An identity type recognition system

Abstract

The present application provides an identity type recognition system, which includes: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit; the image acquisition device is used to acquire image information to be processed; the image analysis unit is used to extract features of the image information to be processed to obtain user features of the target user; the feature determination unit is used to determine identity features based on user features; the identity prediction unit is used to determine the identity type of the target user based on identity features and user features. It can be seen that the present application can use the identity type recognition system to automatically identify the user's identity type, without the need to manually identify the user's identity as in the prior art, avoiding the subjectivity, one-sidedness and limitations caused by manual experience-based evaluation, as well as the problem of identity recognition errors caused by manual negligence, thereby improving the efficiency of user identity type recognition and thus improving user experience.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to an identity type recognition system. Background Art

[0002] In social governance scenarios, there is no good way to identify the identity types of people who appear in the jurisdiction. That is, when a person appears, the system itself can detect the person and identify whether the person exists in the database, but for strangers, it cannot identify whether the person is a takeaway, a resident of the jurisdiction, or a staff member of an enterprise or institution.

[0003] The existing main methods still rely on manual identification and empirical inference, and cannot be automatically labeled. As a result, the process of manually identifying strangers is time-consuming, laborious and inefficient, and the manual identification process is prone to identity errors caused by human negligence. Therefore, an identity type identification solution is urgently needed. Summary of the invention

[0004] The present application provides and implements an identity type identification system to improve the efficiency of user identity type identification, thereby improving user experience.

[0005] The present application provides an identity type recognition system, the system comprising: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit;

[0006] The image acquisition device is used to acquire image information to be processed;

[0007] The image analysis unit is used to extract features from the image information to be processed to obtain user features of the target user;

[0008] The feature determination unit is used to determine the identity feature of the target user according to the user feature;

[0009] The identity prediction unit is used to determine the identity type of the target user according to the identity characteristics and the user characteristics.

[0010] Optionally, the image information to be processed includes images and / or videos, and acquisition locations and acquisition times corresponding to the images and / or videos.

[0011] Optionally, the image analysis unit includes a trained user feature recognition model;

[0012] The image analysis unit is used to input the image information to be processed into the user feature recognition model to obtain the user features of the image information to be processed.

[0013] Optionally, the user characteristics include facial images, facial features, facial attributes, body images, body features, body attributes, behavioral action features, and location attributes.

[0014] Optionally, the identity features include target behavior, action trajectory, and face and body feature sets; the feature determination unit includes a behavior recognition unit, a clustering unit, and a trajectory determination unit;

[0015] The behavior recognition unit is used to perform behavior recognition according to the behavior action characteristics to obtain the target behavior;

[0016] The clustering unit is used to perform clustering processing based on facial features and body features to obtain a facial and body feature set corresponding to the target user;

[0017] The trajectory determination unit is used to determine the movement trajectory of the target user based on the face and body feature set corresponding to the target user and the position attributes, collection positions and collection times corresponding to each face feature and body feature in the face and body feature set.

[0018] Optionally, the behavior recognition unit is used to determine the target behavior corresponding to the behavior action feature according to the behavior action feature and the location attribute, collection location and collection time corresponding to the behavior action feature; and / or,

[0019] The behavior recognition unit includes a trained behavior recognition model; the behavior recognition unit is used to input the behavior feature into the behavior recognition model to obtain the target behavior corresponding to the behavior action feature.

[0020] Optionally, the clustering unit is used to determine the similarity between the facial features, the body features and the features in each historical feature set; and determine the facial and body feature set corresponding to the target user based on the similarity between the facial features, the body features and the features in each historical feature set.

[0021] Optionally, the clustering unit is specifically used to determine the fusion features corresponding to each of the at least two historical feature sets if the similarity between the facial features and / or the body features and the features in at least two historical feature sets is greater than a preset threshold; perform gain correction on the fusion features corresponding to each of the at least two historical feature sets to obtain the corrected fusion features corresponding to each of the at least two historical feature sets; determine the similarity between the facial features and / or the body features and each corrected fusion feature, and use the historical feature set corresponding to the corrected fusion feature with the greatest similarity as the facial and body feature set corresponding to the target user.

[0022] Optionally, the identity prediction unit includes a random forest model, wherein the random forest model includes a plurality of decision tree models, and the dimension weights corresponding to each decision tree model are different;

[0023] The identity prediction unit is used to input the identity feature and the user feature into each decision tree model respectively to obtain the identity prediction type of the target user output by each decision tree model;

[0024] The identity type of the target user is determined according to the predicted identity type of the target user output by each decision tree model.

[0025] Optionally, the system further includes a storage unit for storing the image information to be processed, the user characteristics, the identity characteristics and the identity type of the target user.

[0026] It can be seen from the above technical scheme that the present application provides an identity type recognition system, the system comprising: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit; the image acquisition device is used to acquire image information to be processed; the image analysis unit is used to extract features from the image information to be processed to obtain user features of the target user; the feature determination unit is used to determine identity features based on the user features; the identity prediction unit is used to determine the identity type of the target user based on the identity features and the user features. It can be seen that in this embodiment, after the image acquisition device acquires the image information to be processed, the image analysis unit, the feature determination unit and the identity prediction unit can be used to process the image information to be processed to obtain the identity type of the target user in the image information to be processed. In this way, the present application can use the identity type recognition system to automatically identify the identity type of the user, without the need to manually identify the identity of the user as in the prior art, avoiding the subjectivity, one-sidedness and limitations caused by manual experience evaluation, as well as the problem of identity recognition errors caused by manual negligence, thereby improving the efficiency of user identity type recognition and thus improving user experience.

[0027] The further effects of the above-mentioned non-conventional preferred manner will be described below in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present application or the existing technical solutions, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 A schematic diagram of the structure of an identity type recognition system provided in one embodiment of the present application;

[0030] Figure 2 A schematic diagram of a fusion process of an archive fusion subunit provided in one embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0032] The inventor has found through research that in social governance scenarios, there is no good way to identify the identity type of people who appear in the jurisdiction. That is, when a person appears, the system itself can detect the person and identify whether he exists in the database, but it cannot identify strangers as takeout, residents of the jurisdiction, or employees of enterprises and institutions. The main existing methods still rely mainly on manual identification and empirical inference, and cannot be automatically labeled. As a result, the process of manually identifying the identities of strangers is time-consuming, labor-intensive and inefficient, and the manual identification process is prone to identity recognition errors caused by human negligence. Therefore, there is an urgent need for an identity type recognition solution.

[0033] Therefore, the present application provides an identity type recognition system, in which the system includes: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit; the image acquisition device is used to acquire image information to be processed; the image analysis unit is used to extract features from the image information to be processed to obtain user features of the target user; the feature determination unit is used to determine identity features based on the user features; the identity prediction unit is used to determine the identity type of the target user based on the identity features and the user features. It can be seen that in this embodiment, after the image acquisition device acquires the image information to be processed, the image analysis unit, the feature determination unit and the identity prediction unit can be used to process the image information to be processed to obtain the identity type of the target user in the image information to be processed. In this way, the present application can use the identity type recognition system to automatically identify the identity type of the user, without the need to manually identify the identity of the user as in the prior art, avoiding the subjectivity, one-sidedness and limitations caused by manual experience evaluation, as well as the problem of identity recognition errors caused by manual negligence, thereby improving the efficiency of user identity type recognition and thus improving user experience.

[0034] Various non-limiting implementations of the present application are described in detail below in conjunction with the accompanying drawings.

[0035] See also Figure 1 , an embodiment of the present application provides an identity type recognition system, the system comprising: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit. Wherein, the image acquisition device is connected to the image analysis unit, the image analysis unit is respectively connected to the feature determination unit and the identity prediction unit, and the feature determination unit is connected to the identity prediction unit. It should be noted that the image acquisition device, the image analysis unit, the feature determination unit and the identity prediction unit can all be set on a terminal, or the image acquisition device, the image analysis unit, the feature determination unit and the identity prediction unit can each be a terminal respectively, or some of the image acquisition device, the image analysis unit, the feature determination unit and the identity prediction unit can be set on a terminal.

[0036] In this embodiment, the image acquisition device can be used to acquire image information to be processed. In this embodiment, the image acquisition device can only acquire images, or only acquire videos, or acquire both images and videos. That is to say, the image information to be processed may include images and / or videos, and the acquisition position and acquisition time corresponding to the images and / or videos (i.e., the position and time of acquiring images and videos). In one implementation, the image acquisition device can be a camera and capture device that has been deployed at various points in the jurisdiction. In this way, the camera and capture device that has been deployed at various points in the jurisdiction can be used to input video images of faces, bodies, and behaviors.

[0037] In this embodiment, the image analysis unit may be used to extract features from the image information to be processed to obtain user features of the target user.

[0038] In one implementation, the image analysis unit may include a trained user feature recognition model. In this way, the image analysis unit may be used to input the image information to be processed into the user feature recognition model to obtain the user features of the image information to be processed. It should be emphasized that the user feature recognition model may be trained based on the training sample images and the user features of the users in the training sample images. Among them, the user features may include face images, face features, face attributes, body images, body features, body attributes, behavior and action features, and position features.

[0039] Specifically, after the image analysis unit obtains the image information to be processed, it can first obtain the face image, body image and panoramic large picture (for example, the background area) of the target user in the image information to be processed. Then, the face features and face attributes of the target user can be determined based on the face image of the target user, the body features, body attributes and behavior features of the target user can be determined based on the body image of the target user, and the location features corresponding to the target user can be determined based on the scene large picture.

[0040] The face image can be understood as a face image area of ​​a user in the image information to be processed.

[0041] Facial features can be understood as the characteristics of the face corresponding to the face image. Facial features are usually 128-512 dimensional numerical vectors. They are the result of the face processing by the face recognition deep learning model in the inference stage. The features that distinguish this person's face from others are extracted. The cosine similarity of the facial features of multiple images of a person is significantly higher than the cosine similarity with other people's faces. Usually a threshold is set. If the cosine similarity of several facial features is higher than this threshold, then the facial features can be considered to be the facial features of the same person.

[0042] The facial attributes can be understood as the facial information of the target user corresponding to the facial image. Facial attributes can include at least one of the following information: age, gender, whether wearing glasses, whether wearing sunglasses, whether smiling, whether wearing a mask, race, whether opening eyes, whether opening mouth, whether having a beard, hairstyle, etc. It should be emphasized that due to the differences in lighting angles and expressions, the specific value of the age attribute is difficult to be accurate. Based on experience, multiple age intervals can be divided according to age groups. For example, it can be divided into under 15, 16-25, 26-35, 36-45, 46-55, and over 56, corresponding to teenagers, youth, young and middle-aged, middle-aged, middle-aged and elderly, respectively. It should be noted that the segmentation of age groups can be done without manual sorting but using the results of the facial attribute recognition model, providing a random selectivity of model samples, which helps to reduce the overfitting of the final identity recognition model.

[0043] The human body image can be understood as a human body image region of a user in the image information to be processed (can be understood as a limb image region below the head).

[0044] Human features can be understood as the characteristics of the human body corresponding to the human body image. Human features can be used to identify a person in a short time, also known as the features of re-identification cross-lens tracking. They are usually 512-1024-dimensional feature vectors. The matching of this feature is used to distinguish the matching degree of a person's portrait. It can be used to match the front, side and back of a person in a continuous time. It does not support long-term matching such as changing clothes.

[0045] Human attributes can be understood as the human body information of the target user corresponding to the face image. Human attributes usually include top type, top color, trouser type, trouser color, whether it is a shoulder bag, whether it is a backpack, whether it is a handbag, whether it is holding an umbrella or carrying an umbrella, etc.

[0046] Behavior action features can be understood as the target user's action or behavior state. For example, behavior action features include sitting, standing, lying, squatting, walking, running, etc.

[0047] It is understandable that the image analysis unit can be used to process multiple channels of video and image information. These videos or images all have camera location information (i.e., acquisition location) and its corresponding time (acquisition time). Each captured face or body can be distinguished by time and space identification, that is, the captured face information with time and space identification corresponding to the target user can be determined. Specifically, a captured face information corresponding to the target user can include a face image, a panoramic image, a camera ID, name, longitude and latitude, face features, face appearance time, clustering archive ID, etc., which constitutes a piece of information with time and space identification.

[0048] In this embodiment, the existing face recognition CNN convolutional neural network can be used to extract the characteristic attributes of the face and body. This system does not limit this, and can be combined with the application of the current industry face recognition model, for example, ResNeXt can be used.

[0049] It should be emphasized that in this embodiment, human attributes can also include identification information of clothing and helmets for takeout and express delivery, such as identification of clothing and helmets for JD.com, Meituan, and Ele.me. The behavioral action features can also identify the sitting, standing, lying, walking, squatting, and running of the human target (i.e., the human body of the target user) detected in the picture according to the posture recognition algorithm. The posture recognition of the image analysis unit needs to mark the key joints of the human body and train the deep learning model of human body detection and joint point positioning during the training stage. When performing posture recognition on the image, the human body is first detected and the coordinates of the key nodes are further regressed to obtain the skeleton of the target object, thereby judging its posture.

[0050] In this embodiment, the feature determination unit can be used to determine the identity feature of the target user according to the user feature. In this embodiment, the identity feature of the user can be understood as a feature that can be used to reflect the identity type of the user, for example, the identity feature can include target behavior, action trajectory, and face and body feature set.

[0051] In an implementation of this embodiment, the feature determination unit may include a behavior recognition unit, a clustering unit, and a trajectory determination unit.

[0052] The behavior recognition unit can be used to perform behavior recognition according to the behavior action characteristics to obtain the target behavior. The target behavior can be understood as the behavior performed by the target user, and the target behavior can be understood as the continuous behavior performed by the target user over a period of time, such as leaving the camera range, moving in a straight line, moving in a serpentine manner, standing still, wandering, moving at a specific speed, etc.

[0053] In one implementation of this embodiment, the behavior recognition unit can be used to determine the target behavior corresponding to the behavior action feature according to the behavior action feature and the position attribute, acquisition position and acquisition time corresponding to the behavior action feature. That is to say, the behavior recognition unit can perform time-related behavior recognition according to the action characteristics of the characters in the video, such as identifying the wandering, slow walking, fast walking, stopping, and lying down of the target character according to the passing time and travel distance. It can be understood that in this embodiment, the behavior recognition unit can also track the target user identified in the scene capture of the same camera, and perform behavior recognition according to the local movement trajectory of the target (determined by the position attribute, acquisition position and acquisition time), the moving direction and the moving speed. It only needs to determine whether it is moving in a curved serpentine or straight, or reciprocating, wandering, or standing. Based on the performance considerations of the application scenario, there is no need to perform detailed posture and action recognition, only the direction and speed need to be analyzed.

[0054] In one implementation of this embodiment, the behavior recognition unit may include a trained behavior recognition model. The behavior recognition unit inputs the behavior feature into the behavior recognition model to obtain the target behavior corresponding to the behavior action feature. It is understandable that the trained behavior recognition model may be trained based on the action feature training sample and the behavior corresponding to the action feature training sample.

[0055] In this embodiment, the clustering unit is used to perform clustering processing according to facial features and body features to obtain a facial and body feature set corresponding to the target user. It should be noted that in this embodiment, in order to better distinguish the identity type of each user, the relevant data of the same user can be placed in the virtual file corresponding to the user as much as possible. For the convenience of description, the virtual file corresponding to the user can be referred to as a facial and body feature set. It should be noted that the facial and body feature set can include facial features and body features, or the facial and body feature set can include a facial feature set and a body feature set, wherein the facial feature set includes facial features, and the body feature set includes body features.

[0056] It is understandable that the clustering unit is mainly used for face feature clustering and body feature clustering. Face feature clustering can be understood as achieving long-term face feature matching and making unique marks on a series of photos of the target user. This mark can be used to match the travel snapshot records of all locations of the target user. Body feature clustering data is time-sensitive and long-term data cannot be processed together. Try to allow one person to have multiple files in the long run and reduce the possibility of multiple people having one file.

[0057] Specifically, the clustering unit can be used to determine the similarity between the facial features, the body features and the features in each historical feature set; based on the similarity between the facial features, the body features and the features in each historical feature set, determine the facial and body feature set corresponding to the target user.

[0058] As an example, the clustering unit includes two parts: a virtual archive feature library storage subunit and an archive fusion subunit. The virtual archive feature library stores information such as labels, features, last update time, creation time, and fusion feature number of each historical feature set.

[0059] Next, we will take the clustering process of facial features as an example. It should be emphasized that the clustering process of human features is the same as that of facial features. Figure 2 As shown, the virtual profile feature database is empty at the beginning, and the facial feature data is entered into the database in sequence. The entry rules are as follows:

[0060] If the library is empty or has no similar features, it is stored in the library and a new label is generated (i.e., a historical feature library), and the number of fused features is recorded as 1. The newly generated historical feature library is used as the face and body feature set (or face feature set) corresponding to the target user.

[0061] If the similarity between the newly recorded facial feature and a feature of a historical feature set in the library exceeds a preset threshold, and the number of feature records is unique, then this new feature label adopts the feature label compared in the library, the number of fused features increases by 1, and the feature in the library is fused with the new feature, that is, the facial feature is fused with the feature whose similarity exceeds the preset threshold to obtain the fused feature, and the historical feature set corresponding to the feature is used as the facial and body feature set (or facial feature set) corresponding to the target user.

[0062] If the similarity between the facial features and the features in at least two historical feature sets is greater than a preset threshold, determine the fusion features corresponding to each of the at least two historical feature sets; perform gain correction on the fusion features corresponding to each of the at least two historical feature sets to obtain the corrected fusion features corresponding to each of the at least two historical feature sets; determine the similarity between the facial features and each corrected fusion feature, and use the historical feature set corresponding to the corrected fusion feature with the greatest similarity as the facial and body feature set (or facial feature set) corresponding to the target user.

[0063] Specifically, if the new captured record has feature records with similarity exceeding the threshold and the number of records is greater than 1, the following algorithm is used:

[0064] Assume that the similarity s1, s2, …, sn of the new feature and f1, f2, …, fn in the library exceeds the threshold, and the corresponding archive IDs are vid1, vid2, …, vidn respectively;

[0065] For each virtual file, it is formed by a series of face capture records. These face features are fused into a fused feature to represent this file. We will fused the original record number of this fused feature into a fused feature number. Assume that the fused feature number of vidk (k = 1, 2, .., n) is Ck, and the sum of these n fused feature numbers is Sum = Sigma (Ck), k = 1, 2, ..., n. Calculate the softmax value of these Ck: P_k = exp (Ck) / Sigma (exp (Ci)), i from 1 to n, P_k represents the probability of the occurrence of class k; for example, if the fused feature numbers of four virtual files are [1, 2, 3, 4] respectively, then the softmax value is calculated as [0.0321, 0.0871, 0.2369, 0.6439], and the sum of each value is 1, and it has the effect of amplifying the gap. Calculate the following value for each label vidk: Gain correction is performed on the similarity sk to obtain the modified fused feature Vk, which reduces the influence of excessive similarity caused by the difference of accidental sampling of the image.

[0066] Vk=sk+Delta_k, where Delta_k is the characteristic gain, which can be positive or negative.

[0067] The characteristic gain Delta_k is calculated as follows:

[0068] Delta_k=(1-sk)*(Ck / Sum) +P_k*(-ln(Ik / 5)), where Ik is the time of the last fusion of the label and is sorted from new to old. The most recent value is 1 and the next one is 2.

[0069] Here, Ck / Sum is a positive incentive for the proportion of the number of fused features. The more fused features there are, the greater the increase.

[0070] -ln(Ik / 5) means that the closer the time is, the more weight incentives are given. The corresponding values ​​of Ik=1,2,3,4,5,6 are 1.61,0.92, 0.51, 0.22, 0, -0.18. It can be seen that the label classes that have been merged recently have positive incentives, the second and later have a reducing effect, and the sixth in time have a weakening effect on feature similarity. P_k*(-ln(Ik / 5)) has an inhibitory effect on the number of clusters with fewer clusters.

[0071] The above method can correct the similarity between the facial features in the new face record and the fusion features of each virtual file, and increase the similarity of the category with the most recent update and the largest number of fusion features. It can suppress the occasional sampling differences and reduce the situation of one person being classified into multiple categories.

[0072] Select the largest value of Vk, take the corresponding label as the label of the new feature, add 1 to its fusion count, and update the last time.

[0073] In summary, the clustering unit selection method can shield the impact of the number and fusion time on class merging, promote its merging to the newest and most numerous classes, and ensure its stability.

[0074] It should be noted that the key step in the face recognition algorithm is the extraction of face features. One feature identifies a face image. In the training phase, the definition of the loss function for measuring the training phase has transitioned from the previous softmax to cosface and arcface in recent years. Its essence is a kind of maximization of the classification boundary in cosine space and angle space. Under this mechanism, we use cosine distance (value 1 minus cosine similarity) as the unit to measure its similarity. In the training phase, the extracted features will require that the cosine distance of multiple photos identified as the same person is as small as possible, while the cosine distance of features of different people is as large as possible. Finally, the cosine distance of the face features extracted by the trained face recognition algorithm for the same person should be close to 0, that is, the cosine similarity should be close to 1. Human body features have similar characteristics.

[0075] It should be noted that for human features, if the library is empty or has no similar features, it is stored in the library and a new label (i.e., a historical feature library) is generated, and the number of fused features is recorded as 1. The newly generated historical feature library is used as the face and body feature set (or body feature set) corresponding to the target user.

[0076] If the similarity between the newly recorded human feature and a feature of a historical feature set in the library exceeds a preset threshold, and the number of feature records is unique, then this new feature label adopts the feature label compared in the library, the number of fused features increases by 1, and the feature in the library is fused with the new feature, that is, the human feature is fused with the feature whose similarity exceeds the preset threshold to obtain the fused feature, and the historical feature set corresponding to the feature is used as the face and body feature set (or body feature set) corresponding to the target user.

[0077] If the similarity between the human body features and the features in at least two historical feature sets is greater than a preset threshold, determine the fusion features corresponding to each of the at least two historical feature sets; perform gain correction on the fusion features corresponding to each of the at least two historical feature sets to obtain the corrected fusion features corresponding to each of the at least two historical feature sets; determine the similarity between the human body features and each corrected fusion feature, and use the historical feature set corresponding to the corrected fusion feature with the greatest similarity as the face and body feature set (or human body feature set) corresponding to the target user.

[0078] The trajectory determination unit is used to determine the target user's action trajectory based on the face and body feature set corresponding to the target user, the location attributes corresponding to each face feature and body feature in the face and body feature set, and the collection time, collection location and collection time. It can be understood that the trajectory determination unit can be used to associate a person's records at different times and locations, so that the trajectory of a person in a certain period of time can be characterized, and the travel pattern can also be characterized, such as the time of day, where to start, and where to go after passing through fixed points.

[0079] The identity prediction unit is used to determine the identity type of the target user according to the identity characteristics and the user characteristics. The identity type of the target user can be understood as the identity type of the target user. For example, the identity category can include day shift workers, night shift workers, vendors, takeaways, scavengers, on-duty, unknown, leisure residents, etc. For example, the field of social governance needs to understand the people-centered overview of the jurisdiction, and needs to obtain the approximate occupational categories, age groups, etc. of pedestrians on the road. For occupational categories, we divide them into day shift workers, night shift workers, vendors, takeaways, scavengers, on-duty, unknown, leisure residents, etc. Among them, day shift workers and night shift workers have a lower priority, only for people who go out early and return late or hide during the day and go out at night. For night vendors, the target should be classified as vendors; assuming that the same person appears at night, almost does not move for a long time, stands, stops around pedestrians, wears an apron, etc., then the identity prediction unit can determine the identity type of the target user, vendor, according to the identity characteristics and user characteristics of the user.

[0080] Specifically, the identity prediction unit includes a random forest model, wherein the random forest model includes a plurality of decision tree models, and the dimension weights corresponding to each decision tree model are different. In this embodiment, the random forest model can be pruned and trained in advance according to the collected training data, and the number and depth parameters of the trees in the random forest model can be determined according to the classification effect, for example, the number of decision trees in the random forest is finally set to 12, and the depth of the tree is 7.

[0081] The identity prediction unit can be used to input the identity features and the user features into each decision tree model respectively to obtain the identity prediction type of the target user output by each decision tree model; determine the identity type of the target user according to the identity prediction type of the target user output by each decision tree model. For example, the identity prediction type with the highest frequency of occurrence among the identity prediction types of the target user output by all decision tree models can be used as the identity type of the target user.

[0082] For example, the identity prediction unit obtains user features of the target user, such as facial attributes (age group [supports automatic segmentation], gender, beard, hairstyle, glasses, mask, race, etc.), body attributes (shoulder bag, backpack, handbag, top color, top type, trouser color, trouser type, helmet color, apron, takeaway courier uniform (unknown, JD.com, Meituan, Ele.me)), camera area attributes or location attributes or collection locations (commercial area, office building, residential area, main road, secondary road, secondary branch road, etc.); and obtains the identity features of the target user. The identity prediction unit can input the above user characteristics and identity characteristics into each decision tree model respectively to obtain the identity prediction type of the target user output by each decision tree model; and use the identity prediction type with the highest frequency among the identity prediction types of the target user output by all decision tree models as the identity type of the target user.

[0083] It should be noted that, in the present embodiment, the identity type prediction algorithm selected by the identity prediction unit is not limited to the random forest algorithm, and GBDT can also be used for model training and prediction of newly collected data. As the number of collections increases, the range of features can be expanded, and neural networks can be used for feature selection and automatic extraction of behavioral features. All can be used as replacement implementation methods. Model training can be performed using a large amount of training data (as well as identity features and the user features and the corresponding identity types), and machine learning or deep learning methods can be performed. One implementation is statistical learning through random forests of machine learning, but is not limited to this.

[0084] It should be noted that the system may further include a storage unit. The storage unit may be used to store the image information to be processed, the user characteristics, the identity characteristics and the identity type of the target user, so that other units and devices in the identity type recognition system can read or store data from the storage unit at any time.

[0085] It can be seen from the above technical scheme that the present application provides an identity type recognition system, the system comprising: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit; the image acquisition device is used to acquire image information to be processed; the image analysis unit is used to extract features from the image information to be processed to obtain user features of the target user; the feature determination unit is used to determine identity features based on the user features; the identity prediction unit is used to determine the identity type of the target user based on the identity features and the user features. It can be seen that in this embodiment, after the image acquisition device acquires the image information to be processed, the image analysis unit, the feature determination unit and the identity prediction unit can be used to process the image information to be processed to obtain the identity type of the target user in the image information to be processed. In this way, the present application can use the identity type recognition system to automatically identify the identity type of the user, without the need to manually identify the identity of the user as in the prior art, avoiding the subjectivity, one-sidedness and limitations caused by manual experience evaluation, as well as the problem of identity recognition errors caused by manual negligence, thereby improving the efficiency of user identity type recognition and thus improving user experience.

[0086] That is to say, the identity recognition system in this embodiment includes an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit, and the feature determination unit includes a behavior recognition unit, a clustering unit and a trajectory determination unit, wherein the clustering unit, the image analysis unit and the identity prediction unit are core modules. The image analysis unit is used to analyze the attributes and features of the human face and human body in the video and image, such as age, gender, clothing, backpack, etc. The focus is on the recognition of uniform clothing and helmets for takeout and express delivery. The clustering unit provides the function of marking a person with a unique virtual label, thereby associating a person's multiple appearances in multiple places and performing trajectory characterization. Here, an algorithm is proposed to gradually stabilize and normalize multiple labels for one person, greatly reducing the problem of being unable to uniquely characterize one person. The identity prediction unit uses the multi-dimensional attributes extracted above, combined with the long-term attributes of travel rules and the short-term attributes of behavioral characteristics for identification, to achieve occupational prediction of the current active person, thereby providing data support for the current human-centered overall portrait of the jurisdiction.

[0087] It should be noted that each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. The device and system embodiments described above are merely schematic, and the units described as separate components may or may not be physically separated. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art can understand and implement it without paying creative work.

[0088] The above is only a preferred specific implementation method, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in this application should be included in the protection scope of this application. Therefore, the protection scope of this application should be based on the protection scope of the claims.

Claims

1. An identity type recognition system, characterized in that: The system comprises: an image acquisition device, an image analysis unit, a feature determination unit and an identity prediction unit; The image acquisition device is used to acquire image information to be processed; The image analysis unit is used to extract features from the image information to be processed to obtain user features of the target user; The feature determination unit is used to determine the identity feature of the target user according to the user feature; The identity prediction unit is used to determine the identity type of the target user according to the identity characteristics and the user characteristics; The user characteristics include face images, face features, face attributes, body images, body features, body attributes, behavior and action features, and location attributes; The identity features include target behavior, action trajectory and face and body feature set; the feature determination unit includes a behavior recognition unit, a clustering unit and a trajectory determination unit; The behavior recognition unit is used to perform behavior recognition according to the behavior action characteristics to obtain the target behavior; The clustering unit is used to perform clustering processing based on facial features and body features to obtain a facial and body feature set corresponding to the target user; The trajectory determination unit is used to determine the movement trajectory of the target user according to the face and body feature set corresponding to the target user and the position attributes, collection positions and collection times respectively corresponding to the face features and body features in the face and body feature set; The clustering unit is specifically used to determine the fusion features corresponding to each of the at least two historical feature sets if the similarity between the facial features or the body features and the features in at least two historical feature sets is greater than a preset threshold; perform gain correction on the fusion features corresponding to each of the at least two historical feature sets to obtain the corrected fusion features corresponding to each of the at least two historical feature sets; determine the similarity between the facial features or the body features and each corrected fusion feature, and use the historical feature set corresponding to the corrected fusion feature with the greatest similarity as the facial and body feature set corresponding to the target user.

2. The system according to claim 1, characterized in that The image information to be processed includes images and / or videos, and acquisition locations and acquisition times corresponding to the images and / or videos.

3. The system according to claim 1, characterized in that The image analysis unit includes a trained user feature recognition model; The image analysis unit is used to input the image information to be processed into the user feature recognition model to obtain the user features of the image information to be processed.

4. The system according to claim 1, characterized in that The behavior recognition unit is used to determine the target behavior corresponding to the behavior action feature according to the behavior action feature and the location attribute, collection location and collection time corresponding to the behavior action feature; and / or, The behavior recognition unit includes a trained behavior recognition model; the behavior recognition unit is used to input the behavior action feature into the behavior recognition model to obtain the target behavior corresponding to the behavior action feature.

5. The system according to claim 1, characterized in that The clustering unit is used to determine the similarity between the facial features, the body features and the features in each historical feature set; based on the similarity between the facial features, the body features and the features in each historical feature set, determine the facial and body feature set corresponding to the target user.

6. The system according to claim 1, characterized in that The identity prediction unit includes a random forest model, wherein the random forest model includes a plurality of decision tree models, and the dimension weights corresponding to each decision tree model are different; The identity prediction unit is used to input the identity feature and the user feature into each decision tree model respectively to obtain the identity prediction type of the target user output by each decision tree model; The identity type of the target user is determined according to the predicted identity type of the target user output by each decision tree model.

7. The system according to any one of claims 1 to 3, characterized in that: The system further comprises a storage unit for storing the image information to be processed, the user characteristics, the identity characteristics and the identity type of the target user.

Images