Method for controlling scene detection using a device

Abstract

Embodiments of the present disclosure relate to control scene detection. An embodiment provides a method for controlling a device to perform scene detection from a set of possible reference scenes. The method includes detecting a scene from the set of possible reference scenes at consecutive detection instances using at least one classification algorithm. An initial confidence probability is assigned to each newly currently detected scene. The initial confidence probability is updated according to a first transition probability from a previously detected scene to the newly currently detected scene. Based on at least the updated confidence probability associated with each newly currently detected scene, a filtering operation is performed on these currently detected scenes. The output of the filtering operation continuously delivers the filtered detected scenes.

Description

[0001] Cross - reference to related applications

[0002] This application claims priority to French Patent Application No. 1852333, filed on Mar. 19, 2018, which is incorporated herein by reference. Technical Field

[0003] Embodiments provide a method for controlling scene detection by a device. Background Art

[0004] Broadly speaking, a scene is understood to particularly include the scene characteristics of the environment in which a device is located, whether the device is carried by a potentially mobile user, such as a cellular mobile phone (scenes such as "bus", "train", "restaurant", "office", etc.) or the device is a fixed object that is connected or not connected (such as a radiator in a home automation application), and the scene characteristics of this environment can be of types such as, for example, "wet part", "dry part", "day", "night", "blinds closed", "blinds open", etc.

[0005] A scene can also include the scene characteristics of activities performed by the carrier of the device, such as a smartwatch, and then such a scene can be "walking", "running", etc.

[0006] Certain wireless communication devices, such as, for example, certain types of smartphones or tablets, are now able to perform scene detection, thus making it possible to determine the environment in which the user of the phone or tablet is located. Therefore, this can allow third parties (such as advertisers or cultural institutions) to deliver relevant information related to the location of the user of the device.

[0007] Thus, for example, if the user is located at a given tourist location, it is possible to deliver to him the addresses of restaurants near his location. Similarly, it is also possible to deliver to him information related to certain monuments located near his area.

[0008] "Scene detection" is understood to particularly refer to the discrimination of the scene in which a wireless communication device is located. There are several known solutions for detecting (discriminating) scenes. These solutions use, for example, one or more dedicated sensors that are usually associated with specific algorithms.

[0009] In this particular algorithm, mention can be made of a classifier or a decision tree, which is well-known to those skilled in the art in the context of scene detection. In particular, mention can be made of neural network algorithms known to those skilled in the art, and for such neural network algorithms, reference can be made (for example, for all useful purposes) to the work entitled Neural Network Design (2nd Edition, September 1, 2014) by Martin T Hagan and Howard B Demuth, or to an algorithm known as GMM ("Gaussian Mixture Model") to those skilled in the art, who can refer (for example, for all useful purposes) to the tutorial slides entitled "Clustering with Gaussian Mixtures" by Andrew Moore, which can be found on the website https: / / www.autonlab.org / tutorials / gmm.html.

[0010] These two algorithms are also configured to deliver a confidence probability for each detected scene.

[0011] As a classification algorithm, mention can also be made of a meta-classification algorithm or "meta-classifier", that is to say an algorithm located at a higher level than the layer containing multiple classification algorithms.

[0012] Each classification algorithm provides a decision regarding the detected scene, and this meta-classifier compiles this decision provided by the various classification algorithms in order to deliver a final decision, for example, by majority voting or by average voting.

[0013] Meta-classifiers perform better than traditional classifiers, yet are still subject to possible detection errors. In addition, they are more complex to implement, especially in terms of memory size, and require a more complex learning phase. Summary of the Invention

[0014] Embodiments and example modes of the present invention relate to the real-time detection of scenes of a device, the device being, in particular but not limited to, a wireless communication device, such as a smartphone or a digital tablet equipped with at least one sensor such as an accelerometer, and more particularly to an improvement in the reliability of the decision provided by a classification algorithm before filtering it.

[0015] Embodiments of the present invention can further improve the reliability of the decision provided by a classification algorithm, regardless of its type, and achieve this in an easily implementable manner.

[0016] For example, one embodiment provides a method for a device to perform control scene detection from a set of possible reference scenes. The method includes using at least one classification algorithm to detect scenes from the set of possible reference scenes at consecutive detection instances. Each newly currently detected scene is assigned an initial confidence probability. The initial confidence probability is updated according to a first transition probability from a previously detected scene to the newly currently detected scene. Based on at least the updated confidence probability associated with each newly currently detected scene, a filtering operation is performed on these currently detected scenes. The output of the filtering operation continuously delivers the filtered detected scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Other advantages and features of the present invention will become apparent by examining the various embodiments and examples and the fully non-limiting detailed description of the drawings, in which:

[0018] Figures 1 to 5 Schematically illustrate various embodiments and examples of the present invention. DETAILED DESCRIPTION

[0019] The inventors have observed that for users of a device, certain transitions between a first scene and a second scene are unlikely or even impossible.

[0020] According to one implementation and embodiment, it is proposed to utilize the habits of the user and more specifically the transition probabilities between consecutively detected scenes for the purpose of improving scene detection.

[0021] According to one aspect, a method for a device to perform control scene detection from a set of possible reference scenes is proposed.

[0022] An embodiment provides a method that includes using at least one classification algorithm to detect scenes from the set of possible reference scenes at consecutive detection instances. Each newly currently detected scene is assigned an initial confidence probability. The method further includes updating the initial confidence probability according to a first transition probability from a previously detected scene to the newly currently detected scene, and a filtering operation on these currently detected scenes based on at least the updated confidence probability associated with each newly currently detected scene, the output of the filtering operation continuously delivering the filtered detected scenes.

[0023] Advantageously, the user experience is a parameter to be considered in scene detection. Specifically, the first transition probability makes it possible to quantify the probability that a user transitions from one scene to another.

[0024] Certain transitions from one scene to another are unlikely or even impossible.

[0025] Thus, by taking this parameter into account, the initial confidence probability associated with the currently detected scenario is improved.

[0026] Furthermore, during the filtering operation, for example, the meta-filter, the updated confidence probability associated with the newly currently detected scenario makes it possible to reduce or even eliminate the noise that disturbs the meta-filter before any other decision, thus enabling the meta-filter to have better performance.

[0027] According to one embodiment, the update of the confidence probability associated with the newly currently detected scenario may include multiplying the initial confidence probability by a first transition probability.

[0028] In other words, the initial confidence probability associated with the newly currently detected scenario is weighted by the first transition probability.

[0029] According to one embodiment, the method further includes normalizing the updated confidence probability associated with the newly currently detected scenario.

[0030] According to one embodiment, each transition from a first scenario in the set of possible reference scenarios to a second scenario in the set of possible reference scenarios is assigned a first transition probability having an arbitrary value or an updated value.

[0031] Preferably, the first transition probability is updated and this is done during the total lifetime of the user equipment. This makes it possible to adapt to changes in the experience of the user who owns the equipment.

[0032] According to one embodiment, if at least one transition from the first scenario to the second scenario is performed during a given time interval, the first transition probability is updated, and this update includes calculating a second transition probability for each transition from the first scenario to each possible second scenario in the set of possible reference scenarios.

[0033] The "given time interval" should be understood to refer to, for example, a period corresponding to N transitions observed, for example, on a non-sliding window.

[0034] For example, the number N is defined according to a corpus of reference scenarios.

[0035] For example, it can be done by a counter that counts the transitions up to N transitions.

[0036] According to one embodiment, if the second transition probability is higher than the first transition probability, the update of the first transition probability may include increasing its value by a first set value.

[0037] The first set value can be, for example, between 0.01 and 0.05.

[0038] According to one embodiment, if the second transition probability is lower than the first transition probability, the update of the first transition probability includes decreasing its value by a second set value.

[0039] The second set value can be, for example, between 0.01 and 0.05.

[0040] According to one embodiment, using a differentiable optimization algorithm, each first transition probability that is liable to be updated is updated.

[0041] The classification algorithm can deliver an initial confidence probability associated with the newly currently detected scene.

[0042] If the classification algorithm does not provide an initial confidence probability associated with the newly currently detected scene, it is possible to assign a confidence probability with an arbitrary value to the newly currently detected scene.

[0043] If the value of the initial confidence probability is not available, it is considered that its value is set to 100%. In this case, the weighting of the first transition probability is performed on the probability with a value set to 100%.

[0044] The processing operation of updating the initial confidence probability is compatible with any type of classification algorithm, but is also compatible with any type of filtering processing operation performed based on the probability associated with the scene detected by the classification algorithm.

[0045] When it is stipulated to assign an identifier to each reference scene, the processing operation of updating the initial confidence probability is particularly compatible with the filtering processing operation on the currently detected scene, which is implemented not only based on the updated confidence probability associated with each newly currently detected scene, but also based on the identifier of each newly currently detected scene (each detected scene actually has an identifier because each detected scene is one of each reference scene with an identifier).

[0046] For example, this type of filtering processing operation is described in French Patent Application No. 1754080 filed on May 10, 2017 (U.S. Application No. 15 / 924,608 of the same family filed on March 19, 2018). Therefore, a "meta-filter" that makes it possible to utilize the temporal correlation between two consecutive decisions is proposed in the French patent application. The meta-filter acts on the consecutive decisions delivered by the classification algorithm, regardless of its type, and includes a meta-classifier. These applications are incorporated herein by reference.

[0047] Some of its features will be reviewed here.

[0048] The filtering operation is a sliding time filtering operation on the currently detected scenes on a filtering window of size M, based on the identifier of each newly currently detected scene considered in the window and the updated confidence probability associated with this newly currently detected scene.

[0049] The value of M defines the size of the filter and its latency (on the order of M / 2) and contributes to its accuracy. A person skilled in the art will know how to determine this value according to the target application and the desired performance.

[0050] That being said, a value of M = 15 may be a good compromise.

[0051] This filtering operation is compatible with any type of classification algorithm.

[0052] For a classification algorithm configured to deliver a single detected scene at each detection moment, provisions can be made for a register of size M (1xM), such as a shift register, in order to form a window of size M.

[0053] Thus, in particular, it can be provided that for a storage circuit including a shift register of size M that forms a window of size M, and for each newly currently detected scene, its identifier and the associated updated confidence probability (possibly normalized) are stored in the register, and using the M identifiers and their associated updated confidence probabilities (possibly normalized) that appear in the register, the filtering operation is performed, and one of the possible scenes (i.e., one of the reference scenes) is delivered as the filtered detected scene.

[0054] When the content of the shift register is transferred at each newly currently detected scene, the identifier extracted from it is the earliest identifier in terms of time. However, of course, the identifier that will be delivered after filtering can be any identifier from the register, for example, the identifier of the just detected scene or the identifier of the previous scene.

[0055] That being said, the classification algorithm can be configured to deliver a set of several scenes at each detection moment. For example, this is the case for a meta-classifier, in which case, at each detection moment, the various detected scenes will be used by the various decision trees that make up the meta-classifier, without using a final step such as majority voting.

[0056] This can also be the case for a neural network that uses various scenes and their corresponding confidence probabilities, which are respectively associated with the various neurons of the output layer.

[0057] In the case where a classification algorithm (e.g., a meta-classifier or a neural network) is configured to deliver a set of D scenes at each detection moment, where D is greater than 1, in order to form a filtering window of size M, a larger-sized memory can be provided that is capable of storing a matrix of size DxM (D is the number of rows and M is the number of columns).

[0058] Then, updated confidence probabilities (which may be normalized) associated with the identifiers of these D currently detected scenes can be stored in the storage circuit for forming a window of size M for each new set of the D currently detected scenes. Using the DxM identifiers and their associated updated confidence probabilities (which may be normalized), a filtering processing operation is performed, and one of the possible scenes (i.e., one in the reference scene) is delivered as the filtered detected scene.

[0059] In particular, when the storage circuit includes a shift register of size M, the filtering processing operation can include the definition of an integer J of an entire portion greater than or equal to 2 and less than or equal to M / 2, and for each identifier present in the register constructed from 2J identical identifiers. The filtering processing operation can also include comparing this constructed identifier with the 2J constructed identifiers, and in the case of a non-identical value, including replacing the constructed identifier with one of the constructed identifiers, and for the replacement of the constructed identifier with the constructed identifier, also including the assignment of an updated confidence probability (which may be normalized) calculated based on the confidence probabilities (which may be normalized) of the updated 2J constructed identifiers, for example, the average of the confidence probabilities (which may be normalized) of the 2J constructed identifiers.

[0060] This makes it possible to eliminate, for example, isolated detection errors.

[0061] When the storage circuit is suitable for a classification algorithm that is configured to continuously deliver multiple sets of D detected scenes, the filtering processing operation just outlined above, especially for eliminating isolated errors, is then advantageously applied row by row to the matrix DxM.

[0062] According to a compatible implementation, regardless of the configuration of the classification algorithm and thus regardless of the size of the storage circuit, for each identifier considered more than once in the storage circuit, the filtering processing operation includes: summing the updated confidence probabilities (which may be normalized) associated with it, and the filtered detected scene then becomes a scene whose identifier considered in the storage circuit has the highest total updated confidence probability (which may be normalized).

[0063] According to a possibly more complex variant, which is also compatible, independent of the size of the storage circuit, and makes it possible, in particular, to provide an indication of the variability of the filter, for each identifier considered more than once in the storage circuit, the filtering processing operation consists of, for the sum of the updated confidence probabilities (possibly normalized) associated with it, the formula for the probability density function of the identifier, which is centered on the identifier having the highest total updated confidence probability (possibly normalized), the calculation of the variance of this function, the calculation of the ratio of this highest total confidence probability (possibly normalized) and this variance, the comparison of this ratio with a threshold, and the selection of the filtered detected scenario according to the result of this comparison.

[0064] Thus, this variant makes it possible to determine the confidence of the filter and thus make a decision.

[0065] Thus, if the ratio is below the threshold that may be caused by a large variance, the confidence is low, and then it can be decided to deliver the scenario that actually has the highest total updated confidence probability (except that an information is assigned to it to describe it as "uncertain"), or to deliver the previously filtered detected scenario in terms of time, as the detected scenario at the output of the filter.

[0066] Conversely, if the ratio is higher than or equal to the threshold, which may be caused by a small variance, the confidence is high, and then it can be decided to deliver the scenario that actually has the highest total updated confidence probability (possibly normalized), as the detected scenario at the output of the filter.

[0067] According to another aspect, a device is also proposed that includes a detector configured to detect a scenario from a set of possible reference scenarios at successive detection moments, the detection using a classification algorithm, and each newly currently detected scenario being assigned an initial confidence probability. A processor configured to update the initial confidence probability according to a first transition probability from a previously detected scenario to the newly currently detected scenario. A filter configured to perform a filtering processing operation based on at least the updated confidence probability associated with each newly currently detected scenario and continuously deliver the filtered detected scenario.

[0068] According to one embodiment, the processor is configured to update the initial confidence probability associated with the newly currently detected scenario by multiplying the initial confidence probability by the first transition probability.

[0069] According to one embodiment, the processor is configured to normalize the updated confidence probability associated with the newly currently detected scenario.

[0070] According to one embodiment, each transition from a first scenario in the set of possible reference scenarios to a second scenario in the set of possible reference scenarios is assigned a first transition probability having an arbitrary or updated value.

[0071] According to one embodiment, the processor is configured to update the first transition probability if at least one transition from the first scenario to the second scenario is performed during a given time interval, and the update includes the calculation of a second transition probability for each transition from the first scenario to each possible second scenario in the set of possible reference scenarios.

[0072] According to one embodiment, if the second transition probability is higher than the first transition probability, the processor is configured to update the first transition probability by increasing its value by a set value.

[0073] According to one embodiment, if the second transition probability is lower than the first transition probability, the processor is configured to update the first transition probability by decreasing its value by a set value.

[0074] According to one embodiment, the processor is configured to update each first transition probability that is amenable to being updated using a differentiable optimization algorithm.

[0075] According to one embodiment, a classification algorithm is configured to deliver a confidence probability associated with a newly currently detected scenario.

[0076] According to one embodiment, if the classification algorithm is not configured to deliver a confidence probability associated with a newly currently detected scenario, a confidence probability having an arbitrary value will be assigned to the newly currently detected scenario.

[0077] According to one embodiment, each reference scenario is assigned an identifier, and the filter is configured to perform a filtering processing operation based on the identifier of each newly currently detected scenario and the updated confidence probability associated with each newly currently detected scenario.

[0078] The device can be, for example, a cellular mobile phone or a digital tablet, or any type of smart object, in particular a smartwatch, which is connected or not connected to the Internet.

[0079] In Figure 1 , reference APP represents an electronic device that will be considered, in this non-limiting example, as a wireless communication device equipped with an antenna ANT. The device can be a cellular mobile phone, such as a smartphone, or a digital tablet. Although the present invention is applicable to any type of device and any type of scenario, wireless communication devices will now be more specifically mentioned.

[0080] In this case, the device APP includes a plurality of measurement sensors CPT1 - CPTj, where j = 1 to M.

[0081] As an indication, the sensor CPTj can be selected from the group formed by an accelerometer, a gyroscope, a magnetometer, an audio sensor such as a microphone, a barometer, a proximity sensor or an optical sensor.

[0082] Of course, the device can be equipped with multiple accelerometers and / or multiple gyroscopes and / or multiple magnetometers and / or multiple audio sensors and / or barometers, and / or with one or more proximity sensors, and / or with one or more optical sensors.

[0083] The audio sensor is a useful environmental descriptor. Specifically, if the device is not moving, the audio sensor can help to detect the nature of the environment. Of course, depending on the application, environmental sensors such as accelerometers or even gyroscope or magnetometer types, or audio sensors, or even a combination of these two types of sensors, or other types of sensors such as non-inertial sensors with temperature, heat and brightness sensor types may be used.

[0084] These environmental measurement sensors can in particular be combined with a traditional discrimination algorithm ALC (for example of the decision tree type and intended to process, for example, the filtered raw data originating from these sensors) in a multi-mode method to form a detector MDET, which is thus able to detect, for example, whether the device APP is in one environment or another (restaurant, moving vehicle, etc.) or whether the carrier of the device (for example, a smartwatch) is performing a specific activity (walking, running, cycling, etc.).

[0085] As a non-limiting example, now assume that all environmental sensors CPT1 - CPTM participate in the detection of the scene and provide data to the discrimination algorithm ALC at the time of measurement to make the detection of the scene possible.

[0086] As will be seen in more detail below, the control of the detection of the scenes obtained based on the classification algorithm uses a filtering processing operation on the identifiers of these scenes and uses the updated confidence probabilities associated with these detected scenes.

[0087] This filtering processing operation is implemented in the filter MFL.

[0088] A non-limiting example of a classification algorithm that provides an initial confidence probability for each detected scene will now be described. Such an algorithm is described, for example, in French patent application No. 1752947 (filed on April 5, 2017) and US corresponding application No. 15 / 936,567 (filed on March 27, 2014), and some features of the algorithm will be reviewed here. These applications are incorporated herein by reference.

[0089] The discrimination algorithm implemented in software form in the scene detector MDET is in this case a decision tree, which has undergone a learning phase on the measurement database of the environmental sensors. This decision tree is particularly easy to implement and requires only a few kilobytes of memory and a working frequency of less than 0.01 MHz.

[0090] It is stored in the program memory MM1.

[0091] The decision tree ALC acts on the attribute vector. The tree consists of a series of nodes. Each node is assigned to test an attribute.

[0092] Two branches leave the node.

[0093] In addition, the output of the tree includes leaves corresponding to the reference scenes that the device APP should detect.

[0094] These reference scenes can be, for example but not limited to, "ship", "airplane", "vehicle", "walking" scenes, which represent the environments where the device APP (in this case a phone) may be located.

[0095] The detector MDET also includes an acquisition circuit ACQ, which is configured to acquire the current value of the attribute based on the measurement data originating from the sensors.

[0096] The detector MDET includes a command circuit MCM, which is configured to activate the software module ALC with the current value of the attribute in order to obtain a path within the decision tree and obtain a scene from the set of reference scenes at the output of the path, and the obtained scene forms the detected scene.

[0097] In addition, the discrimination algorithm also delivers an initial confidence probability associated with the detected scene, which will be stored in the memory MM2.

[0098] In this regard, after the detection of the detected scene and based on the knowledge of the detected scene, especially by making additional trips on the path using the knowledge at each node of the detected scene, the initial confidence probability associated with the detected scene is formulated.

[0099] There are other classification algorithms or classifiers different from decision trees and well-known to those skilled in the art for detecting scenarios. The neural network algorithm known to those skilled in the art can be particularly mentioned, and for this neural network algorithm, reference can be made (for example, for all useful purposes) to the work entitled Neural Network Design by Martin T Hagan and Howard B Demuth (2nd Edition, September 1, 2014), or the algorithm known as GMM ("Gaussian Mixture Model") to those skilled in the art. Those skilled in the art can refer to (for example, for all useful purposes) the tutorial slides entitled "Clustering with Gaussian Mixtures" by Andrew Moore, which can be found on the website https: / / www.autonlab.org / tutorials / gmm.html.

[0100] These two algorithms are also configured to deliver an initial confidence probability for each detected scenario.

[0101] The device APP also includes a processor MDC, which is configured to update the initial confidence probability.

[0102] The processor MDC includes a counter MDC1, whose function will be detailed below, and an arithmetic logic unit (ALU) MDC2, which is configured to perform arithmetic operations and comparisons.

[0103] The operations performed by the arithmetic logic unit MDC2 will be detailed in Figure 4 and 5 below.

[0104] The device APP also includes a block BLC capable of interacting with the detector MDET to facilitate processing the detected scenarios and transmitting information via the device's antenna ANT.

[0105] Of course, if the device is not a connected device, the antenna is optional.

[0106] The device also includes a controller MCTRL, which is configured to continuously activate the detector MDET to facilitate a series of scenario detection steps spaced apart from each other by time intervals.

[0107] These various circuits BLC, MDET, MCTRL, MDC, and MFL can be formed, for example, at least partially by software modules within the microprocessor PR of the device APP, such as the microprocessor sold by STMicroelectronics under STM32.

[0108] As Figure 2As shown, the filter MFL is configured to perform a filtering processing operation 10 using, for example, a meta-filter that acts on the successive decisions delivered by a classification algorithm ALC, regardless of their type, and includes a meta-classifier.

[0109] In this case, the filtering processing operation 10 is a sliding time filtering processing operation PC2 on these currently detected scenes over a filtering window of size M, based on the identifier of each newly currently detected scene considered in the window and the updated confidence probability associated with this newly currently detected scene. This filtering processing operation is described in the above patent applications FR1754080 and US15 / 924,608.

[0110] Of course, depending on the envisaged application, the person skilled in the art can easily define the value of M, which particularly defines the waiting time of the filter.

[0111] The output of the filtering processing operation continuously delivers the filtered detected scenes 15.

[0112] In this regard, an identifier ID (e.g., a number) is assigned to each reference scene.

[0113] In this case, it is assumed that there are four reference scenes, namely the scenes "vehicle", "airplane", "ship" and "walking".

[0114] The reference scene "vehicle" has the identifier 1.

[0115] The reference scene "airplane" has the identifier 2.

[0116] The reference scene "ship" has the identifier 3, and the reference scene "walking" has the identifier 4.

[0117] As a result, since each detected scene belongs to one of the reference scenes, the identifier of the currently detected scene is an identifier having the same value as one of these reference scenes.

[0118] The confidence probability PC2 associated with the newly detected scene is in this case a confidence probability updated by weighting 20 with a first transition probability TrPT1 of the initial confidence probability PC1 delivered by the discrimination algorithm ALC.

[0119] This update is performed by the processor MDC.

[0120] If this initial confidence probability PC1 is not delivered by the discrimination algorithm ALC, it is considered to have an arbitrary value, for example set to 100%. In this case, the weighting of the first transition probability TrPT1 is performed on the confidence probability with the set value.

[0121] The first transition probability TrPT1 makes it possible to quantify the probability that a user changes from one scenario to another.

[0122] It is advantageous to consider parameters in scene detection because some transitions from one scene to another are unlikely or even impossible.

[0123] Therefore, by considering the first transition probability TrPT1, the initial confidence probability PC1 associated with the currently detected scene is improved.

[0124] In this regard, as Figure 3 illustrated, the first transition probability TrPT1 can be extracted from a transition table TT stored in the memory MM3.

[0125] The transition table TT is a two-dimensional table. The vertical axis AxV represents the previously detected scene, and the horizontal axis AxH represents the currently detected scene.

[0126] Therefore, each cell CEL of the transition table TT represents the transition probability from the previously detected scene to the currently detected scene.

[0127] For example, the transition probability from the previously detected scene "vehicle" to the currently detected scene "walking" is 0.8, which represents a high transition probability.

[0128] Conversely, the transition probability from the previously detected scene "vehicle" to "boat" is 0.1, which represents a low probability and thus an unlikely transition for a user having the device APP.

[0129] For example, according to the application, the table can be initially predetermined.

[0130] Of course, the habits of the user having the device APP change.

[0131] Therefore, it is worthwhile to update the transition probability TrPT1 that may have arbitrary values in order to ensure good accuracy of the confidence probability PC2.

[0132] This update is also performed by the processor MDC.

[0133] Now refer more specifically to Figure 4 and 5 in order to illustrate the various steps of updating the first transition probability TrPT1 by an algorithm consisting of two parts.

[0134] Before any update, a non-sliding observation window is activated.

[0135] Using a counter MDC1 within a given time interval, each transition instance from the first scene i to the second scene j is accumulated, corresponding to N transitions.

[0136] Advantageously, these N conversions are stored in the memory MM3 in the form of a list.

[0137] Of course, the number N will be chosen by a person skilled in the art according to the corpus of the reference scenarios.

[0138] Once the number N corresponding to the N conversions is reached, the instructions of the algorithm that makes it possible to update the transition table TT are read.

[0139] The first step is step 21 of initializing the value of the first variable i to 1, corresponding to the first reference scenario with identifier 1, and initializing the value of the second variable j to 2, corresponding to the second reference scenario with identifier 2.

[0140] Step 22 makes it possible to check whether i is indeed different from j. This is useful for the next instruction of the algorithm.

[0141] If i is different from j, the list of N conversions is run in order to count, for example using the counter MDC1, the instances N(i->j) of the transition from the first reference scenario with identifier 1 to the second reference scenario with identifier 2.

[0142] The value of N(i->j) is stored in the memory MM3.

[0143] In step 24, the value of N(i) where i equals 1, corresponding to the number of transitions from the first reference scenario with identifier 1 to each possible reference scenario, is incremented by the number of instances N(i->j) of the first transition from the first reference scenario with identifier 1 to the second reference scenario with identifier 2.

[0144] The formula (I) implemented in step 24 is appended.

[0145] The value of N(i) is stored in the memory MM3.

[0146] Next, step 25 makes it possible to check whether j has reached the maximum value, i.e., the number jmax corresponding to the number of reference scenarios, which is 4 in this example.

[0147] Since the maximum value has not been reached, in step 23, the value of j is incremented to 3 and the same operation is performed again until j reaches the maximum value jmax, which is 4 in this example.

[0148] In this case, in step 26, it is checked whether the maximum value imax of i, i.e., 4, has been reached. This makes it possible to check that the algorithm has indeed traversed all possible reference scenarios.

[0149] If this is not the case, all the steps described above are repeated and in step 28 the variable j is re-initialized to 1.

[0150] Thus, once all the reference scenarios have been considered, Figure 5 the second part of the algorithm illustrated in Figure 5 is executed.

[0151] The first step is step 31 which initializes the value of the first variable i to 1, corresponding to the first reference scenario with identifier 1, and initializes the value of the second variable j to 2, corresponding to the second reference scenario with identifier 2.

[0152] Step 32 checks whether N(i) is greater than 0. In other words, it checks whether a transition has occurred from the first reference scenario i (where i equals 1) to any of the second reference scenarios during a given interval.

[0153] If N(i) is greater than 0, the algorithm moves to the following instruction shown in step 33, which makes it possible to check whether i is indeed different from j. This is first useful for the next instruction of the algorithm.

[0154] If i is different from j, the arithmetic logic unit MDC2 calculates in step 34 the second transition probability TrPT2(i->j) from the first reference scenario i (where i equals 1, corresponding to the first reference scenario with identifier 1) to the second reference scenario j (where j equals 2, corresponding to the second reference scenario with identifier 2).

[0155] This additional calculation (II) is the number of instances N(i->j) of the first transition from the first reference scenario with identifier 1 to the second reference scenario with identifier 2 divided by N(i) (corresponding to the number of transitions from the first reference scenario with identifier 1 to each possible reference scenario).

[0156] Thus, in step 331, this second transition probability TrPT2(i->j) is compared with the first transition probability TrPT1(i->j) in the arithmetic logic unit MDC2.

[0157] Thus, if the second transition probability TrPT2(i->j) is higher than the first transition probability TrPT1(i->j), the value of the first transition probability TrPT1(i->j) is increased by the first set value δ1 in step 37 (additional equation III).

[0158] This first set value δ1 can be, for example, between 0.01 and 0.05.

[0159] The first transition probability TrPT1(i->j) is updated and stored in the transition table TT.

[0160] Otherwise, in step 36 (additional formula IV), the value of the first transition probability TrPT1(i->j) is decreased by a second set value δ2.

[0161] The second set value δ2 can be, for example, between 0.01 and 0.05.

[0162] The first transition probability TrPT1(i->j) is updated and stored in the transition table TT.

[0163] Next, step 38 makes it possible to check whether j has reached the maximum value, i.e., the number jmax corresponding to the number of reference scenarios, i.e., 4 in this example.

[0164] If the maximum value jmax has not been reached, in step 39, the value of j is incremented to 3 and the same operation is performed again until the maximum value jmax of j is reached, i.e., 4.

[0165] In this case, in step 40, it is checked whether the maximum value imax of i has been reached, i.e., 4. This makes it possible to check that the algorithm has indeed traversed all possible reference scenarios.

[0166] If this is not the case, all the steps described above are repeated and the variable j is re-initialized to 1.

[0167] Thus, once all reference scenarios have been considered, there are no more instructions to execute.

[0168] As a result, any transition between a first scenario i from the set of possible scenarios and a second scenario j from the set of reference scenarios, observed by a non-sliding observation window during a given interval, makes it possible to update the transition probability from the first scenario i to the second scenario j and thus subsequently improve the confidence probability delivered by the discrimination algorithm ALC.

[0169] Furthermore, the present invention is not limited to these embodiments and various implementations, but includes all its variants.

[0170] Thus, the update of the transition probability TrPT(i->j) can be performed by any other algorithm within the scope of those skilled in the art, in particular by a differentiable optimization algorithm that implements the formula (V) presented in the appendix.

[0171] Appendix

[0172]

[0173]

[0174] TrPT1(i→j)=TrPT1(i→j)+δ 1 (III)

[0175] TrPT1(i→j) = TrPT1(i→j) - δ 2 (IV)

[0176] TrPT1(i→j) = TrPT1(i→j) + δ(TrPT1(i→j) - TrPT2(i→j)) (V)

Claims

1. A method for controlling a device to perform scene detection from a set of possible reference scenes, the method comprises: detecting scenes from the set of possible reference scenes using at least one classification algorithm at consecutive detection instances; assigning an initial confidence probability to each newly currently detected scene; updating the initial confidence probability according to a first transition probability from a previously detected scene to the newly currently detected scene; performing a filtering operation on the currently detected scenes based at least on the updated confidence probabilities associated with each newly currently detected scene, the output of the filtering operation continuously delivering the filtered detected scenes; and assigning an identifier to each reference scene, and the filtering operation on the currently detected scenes is implemented based on the identifier of each newly currently detected scene and the updated confidence probabilities associated with the newly currently detected scene.

2. The method according to claim 1, wherein updating the initial confidence probability associated with the newly currently detected scene comprises: multiplying the initial confidence probability by the first transition probability.

3. The method according to claim 1, further comprises: normalizing the updated confidence probabilities associated with the newly currently detected scene.

4. The method according to claim 1, wherein the classification algorithm delivers the initial confidence probability associated with the newly currently detected scene.

5. The method according to claim 1, wherein, when the classification algorithm does not deliver the initial confidence probability associated with the newly currently detected scene, the newly currently detected scene is assigned an initial confidence probability with an arbitrary value.

6. A method for controlling a device to perform scene detection from a set of possible reference scenes, the method comprises: detecting scenes from the set of possible reference scenes using at least one classification algorithm at consecutive detection instances; assigning an initial confidence probability to each newly currently detected scene; updating the initial confidence probability according to a first transition probability from a previously detected scene to the newly currently detected scene; performing a filtering operation on the currently detected scenes based at least on the updated confidence probabilities associated with each newly currently detected scene, the output of the filtering operation continuously delivering the filtered detected scenes; assigning a first transition probability to each transition from a first scene to a second scene in the set of possible reference scenes, the first transition probability having an arbitrary value or an updated value; and assigning an identifier to each reference scene, and the filtering operation on the currently detected scenes is implemented based on the identifier of each newly currently detected scene and the updated confidence probabilities associated with the newly currently detected scene.

7. The method according to claim 6, further comprises: When a transition from the first scenario to the second scenario is performed during a given time interval, update the first transition probability.

8. The method according to claim 7, wherein the updating comprises: For each transition from the first scenario to each possible second scenario from the set of possible reference scenarios, calculate a second transition probability.

9. The method according to claim 8, wherein updating the first transition probability comprises: When the second transition probability is higher than the first transition probability, increase the value of the first transition probability by a first set value; and wherein updating the first transition probability comprises: when the second transition probability is lower than the first transition probability, decrease the value of the first transition probability by a second set value.

10. The method according to claim 7, wherein updating the first transition probability comprises: Update the first transition probability using a differentiable optimization algorithm.

11. An electronic device, comprising: A detector configured to detect a scenario from a set of possible reference scenarios using a classification algorithm at consecutive detection times, and each newly currently detected scenario is assigned an initial confidence probability; A processor configured to update the initial confidence probability according to a first transition probability from a previously detected scenario to the newly currently detected scenario; and A filter configured to perform a filtering processing operation based at least on the updated confidence probability associated with each newly currently detected scenario, and continuously deliver the filtered detected scenarios; wherein, in the case where each reference scenario is assigned an identifier, the filter is configured to perform a filtering processing operation based on the identifier of each newly currently detected scenario and the updated confidence probability associated with each newly currently detected scenario.

12. The electronic device according to claim 11, wherein the processor is configured to update the initial confidence probability associated with the newly currently detected scenario by multiplying the initial confidence probability by the first transition probability.

13. The electronic device according to claim 11, wherein the processor is configured to normalize the updated confidence probability associated with the newly currently detected scenario.

14. The electronic device according to claim 11, wherein each transition from a first scenario to a second scenario from the set of possible reference scenarios is assigned a first transition probability having an arbitrary value or an updated value.

15. The electronic device according to claim 14, wherein the processor is configured to update the first transition probability when a transition from the first scenario to the second scenario is performed during a given time interval, and the updating comprises: Calculation of a second transition probability for each transition from the first scenario to each possible second scenario from the set of possible reference scenarios.

16. The electronic device according to claim 15, wherein the processor is configured to update the first transition probability by increasing the value of the first transition probability by a first set value when the second transition probability is higher than the first transition probability.

17. The electronic device according to claim 15, wherein the processor is configured to update the first transition probability by decreasing the value of the first transition probability by a second set value when the second transition probability is lower than the first transition probability.

18. The electronic device according to claim 14, wherein the processor is configured to update each first transition probability to be updated using a differentiable optimization algorithm.

19. The electronic device according to claim 11, wherein the classification algorithm is configured to deliver the initial confidence probability associated with the newly currently detected scene.

20. The electronic device according to claim 11, wherein when the classification algorithm is not configured to deliver the initial confidence probability associated with the newly currently detected scene, an initial confidence probability with an arbitrary value is assigned to the newly currently detected scene.

21. The electronic device according to claim 11, wherein the device is a cellular mobile phone or a digital tablet or a smart watch.

Images