in particular if the passenger counting system sensor is a so-called "smart sensor". Such sensors provide, for example, a probability for the correct detection of the object per detection (e.g. per counting event) and classification (e.g. per passenger category, i.e. child or adult). Such a detection probability, already determined by the sensor, can have a significant influence on the counting accuracy of the passenger counting system.

[0023] In a further preferred embodiment, a correlation measure for a correlation between the counting data, the circumstance data and / or the provided counting accuracies is determined. In this respect, the training data sets can be pre-processed in order to make the training of the artificial intelligence more efficient. In particular, it is checked how well the counting data, for example the differences between the number of passengers boarding and alighting, correlate with the provided counting accuracies. It is expedient to determine which circumstances characterized by the circumstance data exist in the event of a deviating correlation, i.e., in the event of a low correlation measure. The training data sets pre-processed in this way can then be presented to the artificial intelligence.

[0024] To prepare the artificial intelligence more specifically for prediction, the training data sets can be filtered based on the correlation measure. For example, those counting data, circumstantial data, and / or counting accuracies are presented to the artificial intelligence for which the determined correlation measure fulfills a predefined correlation condition.

[0025] A second aspect of the invention relates to a method, in particular a computer-implemented method, for generating a training data set for training an artificial intelligence. The artificial intelligence can preferably be trained in such a way, in particular by a method according to the first aspect of the invention, that a counting accuracy of a passenger counting system for a vehicle, in particular a rail vehicle, can be determined. The second aspect of the invention can relate in particular to a method for generating a training data set for use in the method according to the first aspect of the invention.

[0026] To generate the training dataset, passengers boarding and disembarking a vehicle are counted using a passenger counting system over a specified minimum period, and the circumstances under which the boarding and disembarking passengers are counted are recorded using a recording device. The counted boarding and disembarking passengers are then linked to the recorded circumstances.

[0027] In a method, in particular a computer-implemented method according to a third aspect of the invention for predicting a counting accuracy of a passenger counting system of a vehicle, in particular a rail vehicle, input data is provided which contains counting data determined with the aid of a passenger counting system and circumstance data. The counting data characterizes a number of passengers who have boarded and alighted from a vehicle. The circumstance data characterizes the circumstances of the generation of the counting data. A counting accuracy of the passenger counting system is predicted on the basis of the input data with the aid of an artificial intelligence system, trained in particular by the method according to the first aspect of the invention. With this method, robust predictions about the counting accuracy can be made even with sparse data.

[0028] A system according to a fourth aspect of the invention for predicting the counting accuracy of a passenger counting system of a vehicle, in particular a rail vehicle, has an interface via which input data can be provided. The input data contain counting data determined with the aid of a passenger counting system, which characterizes a number of passengers boarding and disembarking from a vehicle, as well as circumstance data characterizing the circumstances of the generation of the counting data. A prediction device is also provided, which is configured to carry out the method according to the third aspect of the invention.

[0029] A prediction device within the meaning of the invention can be designed in hardware and / or software. The prediction device can in particular have a processing unit, preferably one that is data- or signal-connected to a memory and / or bus system. For example, the prediction device can have a microprocessor unit (CPU) or a module thereof and / or one or more programs or program modules. The prediction device can be designed to process instructions that are implemented as a program stored in a memory system, to detect input signals from a data bus and / or to output output signals to a data bus. A memory system can have one or more, in particular different, storage media, in particular optical, magnetic, solid-state and / or other non-volatile media.The program can be designed in such a way that it embodies or is capable of carrying out the methods described here, so that the means can carry out the steps of such methods and thus in particular can predict a counting accuracy of a passenger counting system.

[0030] The properties, features and advantages of the invention described above, as well as the manner in which they are achieved, are explained in more detail in conjunction with the figures in the following description of exemplary embodiments of the invention. Where appropriate, the same reference numerals are used in the figures for the same or corresponding elements of the invention. The exemplary embodiments serve to explain the invention and do not limit the invention to the combinations of features specified therein, including with regard to functional features. Furthermore, all features specified in the exemplary embodiments can be considered in isolation and combined as appropriate with the features of any claim.

[0031] They show, at least partly schematically:

[0032] FIG 1 shows an example of a system for predicting the counting accuracy of a passenger counting system; and

[0033] FIG 2 shows an example of a method for training an artificial intelligence.

[0034] FIG. 1 shows an example of a system 1 for predicting the counting accuracy of a passenger counting system 10 of a vehicle 100. The system 1 has interfaces 2a, 2b via which data can be provided to a prediction device 4. The prediction device 4 is connected to a memory device 3 for data or signals.

[0035] The interfaces 2a, 2b, in particular, allow the provision of input data that can be processed by the prediction device 4 to determine the counting accuracy. This input data contains counting data Z, which characterizes a number of passengers boarding and disembarking from the vehicle 100. The input data also contains circumstance data U, which characterizes the circumstances of the generation of the counting data Z, in particular the passenger count.

[0036] The counting data Z are preferably generated by the passenger counting system 10 and provided via the interface 2a. The counting data Z are expediently based on a detection of passengers boarding and alighting by one or more passenger counting system sensors 11. These sensors 11 are arranged, for example, in the area of ​​doors of the vehicle 100. Such a sensor 11 preferably has an optical sensor, for example a 3D camera for generating images and / or a time-of-light camera for generating point clouds. With the aid of algorithms known per se, objects can be recognized in this sensor data and identified passengers can be counted.

[0037] The circumstance data U are preferably generated and / or read out by a recording device 12 and provided via the interface 2b. The circumstance data U are expediently based on a recording of ambient or environmental conditions and / or properties of the passenger counting system 10. The circumstance data U characterize, for example, i) the stop, ii) the position of the vehicle 100, in particular of the respective counting sensor 11, iii) the height of the platform edge at a station, iv) the weather or v) the time at which the boarding and alighting passengers are counted. Alternatively or additionally, the circumstance data U can also characterize vi) a status of the passenger counting system 10, in particular of the respective passenger counting system sensor 11, and / or vii) a detection probability of the respective sensor 11 when detecting the passengers boarding and disembarking.

[0038] This information on the circumstances can, for example, be retrieved from corresponding databases or received from corresponding services using a land-based detection device 12a, i.e., one not assigned to the vehicle 100. Alternatively or additionally, this information can be read out or determined by a vehicle-based detection device 12b. It is also conceivable that this information is provided at least partially by the passenger counting system 10 itself. In this case, part of the circumstance data U could also be provided via the interface 2a.

[0039] The prediction device 4 expediently comprises a data processing device for processing the counting data Z and the circumstance data U. For example, the prediction device 4 comprises an artificial intelligence that is trained to evaluate the counting data Z and circumstance data U. The artificial intelligence is trained, in particular, to predict the counting accuracy of the passenger counting system 10 based on the counting data Z and the circumstance data U.

[0040] To operate the artificial intelligence, the prediction device 4 can have a processing unit, for example a microprocessor unit. The processing unit is expediently connected to the storage device 3 for data or communication purposes. The storage device 3 can have one or more, in particular different, storage media, in particular optical, magnetic, solid-state and / or other non-volatile media. A program corresponding to the artificial intelligence is expediently stored on the storage device 3, which program embodies or is capable of carrying out a method V for predicting the counting accuracy of the passenger counting system 10, so that the prediction device 4 can carry out steps of this method and can thus in particular predict a counting accuracy of the passenger counting system 10.

[0041] In method V for predicting the counting accuracy of the passenger counting system 10, the artificial intelligence of the prediction device 4 evaluates the provided counting data Z and circumstance data U. The predicted counting accuracy can then be output as output A, for example, via a corresponding interface to a user or another hardware and / or software module.

[0042] The infrastructure provided by the system 1 can also be used in a method E for generating a training data set for training the artificial intelligence. For example, the counting data Z and circumstance data U that are associated with the same journey chain are linked together in such a data set. It is expedient if, in addition, an (independent) comparative count of the passengers boarding and alighting during this journey chain is carried out in order to determine the actual counting accuracy of the passenger counting system 10. The counting data Z and the circumstance data U as well as this actual counting accuracy can then form the basis of a method for training the artificial intelligence, as described in connection with FIG 2.

[0043] FIG 2 shows an example of a method T for training an artificial intelligence in such a way that with the aid of the artificial intelligence a counting accuracy of a passenger counting system of a vehicle , in particular a rail vehicle , can be determined .

[0044] In a method step S1, training data sets are provided. Such training data sets include, for example, counting data generated by a passenger counting system, for which (independent) comparative counts have also been conducted. It is expedient for the counting data to have been generated during a certification of the passenger counting system, during which the counting by the passenger counting system has been verified by an independent count.

[0045] The counting data characterize, for example, the number of passengers boarding and disembarking as counted by the certified passenger counting system.

[0046] The training data sets expediently also contain circumstantial data characterizing the circumstances under which the counting data was generated. The circumstantial data includes, for example, information on the time and / or location of the counting, the weather, the status of the sensors of the passenger counting system, and / or the like.

[0047] The training data sets are preferably each formed from counting data and circumstance data that were generated and / or read out at the trip level. For example, each training data set comprises counting data and circumstance data that were generated, recorded and / or read out during a trip chain. In a further method step S2, a counting accuracy is provided for each of the training data sets. The counting accuracy preferably indicates the accuracy and / or reliability with which the number of passengers boarding and alighting characterized by the counting data was counted by the passenger counting system. In this respect, it is expedient that the counting accuracy for each of the data sets is determined and provided on the basis of the corresponding comparison count.

[0048] In process step S2, the training data sets are preferably also prepared or preprocessed. For example, for each set of counting data, such as for each trip chain, a difference between the counted number of passengers boarding and disembarking can be determined. This difference can form a solid basis for a subsequent estimation of the counting accuracy of the passenger counting system.

[0049] For practical purposes, the training data sets, i.e., the respective counting and circumstance data, are also linked to the provided counting accuracies in process step S2. Based on this link, the artificial intelligence can learn which combination of counting and circumstance data (also referred to as input) results in which counting accuracy (also referred to as output).

[0050] In an optional further method step S3, a correlation measure for a correlation between the counting data, the circumstantial data, and / or the provided counting accuracies is determined. In particular, the extent to which the counting data correlate with a specific counting accuracy and the extent to which the circumstantial data play a role in this are examined. The result of the examination is expediently represented by the correlation measure.

[0051] For example, it is examined under which circumstances the difference between the number of passengers boarding and alighting, possibly in relation to the total number of passengers recorded, at least approximately reflects the corresponding counting accuracy.

[0052] In this respect, it is also possible to examine which circumstantial data or circumstances characterized thereby are conspicuous in the case of a deviating correlation between the difference and the counting accuracy. The correlation measure can be used as the basis for this examination. If the determined correlation measure reaches or falls below a predetermined correlation threshold, deviations between the circumstantial data and other, particularly predetermined, circumstantial data that characterize circumstances in a situation in which the correlation measure exceeds the correlation threshold can be determined.

[0053] On the basis of the determined correlation between the counting data, the circumstance data and / or the counting accuracies, the data with which the artificial intelligence is to be trained are selected in a further optional method step S4. Preferably, the counting data, circumstance data and counting accuracies for which the determined correlation measure fulfills a predetermined correlation condition are selected. For example, the training data sets are selected in which the counting data and circumstance data correlate maximally with the counting accuracy. In particular, the training data sets can be selected in which the correlation of the difference between the number of passengers boarding and alighting with the counting accuracy reaches or exceeds a predetermined correlation threshold and the circumstances determined in method step S3 are assessed as unremarkable.

[0054] As part of this process step, the training data sets can also be processed depending on whether the specified correlation condition is met. For example, circumstantial data characterizing circumstances that have proven to be insignificant with regard to a correlation with counting accuracy can be removed from the data sets. This "slimming down" of the data can significantly increase the efficiency of training the artificial intelligence.

[0055] In a further method step S5, the training data sets, optionally modified as described above, and the associated counting accuracies are presented to an artificial intelligence system. For example, this data is presented to a k-means algorithm, a genetic algorithm, a neural network, or the like. This trains the artificial intelligence system to predict a counting accuracy with high precision and reliability based on counting data and circumstantial data that are generated, recorded, and / or read out, for example, during a journey chain.

[0056] In an optional further procedural step S6, the artificial intelligence training can be evaluated. For example, it can be checked which key performance indicators (KPIs) can be achieved with the training model. This makes it possible, for example, to check the accuracy with which the trained artificial intelligence makes predictions. The evaluation of the training, in particular the key performance indicators, can, if necessary, be used as the basis for a certification of the artificial intelligence or presented as proof of quality.

[0057] In one example, a comparative count on a bus, perhaps by an external service provider, might show: 50 passengers boarded, 49 people alighted (because, for example, one person stays seated with the bus driver until the depot). If the (automated) passenger counting system counts 50 passengers boarding and 49 passengers alighting, the counting accuracy is 100%.

[0058] If the counting accuracy is based on a difference between the number of passengers boarding and disembarking, the same result is obtained. However, based on the usual assumption that the total number of boarding passengers equals the total number of disembarking passengers, a counting accuracy of only approximately 98% could have been estimated (a deviation of one passenger in 50 counting events).

[0059] The artificial intelligence trained according to procedure V can determine the fact "the total number of boarding does not necessarily have to equal the total number of alighting" by linking it with the circumstantial data. In this example, this circumstantial data could be as follows: i) historical data for the route (or similar routes) shows that at the relevant time of day fewer passengers are alighting than have previously boarded (people who remain seated until the end of service); and / or ii) messages from the passenger counting sensor (e.g. error messages from the sensor, messages about the probability of certain counting events and / or the like) do not indicate a counting error.

[0060] Although the invention has been illustrated and described in detail by means of the preferred embodiments, the invention is not limited by the disclosed examples and other variations can be derived therefrom by those skilled in the art without departing from the scope of the invention.

Claims

Patent claims 1. Method (T) for training an artificial intelligence such that the counting accuracy of a passenger counting system (10) of a vehicle (100), in particular a rail vehicle, can be determined, wherein - Training data sets are provided (Sl) , each containing count data (Z) determined using a passenger counting system (10) which characterize a number of passengers boarding and alighting from a vehicle (100), as well as circumstance data (U) which characterize the circumstances of the generation of the count data (Z), - a counting accuracy is provided for each of the training datasets (S2) , and - an artificial intelligence is presented with the training data sets with the corresponding counting accuracy (S5) .

2. Method (T) according to claim 1, wherein the counting accuracy is determined on the basis of a comparative count over a predetermined minimum period.

3. Method (T) according to claim 1 or 2, wherein the counting accuracy is determined on the basis of a comparative count at a predetermined minimum number of journey chains of the vehicle (100) .

4. Method (T) according to one of the preceding claims, wherein the counting data (Z) characterize the number of passengers boarding and alighting during a journey chain.

5. Method (T) according to one of the preceding claims, wherein a difference between the number of passengers boarding and alighting determined using the passenger counting system (10) is determined and linked with the circumstance data (U) and the counting accuracy.

6. Method (T) according to one of the preceding claims, wherein the circumstance data (U) contains information on at least one of the following circumstances: stop, location, status of a passenger counting system sensor (11), platform height, weather, time and / or illuminance.

7. Method (T) according to one of the preceding claims, wherein the circumstance data (U) includes a value for the detection probability of a passenger counting system sensor (11).

8. Method (T) according to one of the preceding claims, wherein a correlation measure for a correlation between the count data (Z), the circumstance data (U) and / or the provided count accuracies is determined (S3) and used as the basis for the presentation.

9. Method (T) according to claim 8, wherein those count data (Z), circumstance data (U) and count accuracies are presented to the artificial intelligence for which the determined correlation measure satisfies a predetermined correlation condition.

10. Method (E) for generating a training data set for training an artificial intelligence, in particular using the method (T) according to one of the preceding claims, such that the counting accuracy of a passenger counting system (10) of a vehicle (100), in particular a rail vehicle, can be determined, wherein - Passengers entering and exiting a vehicle (100) are counted using a passenger counting system (10) over a specified minimum period of time, - circumstances under which the boarding and alighting passengers are counted, are recorded using a recording device (12), and - the counted passengers boarding and alighting are linked to the recorded circumstances.

11. Method (V) for predicting the counting accuracy of a passenger counting system (10) of a vehicle (100), in particular a rail vehicle, wherein - Input data are provided which includes count data (Z) determined using a passenger counting system (10) which characterizes the number of passengers who entered and exited a vehicle (100), as well as circumstance data (U) which characterize the circumstances of the generation of the count data (Z), - the counting accuracy of the passenger counting system (10) is predicted on the basis of the input data using an artificial intelligence, in particular trained according to one of claims 1 to 9.

12. System (1) for predicting the counting accuracy of a passenger counting system (10) of a vehicle (100), in particular a rail vehicle, with - an interface (2a, 2b) via which input data can be provided, which includes count data (Z) determined using a passenger counting system (10), which characterizes the number of passengers who entered and exited a vehicle (100), as well as circumstance data (U) which characterize the circumstances of the generation of the count data (Z), - a storage device (3) on which a computer program product is stored that causes a computer to perform the steps of the method (V) according to claim 11; and - a prediction device (4) which is set up to execute the computer program product stored in the storage device (3).

Images