Object automatic labeling method and system suitable for gate channel

Abstract

The invention discloses an automatic object labeling method and system suitable for a gate channel. The method comprises the following steps: constructing an initial data set based on gate passing characteristics and a priority strategy; constructing an automatic labeling model and performing model training by using the initial data set; and performing target detection on the to-be-labeled image by using the automatic labeling model, and taking pedestrian target information in a target detection result as labeling information of the to-be-labeled image. According to the method, the priority strategy is introduced into the initial data set of the automatic labeling model, so that labeling results have high consistency when shielding and overlapping occur. The model can generate the labeling box according to the preset priority rule regardless of the condition that adults, adults and children or adults carry articles, so that the problem of non-uniform labeling standards caused by human subjective judgment is avoided. And by introducing the priority loss, the automatic labeling model pays more attention to the features of important positions, and the recognition capability in a complex scene is improved.

Description

Technical Field

[0001] The present invention belongs to the field of gate control and machine vision technology, and in particular, relates to an automatic object labeling method and system suitable for gate channels. Background Art

[0002] The turnstile is a device used for area separation and access control. By combining authorization verification and payment transactions, the turnstile can divide the separation area into paid area and non-paid area, authorized area and non-authorized area, so as to achieve effective access management. The turnstile is widely used in various access control scenarios such as automatic ticket vending and checking systems (AFC), building management, and venues. The passage logic technology of the turnstile is the technical basis for evaluating its passage detection performance, and it is also the core component to ensure the passage efficiency and safety reliability of the turnstile.

[0003] With the continuous advancement of science and technology, the access logic technology of gate machines has gradually shifted from traditional sensor solutions to machine vision solutions. In recent years, deep learning methods that use depth images obtained by depth cameras and combine them with target detection and tracking algorithms have begun to attract attention. In the access logic technology of gate machines based on machine vision, the accuracy of target detection and tracking by machine learning models is a key indicator for measuring its access efficiency and safety, and the accuracy of the model depends on high-quality training data sets.

[0004] However, for special equipment application scenarios with limited viewing angles such as gate access, there are currently no publicly available depth image datasets and no universal annotation specifications. This results in the possibility that multiple annotators may apply different annotation criteria during the long annotation process in the annotation of the depth map dataset, resulting in many problems such as difficulty in converging the model training based on the annotated dataset and low confidence. For example, in the channel, there are situations where pedestrians and other pedestrians and objects (such as luggage, carts, etc.) block each other. The target detection model trained based on the traditional annotation method will tend to have non-overlapping detection frames or directly maximize the detection frame, or the rules are not unified, that is, sometimes the hands are annotated as part of the human body, but sometimes they are ignored. This will result in low generalization and accuracy of the model, and in severe cases, it may even cause the model to not converge.

[0005] An excellent model training data set should contain a large amount of accurately annotated image data. The annotation accuracy of image data directly affects the training effect and detection performance of the model. Therefore, providing a high-quality automatic annotation solution for target objects in the gate channel is one of the key means to reduce annotation costs, increase the annotation speed of annotators, and improve model training effects and detection performance. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a method and system for automatic object labeling suitable for gate channels, aiming to train an automatic labeling model. The automatic labeling model learns a priority-based detection frame generation paradigm through a data set combined with the gate passage characteristics, so that the output detection frames will not affect each other's position judgment.

[0007] The technical solution provided by the present invention is as follows: An automatic object labeling method applicable to a gate channel comprises the following steps: S1, construct the initial data set based on the gate traffic characteristics and priority strategy; S2, build an automatic annotation model and use the initial data set for model training; S3. Use the automatic annotation model to perform target detection on the image to be annotated, and use the pedestrian target information in the target detection result as the annotation information of the image to be annotated.

[0008] Furthermore, in step S1, in the depth image of the gate channel from a bird's-eye view collected, all human objects need to be identified and marked during annotation to ensure that the complete human pixels of each person are included in their corresponding annotation boxes as much as possible, while excluding other irrelevant objects and personnel; overlap between annotation boxes is avoided as much as possible, and for unavoidable overlapping areas, the integrity of the annotation boxes of high-priority areas is guaranteed according to the annotation priority strategy, and the annotation boxes of low-priority areas are adjusted or reduced.

[0009] Furthermore, the labeling priority strategy includes a human body part priority strategy: for human body parts, the head has the highest priority. In any case, the head of an adult or child must appear completely in its own labeling box and cannot appear in the labeling box of others; the torso has the second highest priority. The area above the shoulders has a higher priority than the chest, and the chest has a higher priority than the waist; the limbs have the lowest priority and will be discarded when necessary to ensure the complete labeling of high-priority parts.

[0010] Furthermore, the labeling priority strategy includes a location priority strategy: for the location of the object, in the same image, the pedestrian located in the center of the gate channel has a higher labeling priority than the pedestrians located at the edge. The integrity of the high-priority object labeling box should be ensured, and the labeling box of the object with lower priority should avoid occluding the labeling box of the object with higher priority.

[0011] Furthermore, the labeling priority strategy includes a person identity priority strategy: when a child is blocked by an adult or the boundary between the child and the adult is not obvious, the integrity of the adult labeling box is prioritized.

[0012] Furthermore, when labeling, the clothes worn by pedestrians are considered as part of the human body, unless there is a clear color boundary or shape feature between the clothes and the body; if the personal belongings carried by pedestrians have a clear boundary between the human body, these items should be excluded from the labeling box while ensuring the integrity of the adult's body; when children are in close contact with adults and there is no obvious boundary between the child's head and the adult, the child is considered as the adult's personal belongings. If there is a clear boundary between the child and the adult, a separate labeling box should be drawn for the child, and the adult's labeling box should not include the child's head pixels.

[0013] Furthermore, in step S2, the automatic labeling model adopts the network structure of YOLOv8, and the model loss function introduces a priority loss term based on YOLOv8: , Where, L total is the loss function of the automatic labeling model; L cls Represents the classification loss, which is used to measure the difference between the predicted category and the true category; L obj Represents the confidence loss, which is used to measure the confidence of whether the predicted bounding box contains the target; L box represents the bounding box regression loss, which is used to measure the position and size difference between the predicted box and the true box; represents the priority loss, and , , in, P(x,y) is the priority weight calculated based on the center coordinates (x, y) of the bounding box. W and H are the width and height of the image respectively.

[0014] Furthermore, in step S3, the pedestrian target detection results of the automatic annotation model are screened to exclude interference targets; the screening process includes noise detection: calculating the pixel intensity mean and standard deviation of the pedestrian target image area. If the pixel intensity mean or standard deviation exceeds the set threshold, it indicates that the pedestrian target detected by the automatic annotation model is an interference target and should not be included in the data set as an annotation sample; the screening process also includes pattern matching: calculating the correlation between the pedestrian target image area and the known noise pattern. If the correlation exceeds the set threshold, it indicates that the pedestrian target detected by the automatic annotation model is an interference target.

[0015] An object automatic annotation system based on the above method includes an initial data set construction module, an automatic annotation model training module and an automatic annotation module; The initial data set construction module includes an image display module, a manual annotation module and an annotation information storage module. The annotator uses the annotation tool provided by the manual annotation module to annotate the gate channel image displayed in the image display module with pedestrian targets according to the priority strategy, and then the manual annotation module saves the corresponding annotation information to the annotation information storage module; the automatic annotation model training module is used to train the constructed automatic annotation model through the initial data set, including a training parameter setting module and a training process visualization module; the automatic annotation module uses the trained automatic annotation model to perform target detection on the image to be annotated, and uses the pedestrian target information in the target detection result as the annotation information of the image to be annotated.

[0016] Preferably, the automatic labeling module further includes a screening processing module for screening pedestrian targets in the target detection results to exclude interfering targets.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: Reduce the annotation cost of the data set: The automatic annotation model in the present invention uses a pre-learned priority strategy to judge the subject, and can then give a bounding box judgment with high consistency, thereby reducing position errors caused by overlaps, and reducing the annotation review cost, model training cost, field application cost, etc. in the entire project process.

[0018] Stronger adaptability: Gate passage may face a variety of complex scenarios, such as different objects, high density, different overlaps or interactions, etc. The automatic annotation model in the present invention takes these complexities into account based on a specific data set during training, so that it can be standardized and unified in various scenarios, which also makes the model easier to converge.

[0019] Improve the accuracy of data set annotation: False positives and false negatives are one of the main challenges in the field of machine vision, and overlapping areas caused by interference between multiple objects are often areas with high incidence of difficult examples. In order to improve the accuracy of difficult example detection, the automatic annotation model in the present invention can provide consistent annotation references for annotation personnel for different scenes to be annotated, thereby improving the quality of the final data set and indirectly ensuring the smooth operation of the gate system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0021] Figure 1 It is a schematic diagram of the process flow of an automatic object labeling method provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a priority strategy provided by an embodiment of the present invention; Figure 3 is a marking example A provided by an embodiment of the present invention; Figure 4 is a marking example B provided by an embodiment of the present invention; Figure 5 is a marking example C provided by an embodiment of the present invention; Figure 6 is a marking example D provided by an embodiment of the present invention; Figure 7 is a marking example E provided by an embodiment of the present invention; Figure 8 is a marking example F provided by an embodiment of the present invention; Fig. 9 is a marking example G provided by an embodiment of the present invention; Fig.10 is a labeling example H provided by an embodiment of the present invention; Fig.11 is a marking example 1 provided by an embodiment of the present invention; Fig.12 This is a marking example J provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative work are all within the scope of protection of the present invention.

[0023] Embodiment 1 This embodiment provides an object automatic labeling method applicable to gate channels, such as Figure 1 As shown, the method mainly includes the following steps: Construct an initial data set based on gate traffic characteristics and priority strategies; Build an automatic annotation model and use the initial dataset for model training; The automatic annotation model is used to perform target detection on the images to be labeled, and the pedestrian target information in the target detection results is used as the annotation information of the images to be labeled.

[0024] 1. Construction of initial dataset 1. Determination of marking objects In the depth image of the gate channel from a bird's-eye view, all human objects need to be identified and marked during annotation. The marked objects include all human individuals that can be identified in the image, including both adults and children. When annotating, it should be ensured that the complete human pixels of each person are included in their corresponding annotation box as much as possible, while excluding other irrelevant objects and people. The initial data set should include various positions of pedestrians in the gate channel and various occlusion situations, and the data set size should be at least 2,000.

[0025] 2. Setting of annotation priority In order to effectively handle the occlusion and overlap between multiple pedestrians in the gate channel, such as Figure 2 As shown in the figure, the concept of annotation priority is introduced, which is defined from three dimensions: body part, location and person identity.

[0026] 2.1. Human body part priority: The head has the highest priority: in any case, the head of an adult or child must appear completely within its own annotation box and cannot appear in the annotation boxes of others. The integrity of the head is crucial to accurately determine the location and identity of the pedestrian. The torso is second: the area above the shoulders has a higher priority than the chest, and the chest has a higher priority than the waist. When annotating, try to include high-priority torso parts. The limbs have the lowest priority: upper arms and thighs, lower arms and lower legs, hands and feet, in that order. The limbs can be discarded when necessary to ensure complete annotation of high-priority parts.

[0027] 2.2. Location priority: Higher priority for the center: Pedestrians in the center of the gate channel in the image have a higher annotation priority than those at the edge. This strategy helps improve the detection accuracy of the central area of ​​the channel and meets the actual application requirements of the gate channel. In the same image, the integrity of the high-priority object annotation box should be guaranteed, and the lower-priority object annotation box should avoid occluding the higher-priority object annotation box, for example: Figure 3 The pedestrian in the middle and lower position has a higher priority, so in order to avoid overlap, the annotation box of the pedestrian in the upper part does not include the entire torso to the right; Figure 4 The position of the pedestrian on the right side of the center has a higher priority, and the annotation box of the pedestrian on the left side does not include the entire torso to the right.

[0028] 2.3. Personnel identity priority Children have a relatively low priority: Due to their short height, it may be difficult to accurately detect them from a bird's-eye view. The child's head must still be completely included in its annotation box, but when it is blocked by an adult or the boundary with the adult is not obvious, priority is given to ensuring the completeness of the adult's annotation.

[0029] In some embodiments, the labeling of pedestrians' clothing and belongings may follow the following strategies: 3.1. Clothing treatment for adults Clothing is considered part of the body: Clothing worn by adults (such as sleeves, hems, skirts, etc.) should be considered part of their body and included in the marked box, unless there is a clear color boundary or shape feature between the clothing and the body. Figure 5 , the adult’s sleeves, hem of clothes, etc. can be identified, and there is no obvious boundary between them and the adult itself, so they all need to be included in the adult marking box.

[0030] 3.2. Adults carrying small objects Carried objects are not included in the marked box: For small items such as handbags, underarm bags, backpacks, etc. carried by adults, if there is a clear boundary between them and the human body, these items should be excluded from the marked box while ensuring the integrity of the adult's body. Figure 6 When an adult carries an underarm bag, there is a clear color dividing line between the edge of the bag and the person. It can be seen that it is not a part of the human body, so the adult body is directly marked to form a cut on the underarm bag.

[0031] 3.3. Adults carrying large objects Processing based on the degree of occlusion: When an adult carries a large object (such as a baby stroller or an oversized item), the object should be labeled based on the degree of occlusion of the human body. If the large object occludes the area below the adult's waist and has a clear color or shape boundary with the human body, the human body should be included as much as possible. If necessary, the labeling box can appropriately include some object pixels, but it should be based on the premise that it does not affect the position and shape of the human body. Figure 7 , an adult pushes a baby stroller forward and regards the baby stroller as an ordinary object, so the adult's head, shoulders, back, etc. are used as dividing lines for marking, and some baby stroller pixels will be included in the marking box.

[0032] In some embodiments, the following strategies can be used to handle occlusion and overlap that occur during the annotation process: 4.1 Collisions at the shoulder and above Priority judgment: When a conflict occurs at the shoulders and above of multiple people, the priority of the location is judged first. The head and shoulders of pedestrians with higher priority should be completely included in their annotation boxes, just as Figure 8 For personnel with the same priority, the conflict area can be evenly cut and the edges of the annotation boxes can be overlapped, as shown in Fig. 9 shown.

[0033] 4.2 Conflicts at and below the waist Reduce the annotation of low-priority parts: When multiple people's waists and below conflict, priority is given to ensuring that the high-priority areas above the chest are fully annotated. The annotation box of low-priority personnel can be appropriately reduced to avoid interfering with the annotation of high-priority personnel. Fig.10 As shown, the arms of the two people conflict with each other (the priority of the arms is lower than that of the waist, so they are collectively referred to as the waist and below). The person with a higher priority in the center is marked first, and then the people on the outside are marked closely to their borders.

[0034] 4.3 Occlusion of Children and Adults Handle according to the situation: When a child is in close contact with an adult and there is no clear boundary between the child’s head and the adult’s, the child should be considered as the adult’s personal belongings. Fig.11 If there is a clear boundary between children and adults, children should be labeled separately, and the label box of adults should not include the pixels of children's heads. Fig.12 shown.

[0035] 4.4 Overlapping of annotation boxes Avoid overlap as much as possible: The annotation boxes should avoid overlap as much as possible to reduce the computational complexity of the model during inference. For unavoidable overlapping areas, a pre-set priority strategy is used to handle them. The annotation boxes of high-priority personnel should be complete, and the annotation boxes of low-priority personnel can be appropriately adjusted or reduced. 2. Construction and training of automatic annotation model After the initial data set is created, this embodiment trains an automatic annotation model based on the data set. In order to meet the needs of pedestrian detection from the overhead perspective of the gate channel, the model structure is optimized based on the standard target detection model, and a priority loss term is introduced into the loss function to ensure that the model can learn and follow the aforementioned priority-based annotation specifications.

[0036] 1. Model structure This embodiment uses the target detection model based on YOLOv8 as the basic structure. The model mainly includes three parts: Backbone, Neck and Head. The Backbone part is responsible for extracting multi-scale feature representations from the input image, including multi-layer convolutional layers, residual modules and CSP (Cross Stage Partial) modules. Feature maps of different scales are obtained by downsampling layer by layer. The Neck part is used to fuse features of different scales to enhance the model's detection ability for targets of different sizes. The structure combining FPN (Feature Pyramid Network) and PANet (Path Aggregation Network) is adopted to achieve top-down and bottom-up feature fusion. By fusing the multi-scale features extracted by the Backbone part, the Neck part can obtain rich contextual information. The Head part is used to classify and locate the fused features (give the detection box parameters), including prediction modules of multiple scales, each of which includes a detection box parameter regression branch and a category probability output branch, and each branch includes a convolution layer and an activation function. For the specific network structure, see YOLOv8.

[0037] 2. Model training 2.1 Data Preprocessing Before training, the initial dataset is augmented and preprocessed, including random cropping, scaling, flipping, color jittering, etc., to increase data diversity and improve the generalization ability of the model.

[0038] 2.2 Loss Function Design In order to enable the model to learn priority-based annotation specifications, a priority loss term is introduced based on the standard YOLOv8 target detection loss. The total loss function L total It includes the following parts: , in, L cls Represents classification loss, which is used to measure the difference between the predicted category and the true category. In this embodiment, cross entropy loss or focal loss is used to solve the problem of imbalance between positive and negative samples. L obj Represents confidence loss, which is used to measure the confidence of whether the predicted bounding box contains the target. In this embodiment, binary cross entropy loss is used. L boxRepresents the bounding box regression loss, which is used to measure the position and size difference between the predicted box and the real box. In this embodiment, CIoU loss (Complete Intersection over Union Loss) is used to comprehensively consider the position, size and shape of the bounding box, that is: , , in, is the intersection-over-union ratio of the predicted box and the true box, is the Euclidean distance between the center of the predicted box and the center of the true box, c is the diagonal length of the minimum closure area containing the predicted box and the true box, v To measure the difference in aspect ratio between the predicted box and the real box, is the weight coefficient, .

[0039] In order to make the model focus on high priority target locations, the priority loss is defined , encouraging the model to give more accurate predictions in high-priority areas. Since the application scenario is pedestrian detection in the gate channel, the image is taken from a bird's-eye view of the gate channel. The priority of the central area of ​​the gate channel should be higher than that of the edge area, so the priority loss is set as: , , in, L box is the bounding box regression loss, P(x,y) is the priority weight calculated based on the center coordinates (x, y) of the bounding box. W and H are the width and height of the sample image respectively. The closer to the center of the image, the higher the priority. P(x,y) The value range is between 0 and 1, with the center position having the highest priority of 1 and the edge position having the lowest priority of 0.

[0040] 2.3 Model Training Strategy During the training process, multi-scale training, random data augmentation and other methods are used to improve the robustness of the model. Adam or SGD optimizer is used to update the model parameters according to the gradient of the loss function.

[0041] In order to ensure the convergence and stability of the model, the cosine annealing learning rate scheduling strategy is adopted. The initial value of the learning rate is set to lr0, and then the learning rate is gradually reduced during the training process:

[0042] in, is the minimum learning rate, is the current training step number, is the maximum number of training steps.

[0043] 2.4 Model Evaluation Indicators The model performance was evaluated on the validation set, using mAP (mean Average Precision) as the main evaluation indicator. At the same time, the detection accuracy of the model in high-priority areas was focused on, and the AP values ​​under different priorities were calculated. The effectiveness of the priority loss was verified by comparing the performance changes of the model before and after adding the priority loss.

[0044] 3. Automatic Labeling The trained automatic annotation model is used to perform target detection on the image to be annotated, and the pedestrian annotation information (including the position and size of the annotation box) in the target detection result is used as the annotation information of the image to be annotated.

[0045] After using the automatic labeling model to detect targets, in order to further improve the labeling accuracy, pedestrian targets can be screened to eliminate interference (such as large areas of dust, light spots or other non-personnel images captured by the camera) to avoid contamination of the final data set and ensure the accuracy of the data set.

[0046] The screening process mainly includes two methods: noise detection and pattern matching.

[0047] Noise Check: Calculate the mean and standard deviation of the pixel intensity of the pedestrian target image area. If the mean or standard deviation of the pixel intensity exceeds the set threshold, it indicates that the pedestrian target detected by the automatic annotation model is an interference target and is not suitable for inclusion in the dataset as an annotation sample. For the pixel points in the pedestrian target detection box, calculate the mean and standard deviation of their pixel intensity values. If it exceeds the set threshold, or , it is judged as noise and excluded. The formula is as follows: , in, N is the number of pixels in the detection box, I( x i , y i ) is the pixel point ( x i , y i ) intensity value, is the mean threshold, is the standard deviation threshold.

[0048] Pattern Matching: Calculate the pedestrian target image areaI With known noise patterns P noise If the correlation exceeds the set threshold Tpattern, it indicates that the pedestrian target detected by the automatic labeling model is an interference target and is excluded. The formula is as follows: , in, I is the intensity matrix of the pedestrian target image area, P noise is a known noise pattern matrix. This noise pattern may be predefined, such as a matrix form corresponding to a salt and pepper noise pattern, a Gaussian noise pattern, etc. As a correlation calculation function, the Pearson correlation coefficient or the like may be used.

[0049] The above is the main content of the object automatic labeling method applicable to the gate channel provided by this embodiment. The automatic labeling model obtained after the above training process shows excellent performance in the pedestrian detection task under the bird's-eye view of the gate channel. The automatic labeling model can automatically identify and label pedestrians in the image, greatly reducing the workload of manual labeling. Experimental data show that after using this method, the efficiency of the labeling team has increased by about 70%, thereby shortening the project cycle and reducing labor costs. By introducing a priority strategy in the initial data set of the automatic labeling model, the labeling results of the model in dealing with occlusion and overlap are highly consistent. Whether it is an adult and an adult, an adult and a child, or an adult carrying items, the model can generate a labeling box according to a pre-set priority rule, avoiding the problem of inconsistent labeling standards due to human subjective judgment. The automatic labeling model has higher detection accuracy in high priority areas (such as head and torso). By introducing priority loss, the model pays more attention to the features of important positions and improves the recognition ability in complex scenes. According to the evaluation results of the test set: the overall detection accuracy (mAP) reached 95%, the detection accuracy of the head area increased to 98%, and the detection accuracy in occlusion conditions increased by about 45%.

[0050] Embodiment 2 Based on the above method, this embodiment provides an automatic object labeling system suitable for gate channels, which mainly includes: an initial data set construction module, an automatic labeling model training module and an automatic labeling module.

[0051] The initial dataset construction module includes an image display module, a manual annotation module and an annotation information storage module. The annotator uses the annotation tools provided by the manual annotation module to annotate pedestrian targets according to the priority strategy on the gate channel image displayed in the image display module, and then the manual annotation module saves the corresponding annotation information to the annotation information storage module.

[0052] The automatic labeling model training module is used to train the constructed automatic labeling model through the initial data set, including a training parameter setting module and a training process visualization module. The training process visualization module can display the training loss, iteration rounds, etc. in real time.

[0053] The automatic labeling module uses the trained automatic labeling model to perform target detection on the image to be labeled, and uses the pedestrian target information (the position and size of the labeling box) in the target detection result as the labeling information of the image to be labeled. In some embodiments, the automatic labeling module also includes a screening processing module for screening the pedestrian targets in the target detection results to exclude interference targets caused by factors such as dust and light.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may also be combined, the steps may be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for automatically labeling objects in gate channels, characterized in that: The following steps are involved: S1, construct the initial data set based on the gate traffic characteristics and priority strategy; S2, build an automatic annotation model and use the initial data set for model training; S3. Use the automatic annotation model to perform target detection on the image to be annotated, and use the pedestrian target information in the target detection result as the annotation information of the image to be annotated.

2. The automatic marking method according to claim 1, characterized in that: In step S1, in the depth image of the gate channel from a bird's-eye view, all human objects need to be identified and marked during annotation to ensure that each person's complete human pixels are included in their corresponding annotation boxes, while excluding other irrelevant objects and personnel; overlap between annotation boxes is avoided as much as possible. For unavoidable overlapping areas, the annotation box integrity of the high-priority area is guaranteed according to the annotation priority strategy, and the annotation box of the low-priority area is adjusted or reduced.

3. The automatic marking method according to claim 2, characterized in that: The labeling priority strategy includes a human body part priority strategy: for human body parts, the head has the highest priority. In any case, the head of an adult or child must appear completely in its own labeling box and cannot appear in the labeling box of others; the torso has the second highest priority. The area above the shoulders has a higher priority than the chest, and the chest has a higher priority than the waist; the limbs have the lowest priority.

4. The automatic marking method according to claim 2, characterized in that: The labeling priority strategy includes a location priority strategy: for the location of the object, in the same image, the pedestrian located in the center of the gate channel has a higher labeling priority than the pedestrian located at the edge. The integrity of the high-priority object labeling box should be ensured, and the labeling box of the object with lower priority should avoid occluding the labeling box of the object with higher priority.

5. The automatic marking method according to claim 2, characterized in that: The labeling priority strategy includes a person identity priority strategy: when a child is blocked by an adult or the boundary between the child and the adult is not obvious, the integrity of the adult labeling box is prioritized.

6. The automatic marking method according to claim 2, characterized in that: When labeling, the clothes worn by pedestrians are considered as part of the human body, unless there is a clear color boundary or shape feature between the clothes and the body; if the personal belongings carried by pedestrians have a clear boundary between the human body, these items should be excluded from the labeling box while ensuring the integrity of the adult's body; when children are in close contact with adults and there is no obvious boundary between the child's head and the adult, the child is considered as the adult's personal belongings. If there is a clear boundary between the child and the adult, a separate labeling box should be drawn for the child, and the adult's labeling box should not include the child's head pixels.

7. The automatic marking method according to claim 1, characterized in that: In step S2, the automatic labeling model adopts the network structure of YOLOv8, and the model loss function introduces the priority loss term based on YOLOv8: ， Where, L total is the loss function of the automatic labeling model; L cls Represents the classification loss, which is used to measure the difference between the predicted category and the true category; L obj Represents the confidence loss, which is used to measure the confidence of whether the predicted bounding box contains the target; L box represents the bounding box regression loss, which is used to measure the position and size difference between the predicted box and the true box; represents the priority loss, and ， ， in, P(x,y) is the priority weight calculated based on the center coordinates (x, y) of the bounding box. W and H are the width and height of the image respectively.

8. The automatic marking method according to claim 1, characterized in that: In step S3, the pedestrian target detection results of the automatic annotation model are screened to exclude interference targets; the screening process includes noise detection: calculating the pixel intensity mean and standard deviation of the pedestrian target image area. If the pixel intensity mean or standard deviation exceeds the set threshold, it indicates that the pedestrian target detected by the automatic annotation model is an interference target and should not be included in the data set as an annotation sample; the screening process also includes pattern matching: calculating the correlation between the pedestrian target image area and the known noise pattern. If the correlation exceeds the set threshold, it indicates that the pedestrian target detected by the automatic annotation model is an interference target.

9. An automatic object annotation system based on the method according to any one of claims 1 to 8, characterized in that: It includes the initial data set construction module, the automatic annotation model training module and the automatic annotation module; The initial data set construction module includes an image display module, a manual annotation module and an annotation information storage module. The annotator uses the annotation tool provided by the manual annotation module to annotate the gate channel image displayed in the image display module with pedestrian targets according to the priority strategy, and then the manual annotation module saves the corresponding annotation information to the annotation information storage module; the automatic annotation model training module is used to train the constructed automatic annotation model through the initial data set, including a training parameter setting module and a training process visualization module; the automatic annotation module uses the trained automatic annotation model to perform target detection on the image to be annotated, and uses the pedestrian target information in the target detection result as the annotation information of the image to be annotated.

10. The object automatic labeling system according to claim 9, characterized in that: The automatic labeling module also includes a screening processing module for screening pedestrian targets in the target detection results to exclude interfering targets.

Images