Apparatus and method for training model for human recognition

Abstract

An apparatus for training a model for human recognition and a method thereof may include a primary training module configured to perform primary training of the model with respect to a pre-prepared source data set; a target subset generation module configured to generate a target subset by selecting some cameras from a plurality of cameras included on the service robot; a feature vector extraction module configured to extract feature vectors of the target subset by using the model; the marking module is configured to mark the feature vector; and a secondary training module configured to perform secondary training on the model for a target data set by using a result of the marking.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of Korean Patent Application No. 10-2023-0130814 filed in the Korean Intellectual Property Office on September 27, 2023, the entire contents of which are incorporated herein by reference. Technical Field

[0003] The present disclosure relates to an apparatus and method for training a model for human recognition. Background Art

[0004] Unlike industrial robots used in factories, service robots are robots designed to perform specific tasks or provide services and can be used in various fields. For example, service robots can include household robots that provide various services such as home cleaning, medical service robots that help patients receive treatment in the medical field, service robots that serve food, guidance and consulting robots that guide visitors or provide information, security robots responsible for the safety of specific areas or buildings, support robots that support the daily lives of people with disabilities or the elderly, etc. Service robots operate autonomously and are not directly controlled by humans, but interaction with humans is usually important, and it is often necessary to detect and locate humans.

[0005] To this end, service robots can adopt multi-camera systems. Multi-camera systems can allow service robots to recognize three-dimensional environments and recognize humans from different angles. In particular, multi-camera systems can also recognize areas that a single camera may not be able to recognize, and integrate and process multiple data collected from multiple cameras, so that recognition accuracy can be improved, and the environment and humans can be recognized in three-dimensional space, allowing more accurate positioning and interaction. However, although multi-camera systems improve the human recognition capabilities of service robots, an efficient algorithm is required as the amount of data to be processed and the amount of calculation increase. In addition, in multi-camera systems, human re-identification (ReID) is essential, which can recognize humans detected in the field of view of one camera again in the field of view of another camera. Summary of the invention

[0006] The present disclosure relates to an apparatus and method for training a model for human recognition.

[0007] Some embodiments of the present disclosure may provide an apparatus and method for training a human recognition model, which can improve the human recognition and re-recognition performance of a service robot equipped with a multi-camera system.

[0008] According to one embodiment, an apparatus for training a human recognition model may include: one or more processors; and a storage medium storing computer-readable instructions, which, when executed by the one or more processors, enable the one or more processors to provide: a primary training module, which is configured to: perform primary training on the model for a pre-prepared source data set; a target subset generation module, which is configured to: generate a target subset by selecting some cameras from a plurality of cameras installed on a service robot; a feature vector extraction module, which is configured to: extract feature vectors of the target subset by using the model; a labeling module, which is configured to: label the feature vectors; and a secondary training module, which is configured to: perform secondary training on the model for the target data set by using the labeled results.

[0009] In some embodiments, the instructions further enable the one or more processors to have a feature vector clustering module configured to: determine similarity of the feature vectors, and cluster the feature vectors according to the similarity.

[0010] In some embodiments, the feature vector clustering module may further cluster the feature vectors by using a density-based spatial clustering of applications with noise (DBSCAN) technique.

[0011] In some embodiments, the labeling module may also perform pseudo labeling on each cluster clustered by the feature vector clustering module.

[0012] In some embodiments, the instructions further enable the one or more processors to have a weight assignment module configured to assign a weight to each cluster clustered by the feature vector clustering module.

[0013] In some embodiments, the weight assignment module may also assign a relatively higher weight to a given cluster that is determined to have a relatively higher diversity.

[0014] In some embodiments, the secondary training module may also gradually perform training by increasing the reflectance of a given cluster assigned a relatively higher weight.

[0015] In some embodiments, the instructions further enable one or more processors to have a repetition module configured to repeat the target subset generation module, the feature vector extraction module, the feature vector clustering module and the marking module until a selected, set or predetermined condition is met.

[0016] In some embodiments, the instructions also enable one or more processors to have a course sequence generation module, which is configured to generate a course sequence for training the model by using the labeled results repeatedly generated by the repetition module, and the secondary training module can also perform course learning on the model according to the course sequence.

[0017] In some embodiments, the target subset generation module may further generate the target subset by preferentially selecting a given camera from a plurality of cameras, wherein the given camera generates a value having a relatively small difference from the source data set.

[0018] According to one embodiment, a method for training a model for human recognition may include the following steps: primary training of the model for a pre-prepared source data set; generating a target subset by selecting some cameras from a plurality of cameras included on a service robot; extracting feature vectors of the target subset by using the model; labeling the feature vectors; and secondary training the model for the target subset by using the labeling result.

[0019] In some embodiments, the method may further include determining similarity of the feature vectors, and clustering the feature vectors according to the similarity.

[0020] In some embodiments, clustering of the feature vectors may include clustering the feature vectors using a density-based spatial clustering of applications with noise (DBSCAN) technique.

[0021] In some embodiments, labeling may include performing pseudo-labeling on each cluster.

[0022] In some embodiments, the method may further include assigning a weight to each clustered cluster.

[0023] In some embodiments, assigning weights may include assigning a relatively higher weight to a given cluster determined to have a relatively higher diversity.

[0024] In some embodiments, the secondary training may include gradually performing the training by increasing the reflectance of a given cluster to which a relatively higher weight is assigned.

[0025] In some embodiments, the method may further include: repeatedly generating a target subset, extracting feature vectors, clustering the feature vectors, and marking the feature vectors until a set condition is met.

[0026] In some embodiments, the method may further include generating a course sequence for training the model by using repeatedly generated labeling results, and wherein the secondary training may include performing course learning on the model according to the course sequence.

[0027] In some embodiments, generating the target subset may include generating the target subset by preferentially selecting a given camera from a plurality of cameras, wherein the given camera produces values ​​having relatively small differences from the source data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 and Figure 2 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0029] Figure 3 is a schematic diagram showing a model training method for human recognition according to an embodiment of the present disclosure.

[0030] Figure 4 and Figure 5 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0031] Figure 6 and Figure 7 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0032] Figure 8 is a schematic diagram of a computing device for explaining one embodiment of the present disclosure. DETAILED DESCRIPTION

[0033] The exemplary embodiments of the present disclosure will be described in detail below in conjunction with the accompanying drawings so that a person of ordinary skill in the art can easily implement the present disclosure. However, the embodiments of the present disclosure may be implemented in a variety of different forms and are not limited to the exemplary embodiments described herein. In order to clearly explain the present disclosure in the accompanying drawings, parts that are not related to the description may be omitted, and the same reference numerals may represent the same elements throughout the specification. .

[0034] Throughout the specification and claims, unless explicitly stated to the contrary, the word "comprise" and variations such as "include" or "comprising" may be understood to imply the inclusion of the stated elements but not the exclusion of any other elements. Terms including ordinal numbers, such as "first", "second", etc., may be used to describe various elements, but these elements are not necessarily limited by these terms. These terms may only be used to distinguish one element from another.

[0035] The terms such as “-part”, “-group” and “module” described in this specification may refer to a unit that processes at least one function or operation described in this specification, and the unit may be implemented as hardware or software or a combination of hardware and software.

[0036] Figure 1 and Figure 2 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0037] Human recognition through cameras is a major technical factor for service robots that have many continuous interactions with humans. With the development of computer vision and deep learning technologies, human recognition has received increasing attention. Human recognition may include various detailed technologies, such as: facial recognition, which recognizes individuals by analyzing facial features; posture and motion recognition, which recognizes and analyzes human postures or movements; person re-identification, which recognizes the same person in the field of view of multiple cameras; behavior recognition, which classifies or predicts human behavior in video sequences; person segmentation, which separates human outlines from images or videos, and so on. The performance of these technologies depends largely on the learning data or learning methods, and therefore, the learning technology of the model for human recognition is very important. A model training device 1 for human recognition according to one embodiment may include the configurations described in the following various embodiments to improve the performance of human recognition and re-identification of a service robot equipped with a multi-camera system. Reference Figure 1 The model training device 1 for human identification may include a primary training module 11, a target subset generation module 12, a feature vector extraction module 13, a feature vector clustering module 14, a labeling module 15 and a secondary training module 16, any combination or all of which may be multiple and may also include multiple components thereof.

[0038] The primary training module 11 can perform initial training on the model for a pre-prepared source data set. The model can be a model for performing recognition and re-recognition of humans of a service robot equipped with a multi-camera system. The source data set can be data used for initial training of the model, and can generally be configured as data including a large amount of and diverse information. For example, a large image data set (such as ImageNet) can be used as a source data set. The model can learn relatively rich features by learning from the source data set. In other words, the primary training module 11 can also pre-train (or pre-train) the model. The features learned from the source data set may also be useful for the actual target task, namely the recognition and re-recognition task of humans of a service robot equipped with a multi-camera system, and may allow faster secondary learning of a target data set of relatively small and task-specific information.

[0039] The target subset generation module 12 can generate a target subset by selecting some cameras from a plurality of cameras included in the multi-camera installed on the service robot. The target data set can be data for performing actual target tasks using the model pre-trained by the primary training module 11, such as human recognition and re-recognition tasks of a service robot equipped with a multi-camera system, and the target subset can refer to a set of data indicating a specific part or category in the target data set. Specifically, the target subset in the specification can refer to a set of images obtained primarily from a camera to ultimately generate a target data set used by the secondary training module 16, which will be described below. The target subset can be determined as a target data set used by the secondary training module 16 through subsequent operations (such as feature vector extraction and labeling), which will be described below.

[0040] When target subsets are generated for all multiple cameras included in a multi-camera system at the same time, it may be difficult to find commonality due to high complexity. When target subsets are generated for all cameras without considering the difference from the source data set, the learning performance may be insufficient. On the other hand, when target subsets are generated for cameras that differ greatly from the source data set, the learning performance may also deteriorate. In order to solve such performance problems, in some embodiments, the target subset generation module 12 can generate a target subset by preferentially selecting cameras with smaller differences from the source data set from multiple cameras in the multi-camera system.

[0041] The multi-camera system may include a first camera Cam 0, a second camera Cam 1, a third camera Cam 2, and a fourth camera Cam 3. Figure 2 (a), the feature vector can be represented in the embedding space, the feature vector named "source" can represent the source data set, and the feature vectors named "Cam 0", "Cam 1", "Cam 2" and "Cam 3" can represent the target data sets corresponding to the first camera Cam 0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3 respectively. In the model training device 1 for human recognition according to one embodiment, the domain conversion from the source data set "source" to the target data sets "Cam 0", "Cam 1", "Cam 2" and "Cam 3" can be performed, and when the data corresponding to the camera with the smaller difference from the source data set among the first camera Cam 0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3 is preferentially selected as the target subset, the target data set can also be configured as the data on which the model can be trained to ensure better or better learning performance.

[0042] In some embodiments, the difference between the source data set and the camera data to be generated as the target subset can be quantified by measuring the distribution difference. For example, the target subset generation module 12 can measure the difference between the distribution of the source data set and the distribution of the camera data to be generated as the target subset by using the maximum mean difference (MMD) technique, and for example, the difference between the two distributions can be calculated by comparing the means between the samples of the two distributions. The specific calculation equation may include, for example, an equation for the distance between the expected values ​​of the two distributions in a space represented as a function of the reproducing kernel Hilbert space (RKHS), but the scope of the present disclosure is not limited to a specific equation, and may include the use of various equations that can measure the difference between the two distributions. The target subset generation module 12 can calculate the MMD value of each of the first camera Cam 0, the second camera Cam 1, the third camera Cam 2, and the fourth camera Cam 3, and regard the data corresponding to the data of the camera with the smallest MMD value as the data with the smallest difference from the source data set, and preferentially select the data as the target subset. For example, the target subset generation module 12 may calculate the distribution difference between the source data set and each of the first camera Cam 0, the second camera Cam 1, the third camera Cam 2, and the fourth camera Cam 3, and generate the target subset corresponding to the first camera Cam 0 with the smallest distribution difference as the target subset. Then, the generated target subset may be subjected to feature vector extraction and labeling.

[0043] In some embodiments, the target subset generation module 12 may repeatedly generate target subsets for progressive learning in several stages. Figure 2(b), as a result of calculating the distribution difference between the source data set and each of the first camera Cam 0, the second camera Cam 1, the third camera Cam 2, and the fourth camera Cam 3 in the first stage, the target subset generation module 12 can generate a first target subset corresponding to the first camera Cam 0 having the smallest distribution difference with the source data set, and perform feature vector extraction and labeling on the generated first target subset. In a second stage after the first stage, the target subset generation module 12 can then generate a second target subset corresponding to the second camera Cam 1, which has a smaller distribution difference with the source data set, and perform feature vector extraction and labeling on the first target subset and the generated second target subset. In a third stage after the second stage, the target subset generation module 12 can then generate a third target subset corresponding to the third camera Cam 2, which has a smaller distribution difference with the source data set, and perform feature vector extraction and labeling on the first and second target subsets and the generated third target subset. In the fourth stage after the third stage, the target subset generation module 12 can generate a fourth target subset corresponding to the fourth camera Cam 3, which has the largest distribution difference with the source data set, and perform feature vector extraction and labeling on the first to third target subsets and the generated fourth target subset.

[0044] The feature vector extraction module 13 can extract the feature vector of the target subset generated by the target subset generation module 12 by using the model. The feature vector extraction module 13 can extract information representing the data included in the target subset in the form of a vector. The extraction of the feature vector can be performed manually, but a deep learning method that directly learns features from data can also be used.

[0045] The feature vector clustering module 14 can determine the similarity of the feature vectors extracted by the feature vector extraction module 13, and cluster the feature vectors according to the similarity. The feature vector clustering module 14 can cluster the feature vectors by using, for example, a density-based spatial clustering of applications with noise (DBSCAN) technique. DBSCAN is a density-based clustering algorithm that works by treating high-density areas (i.e., areas where data points are close together) as a cluster and processing low-density areas through noise. When DBSCAN is applied to the feature vectors extracted by the feature vector extraction module 13, similar feature vectors can be close to each other, and therefore, these feature vectors can be clustered and treated as a cluster.

[0046] For example, the feature vector clustering module 14 may set a maximum distance between one feature vector and another feature vector, and set a minimum number of feature vectors that exist within the maximum distance relative to one feature vector. When there are more than the minimum number of feature vectors within the maximum distance from the selected feature vector, the feature vector clustering module 14 may select a feature vector and generate a new cluster. In addition, when the selected feature vector has more than the minimum number of adjacent feature vectors within the maximum distance, the feature vector clustering module 14 may: form clusters based on similarities between feature vectors by repeatedly adding all adjacent feature vectors of the selected feature vector within the maximum distance range to the cluster, or adding the selected feature vector to the cluster when the selected feature vector is within the maximum range but has fewer feature vectors than the minimum number.

[0047] refer to Figure 2 (c), the feature vector clustering module 14 can generate: a first cluster "Cluster 0", which is derived from the data corresponding to the first camera Cam 0, the second camera Cam 1 and the third camera Cam 2, and corresponds to a set of feature vectors estimated to be human; a second cluster "Cluster 1", which is derived from the data corresponding to the third camera Cam2 and the fourth camera Cam 3 and corresponds to a set of feature vectors estimated to be another person; and a third cluster "Cluster 2", which is derived from the data corresponding to the second camera Cam1 and corresponds to a set of feature vectors estimated to be another person. The first camera Cam 0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3 may be factors included in the multi-camera system of the service robot that shoot the same person at different positions and angles. Therefore, the first cluster "Cluster 0" corresponding to the first camera Cam 0, the second camera Cam 1 and the third camera Cam 2 may have the highest diversity, and the second cluster "Cluster 1" corresponding to the third camera Cam 2 and the fourth camera Cam 3 may have the second highest diversity; the third cluster "Cluster 2" corresponding only to the second camera Cam 1 may have the lowest diversity. In other words, the first cluster "Cluster 0" includes various backgrounds, poses, and clothing, and may have relatively more learning content, while the third cluster "Cluster 2" is likely to include the same backgrounds, poses, and clothing, and may have relatively less learning content. Regarding the training of human recognition models, this diversity of clusters can be additionally considered, which will be referred to below. Figure 4 and Figure 5 Give a description.

[0048] The labeling module 15 can perform pseudo labeling on each feature vector extracted by the feature vector extraction module 13 or each cluster clustered by the feature vector clustering module 14. Pseudo labeling can assign the label predicted by the model to the unlabeled data as a "pseudo label", and the pseudo labeled data generated as described above can be optionally used for model training together with the original labeled data.

[0049] The secondary training module 16 may perform secondary training on the model for the target dataset by using the result of the labeling performed by the labeling module 15 .

[0050] As described above, as a data set designed to maximize the training effect by considering the detailed characteristics of the data obtained by multiple cameras included in the multi-camera system of the service robot, by applying the above-mentioned training technology, it can be expected to obtain excellent training effects while reducing the large amount of computing time and cost that may occur in unsupervised domain adaptation. In addition, the performance of human recognition and re-recognition of the installed service robot can be improved. Specifically, as an example implementation, when the model trained according to the embodiment of the present disclosure is used as the dataset of Market-1501 and DukeMTMC, and the mean average precision (mAP) and Rank-1 accuracy are used as evaluation indicators, a performance improvement effect of about 30% p (percentage points) is obtained, as shown below. .

[0051] (Table 1)

[0052]

[0053] Figure 3 is a schematic diagram showing a model training method for human recognition according to an embodiment of the present disclosure.

[0054] refer to Figure 3 According to one embodiment, a model training method for human recognition may include: primary training of a model for a pre-prepared source data set (operation S301); generating a target subset by selecting some cameras from a plurality of cameras included in a multi-camera installed on a service robot (operation S302); extracting a feature vector of the target subset by using the model (operation S303); performing labeling on the feature vector (operation S304); and performing secondary training of the model for the target data set by using the labeling result (operation S305). About Figure 3 For more detailed information on the example methods, refer to the or application instructions Figure 1 to Figure 2 and Figures 4 to 8 The description of the embodiments described herein is as follows, therefore, redundant descriptions thereof may be omitted here. For example, Figure 3 The method embodiments described in the Figure 1 to Figure 2 and Figures 4 to 8 This is implemented in accordance with an exemplary embodiment of the present invention.

[0055] Figure 4 and Figure 5 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0056] refer to Figure 4According to an embodiment, the model training device 2 for human recognition may include a primary training module 21, a target subset generation module 22, a feature vector extraction module 23, a feature vector clustering module 24, a labeling module 25, a weight assignment module 26, and a secondary training module 27, any combination or all of which may be multiple, or may include multiple components thereof. For the primary training module 21, the target subset generation module 22, the feature vector extraction module 23, the feature vector clustering module 24, the labeling module 25, and the secondary training module 27, in the absence of contradiction, the above reference may be referred to or applied. Figure 1 The description of the primary training module 11, the target subset generation module 12, the feature vector extraction module 13, the feature vector clustering module 14, the labeling module 15 and the secondary training module 16 are given, and therefore, in reference Figure 4 and Figure 5 While mainly describing the weight allocation module 26.

[0057] The weight assignment module 26 may assign a weight to each cluster clustered by the feature vector clustering module 24. Specifically, the weight assignment module 26 may assign a high weight to a cluster determined to have high diversity. The secondary training module 27 may gradually perform training by increasing the reflectance of the clusters assigned with higher weights.

[0058] Reference together Figure 5 , the feature vector clustering module 24 can generate: a first cluster "cluster a" and a second cluster "cluster b", the first cluster "cluster a" is derived from the data corresponding to the first camera Cam 0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3, and corresponds to a set of feature vectors estimated to be human: the second cluster "cluster b" is derived from the data corresponding to the second camera Cam 1 and the third camera Cam 2 and corresponds to a set of feature vectors estimated to be another person. The first camera Cam0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3 can serve the factor of shooting the same person at different positions and angles included in the multi-camera system of the robot, so the diversity of the first cluster "cluster a" corresponding to the first camera Cam 0, the second camera Cam 1, the third camera Cam 2 and the fourth camera Cam 3 may be higher than the diversity of the second cluster "cluster b" corresponding to the second camera Cam 1 and the third camera Cam 2. In other words, for this example, the first cluster "cluster a" includes more diverse backgrounds, postures and clothing than the second cluster "cluster b", and there may be relatively more content to learn.

[0059] The weight assignment module 26 may assign a relatively high weight (eg, w a =1.11), and assign a relatively low weight (e.g., wb =0.25), the second training module 27 can gradually perform training by increasing the reflectance of the first cluster "cluster a" assigned a relatively high weight compared to the second cluster "cluster b". Accordingly, with respect to the training of the model for human recognition, the model training can be mainly performed based on data with more learning content, and therefore, better or more outstanding training effects can be expected, while training can be performed with less data.

[0060] In some embodiments, for example, the weight assignment module 26 can calculate the weight according to the entropy value of information theory. The entropy of information theory can be used to measure the uncertainty of random variables and can be calculated by considering the probability of possible events and the amount of information of the corresponding events. The weight can be calculated according to a selected, set or predetermined equation in the entropy value. For example, in a simple method, the weight can be calculated by dividing the entropy value of each of the first cluster "cluster a" and the second cluster "cluster b" by the entropy value of all clusters. For example, the weight w a The entropy value H of the first cluster "cluster a" can be a = 2.03 divided by the entropy of all clusters, the weight w b The entropy value H of the second cluster "cluster b" can be b = 0.28 divided by the entropy values ​​of all clusters. The method of calculating the weight by using the entropy value is not limited to the above-described exemplary method.

[0061] Figure 6 and Figure 7 is a schematic diagram showing a model training device for human recognition according to an embodiment of the present disclosure.

[0062] refer to Figure 6 According to an embodiment, the model training device 3 for human identification may include a primary training module 31, a target subset generation module 32, a feature vector extraction module 33, a feature vector clustering module 34, a labeling module 35, a weight allocation module 36, a repetition module 37, a course sequence generation module 38 and a secondary training module 39, any combination or all of which may be multiple, or may include multiple components thereof. For the primary training module 31, the target subset generation module 32, the feature vector extraction module 33, the feature vector clustering module 34, the labeling module 35, the weight allocation module 36 and the secondary training module 39, in the absence of contradiction, the above reference may be referred to or applied. Figure 1 and Figure 4 The description of the primary training modules 11 and 21, the target subset generator modules 12 and 22, the feature vector extraction modules 13 and 23, the feature vector clustering modules 14 and 24, the labeling modules 15 and 25, the weight assignment module 26 and the secondary training modules 16 and 27 given above will therefore be referred to. Figure 6 and 7The repetition module 37 and the course sequence generation module 38 are mainly described.

[0063] The repetition module 37 can repeat the target subset generation module 32, the feature vector extraction module 33, the feature vector clustering module 34 and the labeling module 35 until the selected, set or predetermined conditions are met. In addition, the course sequence generation module 38 can generate a course sequence for training the model by using the labeling results repeatedly generated by the repetition module 37. The secondary training module 39 can perform course learning on the model according to the course sequence. Course learning can be an artificial intelligence learning strategy for solving complex problems, or it can refer to a technique that starts with simple problems and gradually moves on to more difficult problems. Specifically, course learning in the example embodiment can refer to starting learning from data that has a high similarity to the source data set, and gradually continuing learning with more complex and more general data after the learning is stable.

[0064] Also refer to Figure 7 , the repetition module 37 can repeatedly generate target subsets for progressive training. In the first repetition, as a result of calculating the distribution difference between the source data set and each of the multiple cameras Cam 0 to Cam C (C is a non-negative integer), in this example, the target subset generation module 32 can generate a first target subset corresponding to the camera Cam 4 having the smallest distribution difference with the source data set, and the results marked by the feature vector extraction module 33, the feature vector clustering module 34 and the marking module 35 can be added to the course sequence. Next, in the second repetition, in this example, the target subset generation module 32 can generate a second target subset corresponding to the camera Cam1 having a relatively small distribution difference with the source data set, and the results marked by the feature vector extraction module 33, the feature vector clustering module 34 and the marking module 35 can be added to the course sequence. Next, in the third repetition, in this example, the target subset generation module 32 can then generate a third target subset corresponding to the camera Cam 0 having a relatively small distribution difference with the source data set, and the results marked by the feature vector extraction module 33, the feature vector clustering module 34 and the marking module 35 can be added to the course sequence. Finally, in the nth repetition (n is a natural number), the target subset generation module 32 can generate the nth target subset corresponding to the camera Cam C having the largest distribution difference relative to the source data set, and can add the results marked by the feature vector extraction module 33, the feature vector clustering module 34 and the marking module 35 to the course sequence. The secondary training module 39 can perform course learning on the model according to the course sequence generated based on the first to nth target subsets.

[0065] Figure 8 is a schematic diagram of a computing device for explaining one embodiment of the present disclosure.

[0066] refer to Figure 8 , the example model training apparatus and method for human recognition according to the example embodiment may be implemented by using a computing device 50 .

[0067] The computing device 50 may include at least one of a processor 510, a memory 530, a user interface input device 540, a user interface output device 550, and a storage device 560, any combination or all of which may be multiple, or may include multiple components thereof, which communicate via a bus 520. The computing device 50 may also include a network interface 570 electrically connected to the network 40. The network interface 570 may send signals to or receive signals from other entities through the network 40.

[0068] The processor 510 may be implemented in various types, such as a microcontroller unit (MCU), an application processor (AP), a central processing unit (CPU), a graphics processing unit (GPU), a neural processing unit (NPU), and a quantum processing unit (QPU), and may be any semiconductor device that executes commands stored in the memory 530 or the storage device 560. The processor 510 may be configured to implement the above-mentioned Figures 1 to 7 Describes the functions and methods.

[0069] The storage medium may include a memory 530 and a storage device 560, and the storage device 560 may include various types of volatile or non-volatile storage media. For example, the memory 530 may include a read-only memory (ROM) 531 and a random access memory (RAM) 532. In this embodiment, the memory 530 may be located inside or outside the processor 510, and the memory 530 may be connected to the processor 510 through various known implementations.

[0070] In some embodiments, at least some components or functions of the model training apparatus and method for human identification according to example embodiments may be implemented as a program or software running on a computing device 50, and the program or software may be stored on a computer-readable medium. Specifically, a computer-readable medium according to an embodiment may be a computer including a processor 510 that executes a program or command stored in a memory 530 or a storage device 560, and a program for executing the steps included in the model training apparatus and method for human identification according to an embodiment may be recorded thereon.

[0071] In some embodiments, at least some components or functions of the model training apparatus and method for human recognition according to example embodiments may be implemented using hardware or circuits of the computing device 50, or may be implemented as separate hardware or circuits that may be electrically connected to the computing device 50.

[0072] According to the above-mentioned exemplary embodiment, as a data set that maximizes the training effect by considering the detailed features of the data obtained from the multiple cameras included in the multi-camera system of the service robot, by applying the above-mentioned training technology, better or more excellent training effects can be expected while reducing the huge computing time and cost that may occur in unsupervised domain adaptation. In addition, using the embodiments of the present disclosure, the human recognition and re-recognition performance of the installed service robot can be improved.

[0073] Although example embodiments of the present disclosure have been described in detail above, the scope of the present disclosure is not limited thereto, and various modifications and improvements made by those of ordinary skill in the art to which the present disclosure belongs, including their equivalents, may also fall within the scope of the present disclosure.

Claims

1. A device for training a model for human recognition, the device comprising: one or more processors; as well as A storage medium storing computer readable instructions that, when executed by one or more processors, cause the one or more processors to provide: The primary training module is configured to: perform primary training on the model for a pre-prepared source data set; The target subset generation module is configured to: generate a target subset by selecting some cameras from a plurality of cameras installed on the service robot; A feature vector extraction module is configured to: extract a feature vector of the target subset by using the model; A marking module is configured to: mark the feature vector; as well as The secondary training module is configured to perform secondary training on the model by using the labeled results for a target data set.

2. The device according to claim 1, wherein: The instructions further enable the one or more processors to have a feature vector clustering module configured to determine similarities of the feature vectors and cluster the feature vectors according to the similarities.

3. The device according to claim 2, wherein: The feature vector clustering module is further configured to cluster the feature vectors by using a density-based spatial clustering of applications with noise (DBSCAN) technique.

4. The device according to claim 2, wherein: The labeling module is further configured to perform pseudo labeling on each cluster clustered by the feature vector clustering module.

5. The device according to claim 2, wherein: The instructions further enable the one or more processors to have a weight assignment module configured to assign a weight to each cluster clustered by the feature clustering module. 6 . The apparatus according to claim 5 , wherein the weight assignment module is further configured to assign a relatively higher weight to a given cluster determined to have a relatively higher diversity. 7 . The apparatus according to claim 5 , wherein the secondary training module is further configured to: perform the training step by step by increasing the reflectance of a given cluster to which a relatively higher weight is assigned.

8. The apparatus according to claim 2, wherein the instruction further enables the one or more processors to have a repetition module, wherein the repetition module is configured to: repeat the target subset generation module, the feature vector extraction module, the feature vector clustering module and the labeling module until a set condition is met.

9. The apparatus according to claim 8, wherein the instructions further enable the one or more processors to have a course sequence generation module, the course sequence generation module being configured to: generate a course sequence for training the model by using the labeling results repeatedly generated by the repetition module, and in, The secondary training module is also configured to perform course learning on the model according to the course sequence.

10. The device according to claim 1, wherein: The target subset generation module is further configured to generate the target subset by preferentially selecting a given camera from the plurality of cameras, wherein the given camera generates a value having a relatively small difference from the source data set.

11. A method for training a model for human recognition, the method comprising the following steps: Perform preliminary training of the model on the pre-prepared source dataset; generating a target subset by selecting some cameras from a plurality of cameras included on the service robot; extracting a feature vector of the target subset by using the model; marking the feature vector; as well as The model is retrained on the target subset by using the labeled results.

12. The method according to claim 11, further comprising the steps of: determining the similarity of the feature vectors; and The feature vectors are clustered according to the similarities.

13. The method according to claim 12, wherein: Clustering the feature vectors includes clustering the feature vectors by using a density-based spatial clustering of applications with noise (DBSCAN) technique.

14. The method according to claim 12, wherein: The labeling includes: performing pseudo labeling on each cluster of the clusters.

15. The method according to claim 12, further comprising: Assign weights to each cluster.

16. The method according to claim 15, wherein: Assigning weights includes assigning a relatively higher weight to a given cluster determined to have a relatively higher diversity.

17. The method according to claim 15, wherein: The secondary training includes performing training step by step by increasing the reflectance of a given cluster to which a relatively higher weight is assigned.

18. The method according to claim 12, further comprising: The target subset is generated, the feature vector is extracted, the feature vector is clustered, and the feature vector is marked repeatedly until a set condition is met.

19. The method according to claim 18, further comprising: A course sequence for training the model is generated by using the labeling results generated by repetition, and wherein the secondary training includes performing course learning on the model according to the course sequence.

20. The method according to claim 11, wherein: Generating the target subset includes generating the target subset by preferentially selecting a given camera from the plurality of cameras, wherein the given camera produces values ​​having relatively small differences from the source data set.

Images