WO2023123923A1 - Human body weight identification method, human body weight identification device, computer device, and medium - Google Patents

Abstract

The present application provides a human body weight identification method, a human body weight identification device, a computer device, and a medium. The human body weight identification method comprises: acquiring a human body image to be evaluated; extracting a first human body feature of the human body image by means of a pre-trained human body weight identification model, the human body weight identification model being trained on the basis of human body training images captured by a single camera; comparing the first human body feature with a plurality of human body features in a preconfigured database, and determining a second human body feature having a highest similarity measure to the first human body feature; and when the similarity measure is greater than or equal to a specified threshold, causing a human body category corresponding to the second human body feature to be matched as a most similar human body category to the human body image. According to the present application, the human body weight identification and evaluation is performed on the basis of the human body weight identification model trained via a single camera, facilitating matching a most similar human body category for the human body image to be evaluated among the plurality of human body features in the preset database, and consequently reducing the occurrence of misidentification.

Description

Human body weight recognition method, human body weight recognition device, computer equipment and medium

This application claims the priority of the Chinese patent application with the application number 202111682643.5 and the application name "Human body weight recognition method, human body weight recognition device, computer equipment and media" submitted to the China Patent Office on December 30, 2021, the entire content of which Incorporated in this application by reference.

technical field

The present application relates to the technical field of computer vision, in particular to a method for body weight recognition, a body weight recognition device, computer equipment and media.

Background technique

Human body re-identification technology is a technology that uses computer vision technology to perform non-overlapping camera retrieval on a given image, thereby identifying images belonging to the same human body. With the development and progress of artificial intelligence technology, human body recognition technology can track, match and identify the same human body across time and space. It has been widely used in all aspects of social life and has become a research hotspot in the field of computer vision in recent years. one.

In related technologies, when training a human body re-identification model based on a deep learning model, the deep learning model is used to learn the relationship between human bodies to extract features and search, and the performance of the model depends heavily on the quality and quantity of data. However, the training data used to train the human body weight recognition model needs to label pedestrian relationships across cameras, which leads to high training costs and difficulties for the human body weight recognition model.

application content

The technical problem to be solved in this application is to overcome the defects in the prior art that rely too much on data quality and quantity when using the human body weight recognition model to perform human body weight recognition, so as to provide a human body weight recognition method, a human body weight recognition device, computer equipment and medium.

In a first aspect, the present application provides a method for human body weight recognition, the method comprising: acquiring an image of a human body to be detected. The first human body feature of the human body image is extracted through a pre-trained human body weight recognition model, and the human body weight recognition model is trained based on a human body training image collected by a single camera. Comparing the first human body feature with a plurality of human body features in a preset database, and determining a second human body feature with the highest similarity to the first human body feature. When the similarity is greater than or equal to a specified threshold, the human body category corresponding to the second human body feature is matched to the human body category most similar to the human body image.

In this way, when performing human body weight recognition detection on the human body image to be detected, the human body weight recognition model that can be trained based on the human body training image collected by a single camera can be used among the multiple human body features in the preset database as the The detected human body images are matched to the most similar human body category, thereby improving the accuracy of human body category recognition and reducing the occurrence of misidentification.

In a second aspect, the present application also provides a device for body weight recognition, which includes: an acquisition unit configured to acquire an image of a human body to be detected. The extraction unit is configured to extract the first human body feature of the human body image through a pre-trained human body weight recognition model, and the human body weight recognition model is trained based on human body training images collected by a single camera. A comparing unit, configured to compare the first human body feature with a plurality of human body features in a preset database, and determine a second human body feature with the highest similarity to the first human body feature. A determining unit, configured to match the human body category corresponding to the second human body feature to the human body category most similar to the human body image when the similarity is greater than or equal to a specified threshold.

With reference to the second aspect, in the first embodiment of the second aspect, the human body re-identification model is trained based on the human body training images collected by a single camera using the following unit: a first acquisition unit, configured to acquire multiple sets of first human body In the training image set, the first human body training image sets of different groups are collected by different cameras, and the human body types included in different first human body training image sets are the same. The first training unit is configured to input each of the first human body training image sets into the feature extraction network to be trained, train the feature extraction network to extract human body features corresponding to each of the human body categories, and obtain the first feature extraction Model. The second obtaining unit is used to obtain a second human body training image set and a human body training feature set that are the same as the human body category included in the first human body training image set, and the multiple human body training features included in the human body training feature set are each The set of group human body images corresponds to the overall human body characteristics of the first human body training images. The second training unit is configured to input the second human body training image set into the first feature extraction model, train the first feature extraction model according to the human body training feature set, and obtain the human body weight recognition model.

According to the third aspect, the embodiment of the present application further provides a computer device, including a memory and a processor, the memory and the processor are connected to each other in communication, the memory stores computer instructions, and the processor passes Execute the computer instructions, so as to execute the human body weight recognition method in any one of the first aspect and its optional implementations.

According to the fourth aspect, the embodiments of the present application further provide a computer-readable storage medium, the computer-readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the first aspect and optional implementations thereof A human body weight recognition method in any one of the ways.

Description of drawings

The drawings that need to be used in the embodiments of the present application will be introduced below.

Fig. 1 is a flow chart of a method for body weight recognition proposed according to an exemplary embodiment.

Fig. 2 is a flowchart of a method for training a human body weight recognition model according to an exemplary embodiment.

Fig. 3 is a flowchart of a method for training a first feature extraction model according to an exemplary embodiment.

Fig. 4 is a flow chart of another training method for a first feature extraction model proposed according to an exemplary embodiment.

Fig. 5 is a flowchart of a method for acquiring a human body training feature set according to an exemplary embodiment.

Fig. 6 is a flow chart of another method for acquiring a human body training feature set according to an exemplary embodiment.

Fig. 7 is a flowchart of another training method for a human body weight recognition model according to an exemplary embodiment.

Fig. 8 is a flowchart of another method for acquiring a human body training feature set according to an exemplary embodiment.

Fig. 9 is a flowchart of a method for acquiring a human body training image set according to an exemplary embodiment.

Fig. 10 is a structural block diagram of a human body weight recognition device according to an exemplary embodiment.

Fig. 11 is a schematic diagram of a hardware structure of a computer device according to an exemplary embodiment.

Detailed ways

In related technologies, when training the human body weight recognition model, the detection model is used to detect human body pictures from the original pictures captured by multiple cameras and form a base library. Use the existing model to extract human body features from the bottom library pictures, and use the clustering method to classify the bottom library features. Manually check and screen each cluster in the clustering results. Calculate the similarity between clusters in the manual screening results, generate a recommended combination whose similarity between clusters is greater than the set threshold, and then submit it to manual checking and merging, and finally generate manual labeling results. Train a feature extraction network based on manual annotation results, input pedestrian pictures into the feature extraction network to obtain high-dimensional vector feature representations, obtain loss values according to the metric distance and cross-entropy calculated by the features, and select the optimizer to iterate to find the minimum value of the loss function. And continuously update the parameters of the network to achieve the effect of optimized learning, and obtain the final human body weight recognition model.

Since the same human body may appear repeatedly between different cameras, the number of human bodies appearing in each camera is large. Therefore, when using this method to train the human body re-identification model, it is necessary to match the same human body appearing in different cameras and arrange them. The number of combinations is very large, which leads to a large amount of time cost for training.

In order to solve the above problems, an embodiment of the present application provides a human body weight recognition method for use in computer equipment. It should be noted that its execution subject can be a human body weight recognition device, which can be implemented through software, hardware or a combination of software and hardware. Realized as part or all of the storage device in a manner, wherein the computer device can be a terminal or a client or a server, and the server can be a single server, or a server cluster composed of multiple servers. In the embodiment of this application, the Terminals can be smart phones, personal computers, tablet computers, wearable devices, and other smart hardware devices such as smart robots. In the following method embodiments, the execution subject is a computer device as an example for illustration.

The computer device in the embodiment of the present application is applied to a usage scenario of performing human body weight recognition detection on a human body image of an unknown human body type. Through the human body weight recognition method provided by this application, when performing human body weight recognition detection on the human body image to be detected, the human body weight recognition model that can be trained based on the human body training image collected by a single camera, and then multiple people in the preset database In the body feature, the most similar human body category is matched to the human body image to be detected, thereby improving the accuracy of human body category recognition and reducing the occurrence of misidentification.

Fig. 1 is a flow chart of a method for body weight recognition proposed according to an exemplary embodiment. As shown in FIG. 1 , the distributed cluster expansion method includes the following steps S101 to S104.

In step S101, an image of a human body to be detected is acquired.

In the embodiment of the present application, the human body image to be detected can be understood as an image that does not specify that the human body image specifically corresponds to a human body category. Human body category can be understood as a category used to distinguish different human bodies, and each independent individual corresponds to a human body category. For example: Pedestrian A and Pedestrian B are two different human categories. In one example, for the convenience of recording, the form of an identity document (id) can be used to record and distinguish each human body category. For example: id1 corresponds to the first human category, and id2 corresponds to the first human category.

In step S102, the first human body feature of the human body image is extracted by using a pre-trained human body re-identification model.

In the embodiment of the present application, the pre-trained human body weight recognition model is obtained by training based on the human body training images collected by a single camera. The shooting characteristics of the human body improve the accuracy of human body feature extraction, and then when the human body category corresponding to the human body image is subsequently matched, it helps to improve the matching accuracy.

In step S103, the first human body feature is compared with multiple human body features in a preset database, and the second human body feature with the highest similarity to the first human body feature is determined.

In the embodiment of the present application, the human body categories corresponding to the multiple human body characteristics in the preset database are known, and the human body characteristics corresponding to each human body category can be specified. Comparing the first human body feature with multiple human body features in a preset database, so as to determine whether any human body feature similar to the first human body feature is included in the multiple human body features of known human body categories. Since the higher the similarity, the greater the possibility that the two belong to the same human body category, therefore, based on the comparison results, determine the second human body feature with the highest similarity to the first human body feature among multiple human body features, so as to pass the The second human body feature determines the human body category corresponding to the first human body feature. In an example, when determining the similarity between the first human body feature and each human body feature, it may be determined by cosine similarity, Euclidean distance or Minkowski distance.

In step S104, when the similarity is greater than or equal to a specified threshold, the human body category corresponding to the second human body feature is matched to the human body category most similar to the human body image.

In the embodiment of the present application, since the same human body type is photographed in different scenes, there may be differences in the extracted human body features. Therefore, the specified threshold can be understood as an error-tolerant threshold for determining whether the first human body feature and the second human body feature belong to the same human body category. When the similarity is greater than or equal to the specified threshold, it indicates that the first human body feature and the second human body feature are too similar, and the first human body feature and the second human body feature may belong to the same human body category. When the similarity is less than the specified threshold, it indicates that the similarity between the first human body feature and the second human body feature is low, and it can be determined that the first human body feature and the second human body feature do not belong to the same human body category. Therefore, when the similarity is greater than or equal to the specified threshold, the human body category corresponding to the second human body feature can be matched as the most similar human body category of the human body image, so that the final corresponding human body category of the human body image can be determined according to the matched human body category. In one example, if there are multiple second human body features whose similarity is greater than or equal to a specified threshold, the human body category corresponding to the second human body feature with the largest similarity value is determined to be the most similar human body category to the human body image.

Through the above-mentioned embodiment, when performing human body weight recognition detection on the human body image to be detected, the human body weight recognition model that can be trained based on the human body training image collected by a single camera, and then among the multiple human body features in the preset database are the human body weight recognition models to be The detected human body images are matched to the most similar human body category, which helps to improve the accuracy of human body category recognition and reduce the occurrence of misidentification.

In one embodiment, if the similarity between the first human body feature and the second human body feature is less than a specified threshold, it can be determined that the human body category corresponding to the human body image is a new human body category, and then the first human body category of the human body image can be A human body feature and the corresponding human body category are stored in a preset database, so that the accuracy of recognition and tracking can be improved in subsequent human body weight recognition and tracking.

The following examples will illustrate the specific training process of the human weight recognition model.

Fig. 2 is a flowchart of a method for training a human body weight recognition model according to an exemplary embodiment. As shown in Figure 2, the training method of the human body weight recognition model includes the following steps.

In step S201, multiple groups of first human body training image sets are acquired.

In the embodiment of the present application, different sets of first human body training image sets are collected by different cameras, and the types of human bodies included in different first human body training image sets are the same. That is, the human body categories included in the first human body training image sets of different groups are the same, but the first human body training image sets of different groups are collected by different cameras. For example: the first human body training image set of group A is collected by camera A, and the first human body training image set of group B is collected by camera B. The first human body training image set of group A includes the same human body category as that included in the first human body training image set of group B.

In step S202, each first human body training image set is input into the feature extraction network to be trained, and the trained feature extraction network extracts human body features corresponding to each human body category to obtain a first feature extraction model.

In the embodiment of the present application, when training the feature extraction network, each first human body training image set is input into the feature extraction network to be trained respectively, and the first human body training image set collected by each camera is respectively carried out by the feature extraction network. Human body feature extraction, and then according to the extracted human body features, train the feature extraction network to obtain the first feature extraction model. For example: the acquired first human body training image sets are three groups A, B and C respectively. When training the feature extraction network, A, B, and C are respectively input into the feature extraction network for human body feature extraction, and the feature extraction network extracts the human body feature a of each human body category corresponding to A, and extracts the human body feature a corresponding to B. The human body feature b of each human body category and extract the human body feature c of each human body category corresponding to C, and then train the feature extraction network based on a, b and c to extract the human body features corresponding to each human category, and obtain the first feature extraction model.

When training the first feature extraction model, the human body training images used are all images marked with human body categories. When training the feature extraction model, since the first human body training image set collected by each camera is separately input into the feature network model for training, and then when annotating the human body training image in advance, there is no need to label across cameras, only for The current camera can be labeled, which reduces the difficulty of labeling, saves labeling time, and helps to improve labeling efficiency.

In step S203, a second human body training image set and a human body training feature set of the same human body category as included in the first human body training image set are acquired.

In the embodiment of the present application, the multiple human body training features included in the human body training feature set are the total human body features corresponding to the first human body training image in each group of human body image sets. In one example, the first feature extraction model is used to extract human body features from the first human body training image sets respectively, to obtain the human body feature sets corresponding to each first human body training image set, and to combine the human body feature sets to obtain the human body training image set. feature set.

The human body categories included in the human body training feature set are the same as the human body categories included in the first human body training image set, and then the subsequent training of the first feature extraction model can perform targeted training when the first feature extraction model can perform feature recognition across cameras.

In an example, the second human body training image set may be a human body training image set obtained by merging each group of first human body training image sets.

In step S204, the second human body training image set is input to the first feature extraction model, and the first feature extraction model is trained according to the human body training feature set to obtain a human body re-identification model.

In the embodiment of the present application, since the first feature training model is trained on the first human body training image set collected by a single camera, in practical applications, the human body weight recognition model used for human body weight recognition needs to learn from multiple cameras Human body weight recognition is performed on the human body images to be detected in the collected human body images. Therefore, in order to enable the first feature extraction model to realize cross-camera human body re-identification, the acquired second human body training image set is input into the first feature extraction model, and the first feature is trained according to the acquired human body training feature set. The extraction model learns the human body features of the same human body type in human body images collected by different cameras, and then obtains a human body weight recognition model for human body weight recognition.

Through the above-mentioned embodiment, the process of training the human body weight recognition model is divided into two stages. When training the first feature extraction model stage based on the first human body training image collection collected by a single camera, it is helpful to enhance the learning and extraction of the first feature extraction model. The robustness of human body features collected by a single camera, and then the subsequent training of the first feature extraction model to learn the human body features of the same body type in human body images collected by different cameras, helps to enhance the robustness of human body feature extraction Therefore, it is more accurate when performing human feature extraction in actual use.

The following embodiments will illustrate the specific training process of the first feature extraction model.

Fig. 3 is a flowchart of a method for training a first feature extraction model according to an exemplary embodiment. As shown in Fig. 3, the training method of the first feature extraction model includes the following steps.

In step S301, the current group of first human body training image sets are input into the feature extraction network to be trained for human body feature extraction, and the feature extraction results are input into the preset first classification network for feature classification, and the first classification results.

In the embodiment of the present application, the feature extraction network to be trained is a network model for feature extraction. If only based on the human body features output by the feature extraction network, it is impossible to determine whether the extracted human body features are accurate and whether the extracted human body features can reflect Differences between different human body types. Therefore, when training the feature extraction network, combined with the preset first classification network, the feature extraction network performs classification on the output feature extraction results of human body feature extraction for each first human body training image, and obtains the first human body feature corresponding Classification results, so as to adjust the parameters of the feature extraction network according to the first classification results corresponding to the first human body training images and the corresponding human body categories, so as to obtain the first feature extraction model. In one example, the feature extraction network may be a backbone feature extraction network (a type of neural network), and its network architecture includes convolutional layers and pooling layers. Among them, the convolutional layer is used to reduce the dimensionality and feature extraction of the input image, and the pooling layer is used to reduce the model size and enhance the robustness. Since the structure of the backbone feature extraction network is relatively simple and the iteration speed is fast, it is helpful to improve the training efficiency and reduce the training cost during training.

In an example, the first classification network may be a fully connected classification layer connected to the output end of the feature extraction network, and is used to assist the feature extraction network in training human body feature extraction.

In another example, the number of fully connected classification layers is the same as the number of groups of the first human body training image set, and corresponds to each group of first human body training image sets, and the number of classifications calculated by the fully connected classification layer is also the same as the number of groups corresponding to the first human body training image set. A human body training image set includes the same number of human body categories. It can be understood that, when training the feature extraction network, after the current first human body training image set is input to the feature extraction network, the feature extraction result output by the feature extraction network is input to the fully connected classification layer corresponding to the current first human body training image set The feature classification is carried out to obtain the first classification result of extracting the current first human body training image set.

In step S302, based on the first classification result and the human body categories included in the first human body training image set, the training feature extraction network extracts the human body features corresponding to each human body category to obtain the first feature extraction model.

In the embodiment of the present application, according to the first classification result and the human body category included in the first human body training image set, the loss generated when the first classification network performs classification is determined through the loss function, and then the parameters of the feature extraction network are adjusted accordingly , and according to the adjusted parameters, train the feature extraction network to extract the human body features corresponding to each human body category, so as to obtain the first feature extraction model. In one example, the loss function used may include Arcface loss, cross-entropy loss function or triplet loss function.

Through the above embodiments, when training the feature extraction network, training based on the first classification network helps to improve the accuracy of the feature extraction network learning and extracting human body features, and then makes the human body features extracted by the first feature extraction model better. The differences in human body categories can be expressed, which lays a good feature extraction foundation for the subsequent training of the human body weight recognition model, thereby enhancing the robustness of the human body weight recognition model.

In an implementation scenario, the training process of the first feature extraction model may be as shown in FIG. 4 . Fig. 4 is a flow chart of another training method for a first feature extraction model proposed according to an exemplary embodiment. Wherein, the number of groups of the first human body training image set used to train the backbone feature extraction network is n, corresponding to n cameras respectively, n>1, and n is an integer. For ease of description, the process of training the backbone feature extraction network with the first human body training image set corresponding to the camera 1 will be described as an example below.

Input the first human body training image set corresponding to camera 1 into the backbone feature extraction network for human body feature extraction, and input the output feature extraction results into the fully connected classification layer 1 corresponding to camera 1 for feature classification, and obtain the first classification As a result, the classification loss is calculated, and then the parameters of the feature extraction network are adjusted accordingly, and the feature extraction network is trained according to the adjusted parameters to extract the human body features corresponding to each human body category.

The process of training the backbone feature extraction network with the first human body training image set corresponding to any one of cameras 2 to n is the same as the process of training the backbone feature extraction network with the first human body training image set corresponding to camera 1, and will not be carried out here repeat.

After the training for each group of first human body training image sets is completed, a first feature extraction model is obtained.

其中，loss _n为第n个摄像头对应的损失值，P为全连接分类层的预测概率，i,m分别对应单个摄像头的某一人体类别和人体类别的总数量。 In another implementation scenario, if the number of human body categories in the first human body training image set is small, the loss may be calculated using a cross-entropy loss function. Calculated as follows:

Among them, loss _n is the loss value corresponding to the nth camera, P is the predicted probability of the fully connected classification layer, and i and m correspond to a certain human body category and the total number of human body categories of a single camera, respectively.

The following embodiments will illustrate the specific acquisition process of the human body training feature set.

Fig. 5 is a flowchart of a method for acquiring a human body training feature set according to an exemplary embodiment. As shown in Fig. 5, the method for obtaining the human body training feature set includes the following steps.

In step S501, the human body feature sets corresponding to the first human body training image sets are respectively extracted through the first feature extraction model.

In step S502, the first human body feature in the current human body feature set is sequentially matched with the second human body features in the next human body feature set, and the second human body feature with the highest matching degree with the first human body feature in the next human body feature set is determined. human body features.

In the embodiment of the present application, when performing matching and fusion, the first human body features in the current human body feature set are sequentially matched with the second human body features in the next human body feature set, and the next human body feature set and the first human body feature are determined. The second human body feature with the highest similarity.

In step S503, each first human body feature is fused into the second human body feature with the highest matching degree to obtain a new second human body feature, and a pseudo-label corresponding to each new second human body feature is generated.

In the embodiment of the present application, since the human body categories included in each first human body training image set are the same, the human body category corresponding to the second human body feature with the highest similarity with the first human body feature can be determined to be the same as the first human body. The corresponding human body category of the feature is the same human body category, and then the first human body feature and the second human body feature with the highest similarity with the first human body feature can be fused to obtain a new second human body feature, and generate the new second human body Pseudo-labels corresponding to features. Wherein, the pseudo-label can be understood as a human body category corresponding to the new second human body feature.

In step S504, each human body feature set is cycled in turn until the new second human body features corresponding to each human body category in each group of human body feature sets and corresponding pseudo-labels are obtained to obtain a human body training feature set.

In the embodiment of the present application, the second human body features corresponding to each first human body feature in the current human body feature set are sequentially determined in the next human body feature set, and each first human body feature is fused into the corresponding second human body feature to obtain multiple New secondary human traits. After all the first human body features in the current human body feature set are fused, the current human body feature set is canceled, and the next human body feature set including multiple new second human body features is used as a new current human body feature set, and the new Each new second human body characteristic in the current human body characteristic set is used as the first human body characteristic, and is compared with each second human body characteristic in the next human body characteristic set, and the second human body characteristic corresponding to each first human body characteristic is determined. By analogy, each first human body feature is fused into the corresponding second human body feature, until the new second human body feature corresponding to each human body category in each group of human body feature sets and the corresponding pseudo-label are obtained, and the human body training feature set is obtained. That is, any human body training feature in the human body training feature set includes the total human body features of the pseudo-label corresponding to the human body training feature in each human body feature set, and the human body training feature is the finally generated new second human body feature.

Through the above-mentioned embodiment, the human body features collected by each camera are fused into the same human body training feature, which is helpful for the subsequent training of the first feature extraction model, so that the first feature extraction model can fully learn the human body features of each human body category, and then make The obtained human body weight recognition model can reduce the difference in human body feature detection due to different cameras when performing feature extraction, thereby helping to improve the accuracy of detection results.

In an implementation scenario, the acquisition process of the human body training feature set may be as shown in FIG. 6 . Fig. 6 is a flow chart of another method for acquiring a human body training feature set according to an exemplary embodiment. Among them, id represents the human body category, id _m represents the first human body feature in the human body feature set corresponding to camera n, and id _k represents the second human body feature in the human body feature set corresponding to camera p. Compare the first human body feature corresponding to id _m in camera n with the second human body features in the human body feature set corresponding to camera p, and determine the highest similarity to id _m in the human body feature set corresponding to camera p

id _k , merge id _m and id _k , and generate corresponding pseudo-labels.

The following embodiments will illustrate the specific training process of training the human weight recognition model through the first feature extraction model.

Fig. 7 is a flowchart of another training method for a human body weight recognition model according to an exemplary embodiment. As shown in FIG. 7 , the training method of the human body weight recognition model includes the following steps.

In step S701, the second human body training image set is input into the first feature extraction model for human body feature extraction, and the feature extraction result is input into a preset second classification network for feature classification to obtain a second classification result.

In the embodiment of the present application, the second classification network may be a fully connected classification layer connected to the output end of the first feature extraction model, by classifying the output results of the first feature extraction model, and assisting The first feature extraction model performs training for human feature extraction. In an example, the second classification network model and the first classification network model may be the same classification network model. In another example, the second classification network model and the first classification network model may be two different classification network models, which are used in different training stages respectively. Wherein, the training stage includes: the training stage of obtaining the first feature extraction model by training the feature extraction network, and the training stage of obtaining the human body re-identification model stage by training the first feature extraction model.

In step S702, based on the second classification result and the human body training feature set, the first feature extraction model is trained to extract each human body feature in the second human body training image set to obtain a human body re-identification model.

In the embodiment of the present application, according to the second classification result and the pseudo-label corresponding to each human body training feature, the loss generated by the second classification network is determined through the loss function, and then the parameters of the first feature extraction model are adjusted accordingly , and train the first feature extraction model according to the adjusted parameters to extract the human body features corresponding to each human body category, so as to obtain the human body weight recognition model. In one example, the used loss function may include Arcface loss, cross-entropy loss function or triplet loss function.

In an example, when adjusting the parameters of the first feature extraction model, parameter adjustment may be performed by continuously iteratively updating network parameters through gradient backpropagation.

Through the above-mentioned embodiment, training the first feature extraction model based on the human body training features extracted correspondingly from multiple cameras helps the first feature extraction model to fully learn the corresponding human body features of the same human body category in different shooting scenes, thereby making When the human body weight recognition model extracts human body features, it can fully express the human body characteristics corresponding to different human body types, thereby improving the accuracy of human body weight recognition detection.

In an embodiment, before inputting the second human body training image set to the first feature extraction model, each second human body training image in the second human body training image set may be scaled to the same specified size, and then the subsequent extraction of human body features , which helps to save computational cost.

In an implementation scenario, the training process of the human body weight recognition model may be as shown in FIG. 8 . Fig. 8 is a flowchart of another training method for a human body weight recognition model according to an exemplary embodiment.

In step S801, the second human body training image in the second human body training image set is scaled to a specified size.

In step S802, the second human body training image scaled to a specified size is input to the first feature extraction model for feature extraction, and a feature extraction result is obtained.

In step S803, the feature extraction result is input to the fully connected classification layer for classification, and the pseudo-label corresponding to the second human body training image is determined.

In step S804, the fully connected classification layer is determined to determine the classification loss of the pseudo-label corresponding to the second human body training image.

In another implementation scenario, if the output classification loss does not meet the training requirements, then adjust the parameters of the first feature extraction model, and continue training according to the adjusted parameters.

In one embodiment, each of the first human body training images in the first human body training image set is marked with the human body category that first appears in the video captured by the current camera, and the target tracking algorithm is used to extract the multiple corresponding human body categories in the video. The first human body training image is obtained to obtain the first human body training image set, so that when labeling the first human body training image, there is no need to pay attention to the same human body appearing between the cameras, thereby helping to reduce the difficulty of labeling. Moreover, multiple annotators can also annotate multiple videos at the same time, which helps to improve the efficiency of annotation.

In an example, the target tracking algorithm may include a detection tracking algorithm (Tracking By Detection, TBD) or a SORT algorithm (an online real-time multi-target tracking algorithm), which is not limited in this application.

In an implementation scenario, the process of acquiring the first human body training image set may be as shown in FIG. 9 . Fig. 9 is a flowchart of a method for acquiring a human body training image set according to an exemplary embodiment.

In step S901, the video captured by the current camera is obtained.

In step S902, it is judged whether the current human body category appears in the video for the first time.

In step S903, if the current human body type appears in the video for the first time, mark and extract the first human body training image in which the current human body type appears in the first frame image.

In step S904, the first human body training image of the current human body category in other frame images is extracted through target tracking to obtain a first human body training image set.

Based on the same application concept, the present application also provides a human body weight recognition device.

Fig. 10 is a structural block diagram of a human body weight recognition device according to an exemplary embodiment. As shown in FIG. 10 , the human body weight recognition device includes an acquisition unit 1001 , an extraction unit 1002 , a comparison unit 1003 and a determination unit 1004 .

An acquisition unit 1001, configured to acquire a human body image to be detected;

The extraction unit 1002 is used to extract the first human body feature of the human body image through the pre-trained human body weight recognition model, and the human body weight recognition model is trained based on the human body training image collected by a single camera;

A comparison unit 1003, configured to compare the first human body feature with a plurality of human body features in a preset database, and determine a second human body feature with the highest similarity to the first human body feature;

The determining unit 1004 is configured to match the human body category corresponding to the second human body feature to the human body category most similar to the human body image when the similarity is greater than or equal to a specified threshold.

In one embodiment, the human body weight recognition model is trained based on the human body training images collected by a single camera using the following units: the first acquisition unit is used to acquire multiple groups of first human body training image sets, different groups of first human body training images The sets are collected by different cameras, and the categories of human bodies included in different first human training image sets are the same. The first training unit is configured to input each first human body training image set into the feature extraction network to be trained, and train the feature extraction network to extract human body features corresponding to each human body category to obtain a first feature extraction model. The second acquisition unit is used to acquire a second human body training image set and a human body training feature set that are the same as the human body category included in the first human body training image set, and the multiple human body training features included in the human body training feature set are each group of human body image sets The overall human body features corresponding to the first human body training image. The second training unit is configured to input the second human body training image set into the first feature extraction model, train the first feature extraction model according to the human body training feature set, and obtain the human body re-identification model.

In another embodiment, the first training unit includes: a first extraction unit, configured to input the current set of first human body training image sets into the feature extraction network to be trained for human body feature extraction, and input the feature extraction results to The feature classification is performed in the preset first classification network to obtain the first classification result. The first training subunit is configured to train the feature extraction network to extract human body features corresponding to each human body category based on the first classification result and the human body categories included in the first human body training image set, so as to obtain a first feature extraction model.

In yet another embodiment, the second acquiring unit includes: a second extracting unit, configured to respectively extract human body feature sets corresponding to each first human body training image set through the first feature extraction model. The first matching unit is configured to sequentially match the first human body features in the current human body feature set with the second human body features in the next human body feature set, and determine the person with the highest matching degree with the first human body feature in the next human body feature set Secondary human characteristics. The fusion unit is used to fuse the first human body features into the second human body features with the highest matching degree to obtain new second human body features, and generate pseudo-labels corresponding to the human body categories of each new second human body features. The second obtaining subunit is used to cycle through each human body feature set in turn until obtaining a new second human body feature corresponding to each human body category in each group of human body feature sets and corresponding pseudo-labels to obtain a human body training feature set.

In yet another embodiment, the second training unit includes: a third extraction unit, configured to input the second human body training image set into the first feature extraction model for human body feature extraction, and input the feature extraction result into a preset The feature classification is performed in the second classification network to obtain the second classification result. The second training subunit is used to train the first feature extraction model to extract the features of each human body in the second human body training image set based on the second classification result and the human body training feature set to obtain a human body re-identification model.

In yet another embodiment, the second acquiring unit includes: a determining unit configured to determine and mark the first-appearing human body category according to the video captured by the current camera. The first image extraction unit is configured to extract a plurality of first human body training images corresponding to each of the human body categories in the video through a target tracking algorithm, and obtain a first human body training image set corresponding to the current camera.

In yet another embodiment, the second acquisition unit includes: a merging unit, configured to combine the first human body training image sets to obtain a second human body training image set.

The specific limitations and beneficial effects of the above-mentioned human body weight recognition device can refer to the above-mentioned definition of the human body weight recognition method, and will not be repeated here. Each of the above-mentioned modules can be fully or partially realized by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, and can also be stored in the memory of the computer device in the form of software, so that the processor can invoke and execute the corresponding operations of the above-mentioned modules.

Fig. 11 is a schematic diagram of a hardware structure of a computer device according to an exemplary embodiment. As shown in FIG. 11 , the device includes one or more processors 1110 and a memory 1120 , and the memory 1120 includes a persistent memory, a volatile memory, and a hard disk. In FIG. 11 , one processor 1110 is taken as an example. The device may also include: an input device 1130 and an output device 1140 .

The processor 1110, the memory 1120, the input device 1130, and the output device 1140 may be connected via a bus or in other ways, and connection via a bus is taken as an example in FIG. 11 .

The processor 1110 may be a central processing unit (Central Processing Unit, CPU). The processor 1110 can also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application-specific integrated circuits (Application Specific Integrated Circuit, ASIC), field-programmable gate array (Field-Programmable Gate Array, FPGA) or Other chips such as programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above-mentioned types of chips. A general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.

The memory 1120 is a non-transitory computer-readable storage medium, including persistent memory, volatile memory and hard disk, and can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as the The program instruction/module corresponding to the business management method. The processor 1110 executes various functional applications and data processing of the server by running the non-transitory software programs, instructions and modules stored in the memory 1120, that is, implements any one of the above human body re-identification methods.

The memory 1120 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application program required by at least one function; In addition, the memory 1120 may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory 1120 may optionally include memory located remotely relative to the processor 1110, and these remote memories may be connected to the data processing device through a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The input device 1130 can receive input numbers or character information, and generate key signal input related to user settings and function control. The output device 1140 may include a display device such as a display screen.

One or more modules are stored in the memory 1120, and when executed by the one or more processors 1110, perform the methods shown in FIGS. 1-9.

The above-mentioned products can execute the method provided by the embodiment of the present application, and have corresponding functional modules and beneficial effects for executing the method. For technical details that are not described in detail in this embodiment, please refer to the relevant description in the embodiment shown in FIGS. 1-9 for details.

The embodiment of the present application also provides a non-transitory computer storage medium, the computer storage medium stores computer-executable instructions, and the computer-executable instructions can execute the authentication method in any of the above method embodiments. Wherein, the storage medium can be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a flash memory (Flash Memory), a hard disk (Hard Disk Drive) , abbreviation: HDD) or solid-state drive (Solid-State Drive, SSD), etc.; the storage medium may also include a combination of the above-mentioned types of memory.

Apparently, the above-mentioned embodiments are only examples for clear description, rather than limiting the implementation. For those of ordinary skill in the art, other changes or changes in different forms can be made on the basis of the above description. It is not necessary and impossible to exhaustively list all the implementation manners here. However, the obvious changes or changes derived therefrom are still within the protection scope of the invention of the present application.

Claims (10)

A method for human body weight recognition, wherein the method includes: Obtain the human body image to be detected; Extracting the first human body feature of the human body image through a pre-trained human body weight recognition model, the human body weight recognition model is trained based on a human body training image collected by a single camera; Comparing the first human body feature with a plurality of human body features in a preset database, and determining a second human body feature with the highest similarity to the first human body feature; When the similarity is greater than or equal to a specified threshold, the human body category corresponding to the second human body feature is matched to the human body category most similar to the human body image. The method according to claim 1, wherein the human body weight recognition model is trained based on a human body training image collected by a single camera, comprising: Obtain multiple sets of first human body training image sets, the first human body training image sets of different groups are collected by different cameras, and the human body categories included in different first human body training image sets are the same; Input each of the first human body training image sets into the feature extraction network to be trained, train the feature extraction network to extract the human body features corresponding to each of the human body categories, and obtain the first feature extraction model; Acquire a second human body training image set and a human body training feature set that are the same as the human body category included in the first human body training image set, and the multiple human body training features included in the human body training feature set are each group of human body image sets corresponding to the first Total human features of human training images; The second human body training image set is input to the first feature extraction model, and the first feature extraction model is trained according to the human body training feature set to obtain the human body re-identification model. The method according to claim 2, wherein the first human body training image sets are respectively input into the feature extraction network to be trained, and the feature extraction network is trained to extract human body features corresponding to each of the human body categories , get the first feature extraction model, including: Inputting the first human body training image set of the current group into the feature extraction network to be trained for human body feature extraction, and inputting the feature extraction result into the preset first classification network for feature classification to obtain the first classification result; Based on the first classification result and the human body categories included in the first human body training image set, train the feature extraction network to extract the human body features corresponding to each of the human body categories to obtain a first feature extraction model. The method according to claim 3, wherein obtaining the human body training feature set comprises: Using the first feature extraction model, respectively extracting human body feature sets corresponding to each of the first human body training image sets; sequentially matching the first human body features in the current human body feature set with the second human body features in the next human body feature set, and determining the second human body feature with the highest matching degree between the next human body feature set and the first human body feature; Fusing each of the first human body features into the second human body feature with the highest matching degree to obtain a new second human body feature, and generating a pseudo-label corresponding to the human body category of each of the new second human body features; Each of the human body feature sets is cycled in turn until the new second human body features corresponding to each human body category in each group of human body feature sets and corresponding pseudo-labels are obtained, so as to obtain a human body training feature set. The method according to claim 4, wherein the second human body training image set is input to the first feature extraction model, and the first feature extraction model is trained according to the human body training feature set to obtain the The human body weight recognition model includes: Inputting the second human body training image set into the first feature extraction model for human body feature extraction, and inputting the feature extraction result into a preset second classification network for feature classification to obtain a second classification result; Based on the second classification result and the human body training feature set, train the first feature extraction model to extract each human body feature in the second human body training image set to obtain the human body re-identification model. The method according to any one of claims 2-5, wherein said obtaining multiple sets of first human body training image sets comprises: According to the video captured by the current camera, determine and mark the human body category that appears for the first time; A plurality of first human body training images corresponding to each of the human body types in the video are extracted through a target tracking algorithm to obtain a first human body training image set corresponding to the current camera. The method according to claim 6, wherein said obtaining a second human body training image set identical to the human body category included in said first human body training image set comprises: Merge each of the first human body training image sets to obtain a second human body training image set. A human body weight recognition device, wherein the device includes: an acquisition unit, configured to acquire a human body image to be detected; The extraction unit is used to extract the first human body feature of the human body image through a pre-trained human body weight recognition model, and the human body weight recognition model is trained based on a human body training image collected by a single camera; A comparing unit, configured to compare the first human body feature with a plurality of human body features in a preset database, and determine a second human body feature with the highest similarity to the first human body feature; A determining unit, configured to match the human body category corresponding to the second human body feature to the human body category most similar to the human body image when the similarity is greater than or equal to a specified threshold. A computer device, including a memory and a processor, the memory and the processor are connected in communication with each other, and computer instructions are stored in the memory, and the processor implements the rights by executing the computer instructions The human body weight identification method described in any one of claims 1-7. A computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the human body re-identification method according to any one of claims 1-7.

Images