First view behavior recognition method and system based on uncertainty estimation

Abstract

The invention belongs to the technical field of behavior recognition, and discloses a first visual angle behavior recognition method and system based on uncertainty estimation, and the method comprises the steps: obtaining a first visual angle video, carrying out the preprocessing of the first visual angle video, and obtaining a training sample; performing multi-stage different-scale feature extraction on the training sample, and integrating features in different stages to obtain a first feature; classifying and predicting the first feature, quantifying the uncertainty of a classification result, obtaining a classified evidence vector, defining a first loss function, minimizing the difference between an evidence parameter predicted by a model and a real situation, obtaining a prediction label, defining a second loss function, and minimizing the difference between the prediction label and a real label; and summarizing loss functions, optimizing model parameters, and obtaining a trained behavior recognition model. The method has more advantages in coping with uncertainty and unknown actions, the accuracy of the open set first visual angle behavior recognition task is remarkably improved, and application in the open environment is more reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of behavior recognition, and in particular to a first-person perspective behavior recognition method and system based on uncertainty estimation. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the widespread use of wearable technologies such as smart glasses, egocentric (i.e. first-person) video action recognition (VAR) is gradually becoming a major turning point in the field of video action recognition. These devices fundamentally change the way data is collected, providing a first-person perspective, which is particularly important in fields such as augmented reality and robotics. With this perspective, the system can have a deeper understanding of the user's interactive behavior, thereby improving the user experience, covering a wide range of application scenarios from personal assistants to medical monitoring and interactive games.

[0004] However, first-person video action recognition faces unique challenges, especially in the context of open set recognition. Open set recognition requires the system to not only recognize known action categories, but also detect and mark new actions that have never been seen before, classifying them as "unknown". This requirement stems from the limitations of closed set environments, where classifiers are usually trained and tested only on predefined categories and lack the ability to cope with emerging, unseen categories. In the real world, with the continuous development of technology and changes in applications, new actions emerge in an endless stream, which places higher demands on the adaptability of action recognition systems.

[0005] Most current research on action recognition from ego-view videos focuses on closed-set scenarios, and thus often fails to effectively handle the complexity of open-set environments. This shortcoming further emphasizes the need to develop new recognition methods that can maintain high accuracy and reliability in recognizing a wide range of actions, especially when encountering novel and unseen actions.

[0006] There are obvious limitations in directly applying third-person video action recognition methods to first-person behavior recognition. Although third-person video action recognition has made significant progress driven by extensive research, complex models, and dataset development, these methods are difficult to directly apply to first-person video action recognition. The main challenges lie in the differences in data characteristics between the two and the different understandings of action context caused by different perspectives. First-person videos are usually not pre-processed, and require simultaneous recognition of actions and objects due to perspective limitations and motion blur. For example, when processing the action of "cutting vegetables", the system needs to recognize not only the verb "cut" but also the noun "vegetables", which usually does not require such detailed processing in third-person video action recognition. Therefore, first-person behavior recognition methods need to be further developed to overcome these unique challenges. Summary of the invention

[0007] In order to solve the above problems, the present invention proposes a first-person perspective behavior recognition method and system based on uncertainty estimation, which can not only recognize actions and objects in videos, but also evaluate the certainty of these recognition results. It has more advantages in dealing with uncertainty and unknown actions, and significantly improves the accuracy of open set first-person perspective behavior recognition tasks, making its application in open environments more reliable.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a first-person perspective behavior recognition method based on uncertainty estimation, comprising the following steps:

[0010] Obtain a first-person perspective video, pre-process the first-person perspective video, and obtain a training sample;

[0011] Perform multi-stage and different-scale feature extraction on the training samples, integrate the features at different stages, and obtain the first feature;

[0012] Perform classification prediction on the first feature, quantify the uncertainty of the classification result, obtain the classification evidence vector, define the first loss function, minimize the difference between the evidence parameter predicted by the model and the actual situation, obtain the predicted label, define the second loss function, minimize the difference between the predicted label and the actual label;

[0013] Summarize the loss function, optimize the model parameters, and obtain the trained behavior recognition model.

[0014] As an optional implementation, the first-person-view video is preprocessed, including labeling the acquired first-person-view video with a real label.

[0015] As an optional implementation, multi-stage and different-scale feature extraction is performed on the training samples, specifically:

[0016] In each feature extraction stage, dynamic position embedding is used to capture the spatial and temporal position information in the video, a multi-head relation aggregator is used to aggregate the relation information of different time frames, and a feed-forward network is used to further extract and enhance features.

[0017] As an optional implementation, there are four feature extraction stages in total, each feature extraction stage has a different feature extraction scale, and a hierarchical feature representation is obtained by integrating features from different stages.

[0018] As an optional implementation, the first loss function is defined as:

[0019]

[0020] in, represents the sum of the parameters in the Dirichlet set in K categories, representing the overall strength of the evidence, is a constant, log(α j ) is the evidence parameter α for the model prediction j The logarithmic transformation of t j is the binary indicator in the one-hot encoded vector corresponding to the action label y, j is the classification category, and K represents the number of categories.

[0021] As an optional implementation, the second loss function is defined as:

[0022]

[0023] Among them, t j is the binary indicator in the one-hot encoded vector corresponding to the action label y, j is the classification category, K represents the number of categories, is the original prediction value of the model for the jth category.

[0024] In a second aspect, the present invention provides a first-person perspective behavior recognition system based on uncertainty estimation, comprising:

[0025] The data acquisition and preprocessing module is configured to: acquire the first-person perspective video, preprocess the first-person perspective video, and obtain a training sample;

[0026] The feature extraction module is configured to: perform multi-stage and different-scale feature extraction on the training sample, integrate the features at different stages, and obtain the first feature;

[0027] The classification module is configured to: perform classification prediction on the first feature, quantify the uncertainty of the classification result, obtain the classification evidence vector, define a first loss function, minimize the difference between the evidence parameter predicted by the model and the actual situation, obtain the predicted label, define a second loss function, and minimize the difference between the predicted label and the actual label;

[0028] The model training module is configured to: summarize the loss function, optimize the model parameters, and obtain a trained behavior recognition model.

[0029] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0030] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0031] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present disclosure proposes a first-person action recognition method and system based on uncertainty estimation, which is based on the OpenEAR framework and can effectively handle the diverse and unplanned content commonly seen in the first-person perspective. Through multi-scale feature extraction, complex actions in first-person videos can be understood, and local details and global contextual information in the video can be effectively captured through a four-stage feature extraction process. By combining with a two-headed evidence deep network classifier, good performance is achieved in reliable, gerund-distinguished recognition of open-set first-person video actions. Not only can actions and objects in the video be identified, but also the certainty of these recognition results can be evaluated. This capability makes the model more advantageous in dealing with uncertainty and unknown actions, and can maintain a high recognition accuracy in complex and changing environments. The accuracy of the open-set first-person action recognition task is significantly improved, making its application in an open environment more reliable.

[0034] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0036] Figure 1 This is a flowchart of a first-person perspective behavior recognition method based on uncertainty estimation provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0040] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment provides a first-person perspective behavior recognition method based on uncertainty estimation, comprising the following steps:

[0043] S1 obtains a first-person perspective video, preprocesses the first-person perspective video, and obtains a training sample;

[0044] S2 extracts features of different scales at multiple stages from the training samples, integrates the features at different stages, and obtains the first feature;

[0045] S3 performs classification prediction on the first feature, quantifies the uncertainty of the classification result, obtains the classification evidence vector, defines the first loss function, minimizes the difference between the evidence parameter predicted by the model and the actual situation, obtains the predicted label, and defines the second loss function, minimizes the difference between the predicted label and the actual label;

[0046] S4 summarizes the loss function, optimizes the model parameters, and obtains a trained behavior recognition model.

[0047] Preprocessing the first-person perspective video includes labeling the acquired first-person perspective video with real labels.

[0048] In the field of first-person video analysis, the temporal consistency of user behaviors provides a unique opportunity for action recognition. Given that the background usually remains static, the user's movement is mainly based on their own motion, and the interactive objects do not change frequently, are relatively fixed, and are relatively stable, we take full advantage of these consistencies to enhance the performance of our model. The model involved in the present disclosure includes a unified EAR Transformer module, which excels in reducing redundant information between video frames and capturing the inherent temporal relationships in the data. The EAR Transformer module consists of three key components: dynamic position embedding (DPE), multi-head relationship aggregator (MHRA), and feed-forward network (FFN). When the first-person video (denoted as ) is input into the system, the model's learning algorithm is deployed in a structured manner, using the above components to process the video data in sequence. Specifically, the dynamic position embedding (DPE) is responsible for capturing the spatial and temporal position information in the video, the multi-head relation aggregator (MHRA) aggregates the relation information from different time frames, and the feed-forward network (FFN) further extracts and strengthens the features.

[0049] Through this structured processing process, the model disclosed in the present invention can make full use of the inherent temporal consistency in the video, thereby achieving higher accuracy and robustness in the action recognition task. The specific process can be expressed as follows:

[0050]

[0051] Among them, DPE is applied to the input first-person video X through 3D deep convolution (DWConv) ev , which is characterized by zero padding to preserve the spatial dimension. The DWConv operation is used to dynamically encode position information while taking into account the temporal dimension, which is crucial for understanding the frame sequence in first-person videos.

[0052] MHRA performs label relation learning via a multi-head fusion strategy, which can be formulated as:

[0053]

[0054] Given an input tensor First, reshape it into a sequence of tokens, i.e. Where L = T × H × W. n (·) denotes the function of the relation aggregator (RA) in the nth head. represents a learnable matrix that integrates the outputs of N different heads. In the relational aggregator (RA), each head is responsible for encoding the context of the token and learning the correlation between tokens. After the linear transformation, the token becomes the context Enables the Relationship Aggregator (RA) to tag relevance The key part of the relation aggregator (RA) is to label the relevance The learning mechanism of CNN is crucial for understanding the relationship between tags in video sequences.

[0055] In the MHRA architecture, learning of token (label) affinities is performed hierarchically in the depth of the network, leveraging both local and global contextual cues. In the shallow (early) layers of the network, label affinities are treated as a learnable parameter matrix that operates within a local 3D neighborhood. In this case, for a given anchor label X i , the relation aggregator (RA) learns the relationship between the tag and the tubule The local spatiotemporal affinity between other markers in is as follows:

[0056]

[0057] in represents the set of learnable parameters, and X j means Any adjacent token within . (ij) represents the relative index of the token (token), which is used to determine the weight of the aggregation.

[0058] In contrast, at deeper layers of the network, token affinity learning is extended to a global scale. Instead of local communities, the network examines the content similarity of all tokens across the entire video sequence.

[0059] In contrast, in the deep (later) layers of the network, the learning of tag affinities is extended to a global scale. At this point, the network is no longer limited to the local neighborhood, but analyzes the content similarity of all tags in the entire video sequence:

[0060]

[0061] In this formula, X j It can be any marker in the entire 3D global space, which covers the entire time and space dimensions T×H×W. Function and is a linear transformation that projects the tags into a space where their content-based similarity can be computed. With this global view, the network is able to understand the broader context and relationships across the entire video, which is particularly critical for recognizing actions in first-person videos, where scenes and user interactions can change dramatically over time.

[0062] When analyzing first-person videos, our goal is to convert video X ev The action described in the video is derived by combining the verb and noun categories. Figure 1 As shown in Figure 2, our approach uses two classifiers to independently predict the verb and noun categories:

[0063] L v =CLS v (AvgPooling(Z))

[0064] L n =CLS n (AvgPooling(Z))

[0065] in, Represents the output features from the EAR Transformer. These features are initially aggregated by average pooling, converting the dimensions to Then two independent classifiers (one for verbs, CLS v and a classifier CLS for nouns n Map these summarized features to obtain the corresponding verb and noun categories. The logits predicted by the verb and noun classifiers are L v and L n , the final verb and noun classification probabilities are expressed as P v =softmax(L v ) and P n =softmax(L n ).

[0066] To improve classification predictions, we adopt a modified version of the soft target cross entropy loss, which is defined as follows:

[0067]

[0068] Among them, K∈{M,N} and *∈{v,n}, j is the classification category, is the original prediction value (also called logits) of the model for the jth category, K represents the number of categories. In the specific calculation of the current formula, the value of K is M or N, M is the number of verb categories, and N is the number of noun categories. In this loss function, tj is a binary indicator in the one-hot encoded vector corresponding to the action label y. The indicator t is only valid if the true action label y matches the category j. j The value is 1. This loss function is used to guide the network to minimize the difference between the predicted label and the true label, thereby improving the accuracy of action recognition in the first-person video classification task.

[0069] We adopt the principle of Evidence Deep Learning (EDL) for K-class uncertainty-aware classification. The evidence deep neural network is used as part of the classifier to enhance the model's ability to cope with open set scenarios by quantifying classification uncertainty, while satisfying the classification of gerunds and applying uncertainty guidance to the final classification results separately, which are applied separately after obtaining verb and noun features. By introducing Dirichlet distribution, classification uncertainty is quantified to enhance the model's ability to cope with open set scenarios. This method enables the classifier to not only distinguish known categories, but also identify and reject samples of unknown categories. The classification uncertainty is effectively quantified, which is particularly critical in applications where confidence in the prediction is as important as the prediction result itself. We assume that the class probability It follows the Dirichlet distribution Dir(P|α), where represents the concentration parameter of the distribution, or the "Dirichlet strength". The learning process consists of minimizing the negative log-likelihood loss function:

[0070]

[0071] in, represents the sum of the parameters in the Dirichlet set in K categories, representing the overall strength of the evidence, is a constant, log(α j ) is the evidence parameter α for the model prediction j The logarithmic transformation of t j is a binary indicator in the one-hot encoded vector corresponding to the action label y, and j is the classification category, ranging from 1 to K.

[0072] In the testing phase, for a given video X ev , the verb / noun classification branch generates a non-negative evidence vector Specifically:

[0073] The features extracted by EAR Transformer from the input video are input into the deep neural network (DNN), and the feature vector h output by the last layer of DNN is i is used to predict evidence. Specifically, the feature vector h i Transformed into evidence vector e through an exponential evidence function i , that is, e i =exp(ωT )h i , where ω is the learnable weight of the DNN layer.

[0074] This evidence vector not only provides the logits of the output probability, but also conforms to the framework of subjective logic and evidence theory. The expected value of the classification probability is calculated as Where α = e + 1, following the principles of evidence theory and subjective logic. The measurement of classification uncertainty is given by the formula This provides a quantitative assessment of the confidence in the classification result, where higher uncertainty values ​​indicate lower confidence and vice versa. In our experiments, the threshold for uncertainty is set to 0.1. This means that when u ≥ 0.1, the model believes that the prediction result is uncertain and needs further processing. On the contrary, when u < 0.1, the model is confident that the prediction result is correct.

[0075] Example 2

[0076] This embodiment provides a first-person perspective behavior recognition system based on uncertainty estimation, including:

[0077] The data acquisition and preprocessing module is configured to: acquire the first-person perspective video, preprocess the first-person perspective video, and obtain a training sample;

[0078] The feature extraction module is configured to: perform multi-stage and different-scale feature extraction on the training sample, integrate the features at different stages, and obtain the first feature;

[0079] The classification module is configured to: perform classification prediction on the first feature, quantify the uncertainty of the classification result, obtain the classification evidence vector, define a first loss function, minimize the difference between the evidence parameter predicted by the model and the actual situation, obtain the predicted label, define a second loss function, and minimize the difference between the predicted label and the actual label;

[0080] The model training module is configured to: summarize the loss function, optimize the model parameters, and obtain a trained behavior recognition model.

[0081] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0082] In further embodiments, there is also provided:

[0083] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0084] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0085] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0086] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0087] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0088] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in embodiment 1 is implemented.

[0089] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0090] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0091] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0092] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0093] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A first-person perspective behavior recognition method based on uncertainty estimation, characterized in that: The following steps are involved: Obtain a first-person perspective video, pre-process the first-person perspective video, and obtain a training sample; Perform multi-stage and different-scale feature extraction on the training samples, integrate the features at different stages, and obtain the first feature; Perform classification prediction on the first feature, quantify the uncertainty of the classification result, obtain the classification evidence vector, define the first loss function, minimize the difference between the evidence parameter predicted by the model and the actual situation, obtain the predicted label, define the second loss function, minimize the difference between the predicted label and the actual label; Summarize the loss function, optimize the model parameters, and obtain the trained behavior recognition model.

2. The first-person perspective behavior recognition method based on uncertainty estimation according to claim 1, characterized in that: The first-person perspective video is preprocessed, including annotating the acquired first-person perspective video with real labels.

3. The first-person perspective behavior recognition method based on uncertainty estimation according to claim 1, characterized in that: Perform multi-stage and different-scale feature extraction on the training samples, specifically: In each feature extraction stage, dynamic position embedding is used to capture the spatial and temporal position information in the video, a multi-head relation aggregator is used to aggregate the relation information of different time frames, and a feed-forward network is used to further extract and enhance features.

4. The first-person perspective behavior recognition method based on uncertainty estimation according to claim 3, characterized in that: There are four feature extraction stages in total, and the feature extraction scale of each feature extraction stage is different. By integrating the features of different stages, a hierarchical feature representation is obtained.

5. The first-person perspective behavior recognition method based on uncertainty estimation according to claim 1, characterized in that: The first loss function is defined as: in, represents the sum of the parameters in the Dirichlet set in K categories, representing the overall strength of the evidence, is a constant, log(α j ) is the evidence parameter α for the model prediction j The logarithmic transformation of t j is the binary indicator in the one-hot encoded vector corresponding to the action label y, j is the classification category, and K represents the number of categories.

6. The first-person perspective behavior recognition method based on uncertainty estimation according to claim 1, characterized in that: The second loss function is defined as: Among them, t j is the binary indicator in the one-hot encoded vector corresponding to the action label y, j is the classification category, K represents the number of categories, is the original prediction value of the model for the jth category.

7. A first-person perspective behavior recognition system based on uncertainty estimation, characterized in that: include: The data acquisition and preprocessing module is configured to: acquire the first-person perspective video, preprocess the first-person perspective video, and obtain a training sample; The feature extraction module is configured to: perform multi-stage feature extraction of the training sample at different scales, integrate the features at different stages, and obtain the first feature; The classification module is configured to: perform classification prediction on the first feature, quantify the uncertainty of the classification result, obtain the classification evidence vector, define a first loss function, minimize the difference between the evidence parameter predicted by the model and the actual situation, obtain the predicted label, define a second loss function, and minimize the difference between the predicted label and the actual label; The model training module is configured to: summarize the loss function, optimize the model parameters, and obtain a trained behavior recognition model.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images