Method and apparatus for motion type classification based on indirect incremental learning algorithm
By using an indirect incremental learning algorithm to acquire motion data from sensors and construct a new classifier, the problem of simultaneous increase in features and categories in action recognition tasks is solved, and efficient classification is achieved with limited labeled samples.
Patent Information
- Application Number
- CN202310464681.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-26
- Publication Date
- 2026-01-02
- Estimated Expiration
- 2043-04-26
AI Technical Summary
Existing technologies struggle to effectively handle situations where features and categories increase simultaneously in action recognition tasks, especially when labeled samples are limited, to ensure classification performance.
An indirect incremental learning algorithm is adopted to acquire raw and incremental motion data by setting sensors at different parts of the moving object, construct a new classifier, and use the cross-entropy loss function, regularization term and indirect similarity constraint term to construct the objective function for incremental learning.
With the addition of features and categories, the consistency and diversity of classification results are guaranteed, enabling effective identification of motion types.
Smart Images
Figure CN116502137B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer, in particular to a motion type classification method and device based on indirect incremental learning algorithm. BACKGROUND
[0002] With the emergence of new data collection methods in many dynamic environment applications, the feature space of samples is constantly expanding. With the addition of new types of features of data, more observations can be obtained, leading to the possibility of increasing the number of classes. For example, in the task of action recognition, we can identify different warm-up exercises by using old features generated by sensors during warm-up. At the same time, we can better identify new formal movements such as cycling and rowing by obtaining new attributes through newly added sensors. Learning with simultaneous feature and class increase is essential, but few people have studied it, especially when there are limited labeled samples with complete observations but the classification performance needs to be guaranteed. SUMMARY
[0003] Therefore, it is necessary to provide a motion type classification method and device based on indirect incremental learning algorithm with simultaneous feature and class increase in view of the above technical problems.
[0004] A motion type classification method based on indirect incremental learning algorithm, the method comprises:
[0005] obtaining original motion data according to sensors arranged on different parts of a motion object;
[0006] training an original classifier according to the original motion data to obtain a trained original classifier;
[0007] obtaining incremental motion data according to sensors arranged on other parts of the motion object, and constructing a new classifier capable of processing the incremental motion data according to the classifier trained on the original motion data and the incremental motion data, wherein the incremental motion data contains an increased class;
[0008] regularizing the new classifier capable of processing the incremental motion data to obtain a regularization term, and constructing an indirect similarity constraint term according to the classifier trained on the original motion data and the new classifier trained on the incremental motion data after the class is increased;
[0009] performing incremental learning on the trained original classifier according to a target function of the indirect incremental learning algorithm constructed according to a cross-entropy loss function, the regularization term, the indirect similarity constraint term, and training samples to obtain a trained incremental classifier;
[0010] Classifying the motion type using the trained incremental classifier.
[0011] In one embodiment, further comprising: constructing an indirect similarity constraint term according to the original motion data and the original motion data after adding the new class, wherein the indirect similarity constraint term is:
[0012]
[0013] wherein, denotes a feature representation of the motion data, n1 denotes a number of the original motion data, denotes a classifier parameter of the original motion data belonging to the jth class, denotes a feature representation of the incremental motion data, denotes a classifier parameter of the incremental motion data belonging to the jth class, C denotes a classification class of the original motion data, and C+1 denotes a classification class containing the new class, and denote a classifier parameter of the original motion data and the incremental motion data belonging to the kth class, respectively, denotes a probability of i-n1 data belonging to the jth class;
[0014] constructing a cross-entropy loss function according to the probability of i-n1 data belonging to the jth class, wherein the cross-entropy loss function is:
[0015]
[0016] wherein, n2 denotes a number of the incremental motion data, is a true label of the i-n1 data belonging to the jth class.
[0017] In one embodiment, further comprising: constructing an indirect similarity constraint term according to the original motion data and the original motion data after adding the new class, wherein the indirect similarity constraint term is:
[0018]
[0019] wherein, denotes a feature representation of the original motion data after adding the new class, denotes a classifier parameter of the original motion data, denotes a classifier parameter of the first C columns of the original motion data after adding the new class, wherein C denotes a classification class.
[0020] In one embodiment, further comprising: constructing a target function of the indirect incremental learning algorithm according to the cross-entropy loss function, the regular term, and the indirect similarity constraint term, comprising:
[0021]
[0022] Wherein, alpha, beta represent coefficients, Regular term is represented, Indirect approximation term is represented.
[0023] In one embodiment, the method further comprises solving the objective function using a gradient descent algorithm.
[0024] A motion type classification device based on an indirect incremental learning algorithm, the device comprising:
[0025] A data acquisition module configured to acquire original motion data from sensors arranged on different parts of a motion object;
[0026] An initial training module configured to train an original classifier based on the original data to obtain a trained original classifier;
[0027] An incremental data construction module configured to acquire incremental motion data from sensors arranged on other parts of the motion object, and to construct a new classifier capable of processing the incremental motion data based on the classifier trained based on the original motion data and the incremental motion data; wherein the incremental motion data includes an increased category;
[0028] An incremental training module configured to perform regularization processing on the new classifier capable of processing the incremental motion data to obtain a regularization term, to construct an indirect similarity constraint term based on the classifier trained based on the original motion data and the new classifier trained based on the incremental motion data after the category is increased, and to perform incremental learning on the trained original classifier based on a target function of an indirect incremental learning algorithm constructed based on a cross-entropy loss function, the regularization term, the indirect similarity constraint term, and the training sample to obtain a trained incremental classifier;
[0029] A classification module configured to perform motion type classification using the trained incremental classifier.
[0030] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:
[0031] Acquiring original motion data from sensors arranged on different parts of a motion object;
[0032] Training an original classifier based on the original motion data to obtain a trained original classifier;
[0033] Acquiring incremental motion data from sensors arranged on other parts of the motion object, and constructing a new classifier capable of processing the incremental motion data based on the classifier trained based on the original motion data and the incremental motion data; wherein the incremental motion data includes an increased category;
[0034] regularizing the new classifier capable of processing the incremental motion data to obtain a regularization term, and constructing an indirect similarity constraint term according to the classifier trained on the original motion data and the new classifier trained on the incremental motion data after the added category;
[0035] performing incremental learning on the trained original classifier according to an objective function of an indirect incremental learning algorithm constructed according to a cross-entropy loss function, the regularization term, the indirect similarity constraint term, and training samples, to obtain a trained incremental classifier;
[0036] performing motion type classification by using the trained incremental classifier.
[0037] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the following steps:
[0038] obtaining original motion data according to sensors arranged on different parts of a motion object;
[0039] training an original classifier according to the original motion data to obtain a trained original classifier;
[0040] obtaining incremental motion data according to sensors arranged on other parts of the motion object, and constructing a new classifier capable of processing the incremental motion data according to the classifier trained on the original motion data and the incremental motion data, wherein the incremental motion data contains an added category;
[0041] regularizing the new classifier capable of processing the incremental motion data to obtain a regularization term, and constructing an indirect similarity constraint term according to the classifier trained on the original motion data and the new classifier trained on the incremental motion data after the added category;
[0042] performing incremental learning on the trained original classifier according to an objective function of an indirect incremental learning algorithm constructed according to a cross-entropy loss function, the regularization term, the indirect similarity constraint term, and training samples, to obtain a trained incremental classifier;
[0043] performing motion type classification by using the trained incremental classifier.
[0044] The aforementioned motion type classification method, device, computer equipment, and storage medium based on the indirect incremental learning algorithm, when faced with an increase in features and categories, firstly regularizes the classifier corresponding to the incremental data, extracting a regularization term. Then, based on the similarity between the classifier learned from the original motion data and the classifier to be learned from the incremental data, an indirect similarity constraint term is constructed. The reason for extracting the indirect similarity constraint is to ensure that the classifier learned from the original motion data and the classifier learned from the incremental data retain a certain degree of similarity, while also requiring a certain degree of difference. Ideally, the classification results should maintain consistency among the multiple categories in the initial training, while remaining inconsistent with the new categories that appear in the incremental training. Therefore, the objective function of the indirect incremental learning algorithm, constructed based on the cross-entropy loss function, the regularization term, and the indirect similarity constraint term, and then used to train the original classifier using training samples, can ensure consistency among most classification results while allowing for a small number of differences, thereby enabling the recognition of incremental motion types. Attached Figure Description
[0045] Figure 1 This is a flowchart illustrating a motion type classification method based on an indirect incremental learning algorithm in one embodiment.
[0046] Figure 2 This is a structural block diagram of a motion type classification device based on an indirect incremental learning algorithm in one embodiment.
[0047] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0049] In one embodiment, such as Figure 1 As shown, a motion type classification method based on an indirect incremental learning algorithm is provided, including the following steps:
[0050] Step 102: Obtain raw motion data based on sensors installed on different parts of the moving object.
[0051] The motion data collected varies depending on the different parts of the moving object. Raw motion data refers to the motion data collected at a certain stage.
[0052] It is worth mentioning that for the exercise type, 11 warm-up exercises (jogging, jumping, etc.) are started before the formal exercise (i.e. cycling, rowing, etc.). During the warm-up, multiple sensors are placed on the body (left forearm, right forearm, left upper arm, right upper arm, back, etc.). At the end of the warm-up, other multiple sensors are placed (left lower leg, right lower leg, left thigh, right thigh, etc.), which generates new features and thus better identifies the subsequent formal exercise (new class).
[0053] Step 104, training the original classifier according to the original exercise data to obtain a trained original classifier.
[0054] Step 106, obtaining incremental exercise data according to the sensors arranged on other parts of the exercise object, and constructing a new classifier capable of processing the incremental exercise data according to the classifier trained on the original exercise data and the incremental exercise data.
[0055] The training sample contains an increased class.
[0056] Step 108, regularizing the new classifier capable of processing the incremental exercise data to obtain a regularization term, and constructing an indirect similarity constraint term according to the classifier trained on the original exercise data and the new classifier trained on the incremental exercise data after the class is increased.
[0057] Step 110, performing incremental learning on the trained original classifier according to the objective function of the indirect incremental learning algorithm constructed according to the cross-entropy loss function, the regularization term and the indirect similarity constraint term, and the training sample, to obtain a trained incremental classifier.
[0058] Step 112, classifying the exercise type by using the trained incremental classifier.
[0059] In the motion type classification method based on the indirect incremental learning algorithm, when facing feature increase and class increase, first, the classifier corresponding to the incremental data is regularized, and a regularization term is extracted. Then, a similarity constraint term is constructed according to the similarity of the classifier learned on the original motion data and the classifier to be learned on the incremental data. The reason for extracting the similarity constraint is to ensure that the classifier learned on the original motion data and the classifier learned on the incremental data retain a certain similarity and also have a certain difference. The most ideal case is that, in the classification result, consistency is maintained for the multiple classes on the initial training, and inconsistency is maintained for the new classes appearing on the incremental training. Therefore, the objective function of the indirect incremental learning algorithm is constructed based on the cross-entropy loss function, the regularization term and the indirect similarity constraint term. Then, the original classifier is trained by using the training samples, so that consistency is ensured between most classification results, and a small part of the classification results are allowed to have difference, so that the incremental type recognition of the motion type can be realized.
[0060] In addition, compared with traditional classification problems, there are at least two challenges in analyzing such data. (1) In these applications, the feature type and the number of classes increase simultaneously, which makes the problem very complex, especially when the increase of classes is caused by the increase of feature types, and they are coupled together. (2) It is difficult to design a better algorithm with performance guarantee relative to the two simple implementations, i.e., only using data with previous features or ignoring previous data without complete observations, because the increase of features and classes violates the independent and identically distributed assumption in the traditional learning paradigm.
[0061] In one embodiment, the initial training and the incremental training of the application are respectively named as the preparation stage and the adaptation stage, and the data form in the P stage is: Where n1 and d1 are the sample number and the feature dimension. The i-th row of data is The corresponding label is C is the number of classes. The corresponding label matrix is If the k-th sample belongs to the l-th class, Y1(k, l) = 1, otherwise Y1(k, l) = 0. Similarly, in the A stage, the data form is Where Corresponding to the original feature part, Corresponding to the new feature part, n2 is the data sample number of the A stage, d2 is the feature dimension newly added in the A stage. The label matrix of X2 is The data form in the test stage is According to the above symbol definition, our main task is to classify C+1 class data X According to the above symbol definition, our main task is to classify C+1 class data Xte The classification is performed. Of course, we can also use only the data of {X2, Y2} of the A stage to train the classifier. However, in practical applications, the data quantity n2 of the A stage is usually very small and far less than the data quantity n1 of the P stage, and only using a small part of the data of the A stage to train the classification model usually leads to unreliable performance of the model. Therefore, it is better to train the model in combination with the data of the A stage, so as to maximize the use of all data information of the two stages. Note that, unless otherwise specified, the matrix norm used in the present method is the Frobenius norm.
[0062] In one embodiment, the probability that i-n1 data in the training data belongs to the j-th class is:
[0063]
[0064] wherein, represents the original motion data, n1 represents the number of original motion data, represents the feature representation of the j-th column in the original motion data, represents the incremental motion data, represents the feature representation of the j-th column in the incremental motion data, C represents the classification category of the original motion data, and C+1 represents the classification category containing the incremental category, and respectively represent the feature representation of the k-th column in the original motion data and the incremental motion data, represents the probability that i-n1 data belongs to the j-th class; according to the probability that i-n1 data belongs to the j-th class, a cross-entropy loss function is constructed as:
[0065]
[0066] wherein, n2 represents the number of incremental motion data, is the true label of the i-n1 data belonging to the j-th class.
[0067] In one embodiment, according to the classifier trained on the original motion data and the classifier on the incremental motion data after the incremental category, a similarity constraint term is constructed as:
[0068]
[0069] wherein, represents the original motion data after the incremental category, represents the feature representation of the original motion data, represents the first C columns of the feature representation of the original motion data after the incremental category, wherein C represents the classification category.
[0070] Specifically, the model learned in the P phase To help learning the classification model W2 in the A phase, consider and are both classifiers defined on the old features, so they can inherit the historical information in W1 by constraining to be similar to However, since and have different sizes, it is difficult to measure their similarity directly in the form of a norm. It can be known that and should have similar classification effects on the first C classes of data on the old features, so the matrix composed of the first C columns can be defined as By approximating to , the historical information is inherited. There are two main approximation methods, which are and The first is a direct approximation method, and the second is an indirect approximation method. Regarding the second approximation method, it makes up for the shortcoming of the first method that does not consider the C+1th class. Specifically, the second method uses the data in the A phase while contains information of the C+1th class, so the information is inherited by approximating to in the classification results on . At the same time, uses the norm , which can impose sparse constraints on each row of the matrix, so it can satisfy both similarity and difference, that is, not only constrains to be consistent with in most classification results, but also allows small differences in some classification results. The most ideal case is that the rows of the classification results that are consistent correspond to the first C classes of samples, and the rows that are inconsistent correspond to the C+1th class of samples.
[0071] In one embodiment, the objective function SAFC-D of the direct incremental learning algorithm constructed according to the cross-entropy loss function, the regularization term, and the direct similarity constraint term is:
[0072]
[0073] wherein α and β represent coefficients, represents the regularization term.
[0074] In another embodiment, the objective function SAFC-ID of the indirect incremental learning algorithm is constructed according to a cross-entropy loss function, a regularization term, and an indirect similarity constraint term, and includes:
[0075]
[0076] wherein α and β represent coefficients, represents the regularization term.
[0077] Specifically, the gradient descent algorithm is used for solving, and first, for the algorithm SAFC-D, the objective function is represented as:
[0078]
[0079] Let J1(W) represents the first term of the above objective function. We first derive the gradient of the objective function with respect to W. Without loss of generality, we let Note that For any W pq (1≤p≤d,1≤q≤C+1), we have
[0080]
[0081]
[0082] Here, the reason why the second term is equal to 0 is that the gradient is not equal to 0 only when u=i. We have
[0083]
[0084] Substitute (8) into (7), and combine We have
[0085]
[0086] Since Substitute the above formula into (6), and we have
[0087]
[0088] For more concise description, we represent the gradient matrix as:
[0089]
[0090] Since the regularization term is only added to part of the variables, we decompose the matrix into
[0091]
[0092] wherein is column C+1 of A. Combining the above equation and the objective function, we have
[0093]
[0094] where denotes the sub-matrix consisting of the first d1 rows and the last column of A. Similarly, we have
[0095]
[0096]
[0097] where denotes the sub-matrix consisting of the first d1 rows and the last column of A. denotes the sub-matrix consisting of the last d2 rows of A. In summary, we have
[0098]
[0099] From this equation, we have the following update rule:
[0100]
[0101] where η is the update step size.
[0102] Second, for the algorithm SAFC-ID, its objective function is represented as:
[0103]
[0104] The gradients of the above function with respect to and are exactly the same as before, the only difference is the derivation of the gradient with respect to . First, for any matrix A, consider the definition of the norm, we have the following equation:
[0105]
[0106] where D is a diagonal matrix whose i-th diagonal element is
[0107]
[0108] where A i,· is the i-th row of matrix A.
[0109] Based on the above equation, we have
[0110]
[0111] Where D is the previously defined diagonal matrix, and the definition of A becomes...
[0112] Combining formulas (14), (15) and (21), we can obtain a gradient matrix in a form similar to that of formula (16), and the update rule is the same as that of formula (17).
[0113] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0114] In one embodiment, such as Figure 2 As shown, a motion type classification device based on an indirect incremental learning algorithm is provided, comprising: a data acquisition module 202, an initial training module 204, an incremental data construction module 206, an incremental training module 208, and a classification module 210, wherein:
[0115] The data acquisition module 202 is used to acquire raw motion data based on sensors installed on different parts of the moving object;
[0116] The initial training module 204 is used to train the original classifier based on the original data to obtain a trained original classifier.
[0117] The incremental data construction module 206 is used to acquire incremental motion data based on sensors installed on other parts of the moving object, and to construct a new classifier capable of processing incremental motion data based on the classifier trained on the original motion data and the incremental motion data; wherein, the incremental motion data includes added categories;
[0118] The incremental training module 208 is used to perform regularization processing on the new classifier capable of processing incremental motion data to obtain a regularization term; construct an indirect similarity constraint term based on the classifier trained on the original motion data and the new classifier trained on the incremental motion data after adding categories; and perform incremental learning on the trained original classifier based on the cross-entropy loss function, the regularization term, the objective function of the indirect incremental learning algorithm constructed by the indirect similarity constraint term, and the training samples to obtain a trained incremental classifier.
[0119] The classification module 210 is configured to perform motion type classification by using the trained incremental classifier.
[0120] In one embodiment, the incremental training module 208 is further configured to use the probability that i-n1 data in the training data belong to the jth class as:
[0121]
[0122] wherein, represents a feature representation of the motion data, and n1 represents the number of original motion data, represents the classifier parameters of the original motion data belonging to the jth class, represents a feature representation of the incremental motion data, represents the classifier parameters of the incremental motion data belonging to the jth class, C represents the classification categories of the original motion data, and C+1 represents the classification categories including the added category, and respectively represent the classifier parameters of the original motion data and the incremental motion data belonging to the kth class, represents the probability that i-n1 data belong to the jth class;
[0123] According to the probability that i-n1 data belong to the jth class, a cross-entropy loss function is constructed as:
[0124]
[0125] wherein, n2 represents the number of incremental motion data, is the true label of the i-n1 data belonging to the jth class.
[0126] In one embodiment, the incremental training module 208 is further configured to construct an indirect similarity constraint term according to the original motion data and the original motion data after adding the category as:
[0127]
[0128] wherein, represents a feature representation of the original motion data after adding the category, represents the classifier parameters of the original motion data, represents the classifier parameters of the first C columns of the original motion data after adding the category, wherein C represents the classification categories.
[0129] In one embodiment, the incremental training module 208 is further configured to construct a target function of the indirect incremental learning algorithm according to the cross-entropy loss function, the regular term, and the indirect similarity constraint term, including:
[0130]
[0131] wherein a, b represent coefficients, denotes a regularization term.
[0132] In one of the embodiments, the incremental training module 208 is further configured to solve the objective function by using a gradient descent algorithm.
[0133] The specific limitation of the motion type classification apparatus based on the indirect incremental learning algorithm can refer to the limitation of the motion type classification method based on the indirect incremental learning algorithm in the above, which will not be described here. Each module in the above motion type classification apparatus based on the indirect incremental learning algorithm can be realized by software, hardware and combination thereof in whole or in part. The above each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to be called and executed by the processor to perform the operation corresponding to each module.
[0134] In one embodiment, a computer device is provided, which can be a terminal, and its internal structure diagram can be as shown in Figure 3 The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a motion type classification method based on an indirect incremental learning algorithm. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.
[0135] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.
[0136] In one embodiment, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method in the above embodiments.
[0137] In one embodiment, a computer readable storage medium is provided, having stored thereon a computer program, which, when executed by a processor, implements the steps of the method of the above embodiments.
[0138] A person of ordinary skill in the art can understand that all or part of the processes in the above embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above embodiments. Any reference to memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0139] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.
[0140] The above embodiments only express several implementation manners of the present application, and the description is specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the patent of the present application should be subject to the appended claims.
Claims
1. A method of classifying a type of motion based on an indirect incremental learning algorithm, characterized by, The method comprises: obtaining original motion data according to sensors arranged on different parts of a motion object; training an original classifier according to the original motion data to obtain a trained original classifier; obtaining incremental motion data according to sensors arranged on other parts of the motion object, and constructing a new classifier capable of processing the incremental motion data according to the classifier trained according to the original motion data and the incremental motion data, wherein the incremental motion data contains an added category; regularizing the new classifier capable of processing the incremental motion data to obtain a regularization term, and constructing an indirect similarity constraint term according to the classifier trained according to the original motion data and the new classifier trained according to the incremental motion data after the category is added; performing incremental learning on the trained original classifier according to a target function of an indirect incremental learning algorithm constructed according to a cross-entropy loss function, the regularization term, the indirect similarity constraint term, and training samples to obtain a trained incremental classifier; performing motion type classification by using the trained incremental classifier. The indirect similarity constraint term constructed according to the classifier trained according to the original motion data and the new classifier trained according to the incremental motion data after the category is added further comprises: constructing the indirect similarity constraint term according to the original motion data and the original motion data after the category is added as: wherein, represents a feature representation of the original motion data after adding the class, represents a classifier parameter of the original motion data, represents a classifier parameter of the first C columns of the original motion data after adding the class, wherein C represents the class of the classification.
2. The method of claim 1, wherein, The step of constructing the cross-entropy loss function comprises: Utilizing the probability that a data in the training data belongs to the first class is: class is: in, Feature representation of motion data, Indicates the amount of raw motion data. This represents the classifier parameters belonging to the j-th column in the original motion data. The feature representation of the incremental motion data, This represents the classifier parameter belonging to the j-th column in the incremental motion data, where C represents the classification category of the original motion data, and C+1 represents the classification category containing the incremental category. and These represent the classifier parameters belonging to the k-th column in the original motion data and the incremental motion data, respectively. express The data belongs to the first The probability of a class; According to the described The probability that the data belongs to the first Class, the cross-entropy loss function is constructed as: wherein, denotes the number of incremental motion data, is the true label of the th data belongs to the th class.
3. The method of claim 1, wherein, constructing a target function of an indirect incremental learning algorithm according to a cross-entropy loss function, the regularization term, and the indirect similarity constraint term comprises: constructing a target function of an indirect incremental learning algorithm according to a cross-entropy loss function, the regularization term, and the indirect similarity constraint term comprises: wherein , denotes a coefficient, denotes a regularizer.
4. The method of claim 3, wherein, The method further comprises: solving the target function by using a gradient descent algorithm.
5. A motion type classification device based on an indirect incremental learning algorithm, characterized in that, The device comprises: a data acquisition module configured to obtain original motion data according to sensors arranged on different parts of a motion object; an initial training module configured to train an original classifier according to the original motion data to obtain a trained original classifier; an incremental data construction module configured to obtain incremental motion data according to sensors arranged on other parts of the motion object, and to construct a new classifier capable of processing the incremental motion data according to the classifier trained according to the original motion data and the incremental motion data, wherein the incremental motion data contains an added category; an incremental training module configured to regularize the new classifier capable of processing the incremental motion data to obtain a regularization term, to construct an indirect similarity constraint term according to the classifier trained according to the original motion data and the new classifier trained according to the incremental motion data after the category is added, and to perform incremental learning on the trained original classifier according to a target function of an indirect incremental learning algorithm constructed according to a cross-entropy loss function, the regularization term, and the indirect similarity constraint term, and training samples to obtain a trained incremental classifier; a classification module configured to perform motion type classification by using the trained incremental classifier. The incremental training module is further configured to construct an indirect similarity constraint term according to a classifier trained on the original motion data and a new classifier on the incremental motion data after adding the new category, and the indirect similarity constraint term is: wherein, represents a feature representation of the original motion data after adding the class, represents a classifier parameter of the original motion data, represents a classifier parameter of the first C columns of the original motion data after adding the class, wherein C represents the class of the classification.
6. The apparatus of claim 5, wherein, The incremental training module is also used to utilize the training data in... The data belongs to the first The probability of class is: wherein, represents a feature representation of motion data, represents a number of original motion data, represents a classifier parameter belonging to the j-th column in the original motion data, represents a feature representation of the incremental motion data, represents a classifier parameter belonging to the j-th column in the incremental motion data, C represents a classification category of the original motion data, and C+1 represents a classification category including an added category, and respectively represent a classifier parameter belonging to the k-th column in the original motion data and the incremental motion data, represents a probability that the data belongs to the C-th category, and a probability that the data belongs to the C-th category. According to the described The probability that the data belongs to the first class, the cross-entropy loss function is constructed as: wherein, denotes the number of incremental motion data, is the true label of the th data belongs to the th class. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor implements the steps of the method of any one of claims 1 to 4 when executing the computer program.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 4.