Lightweight personnel wearing equipment intelligent identification method under small sample condition

Abstract

The invention relates to a lightweight personnel wearing equipment intelligent identification method under a small sample condition. The method comprises the following steps: collecting a video frame image of an industrial production environment; carrying out human body posture estimation on the video frame image, judging whether a person exists or not, if so, extracting human body key points, and carrying out smooth processing to form human body skeleton information, and if not, re-collecting the video frame image; performing target detection on the wearing equipment area of the human body, performing fusion with set key wearing area cutting, and associating to a corresponding human body target to obtain a wearing equipment area image; carrying out wearing feature extraction on the wearing equipment area image, and constructing a wearing feature library; and obtaining the wearing features of the to-be-identified wearing equipment area image, carrying out retrieval matching on the wearing feature library, and carrying out multi-frame integration optimization by adopting a hysteresis comparison strategy to obtain a wearing identification node of the wearing equipment. Compared with the prior art, the method has the advantages of real-time light weight, accurate result and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a lightweight personnel wearing equipment intelligent identification method under small sample conditions. BACKGROUND

[0002] In modern industrial production, safety production problems have always been the core issue of high concern of enterprises and society. With the increasingly complex industrial production environment and the continuous expansion of equipment scale, the potential safety risks in the production process have increased significantly, and the correct wearing of safety protective equipment by personnel is crucial to reducing the incidence of safety accidents. Traditional supervision methods mainly rely on manual inspection, which has problems such as high labor intensity, high misjudgment rate, poor real-time performance, and high supervision cost. In recent years, intelligent monitoring technology based on deep learning has been gradually applied to the field of safety production management, which can automatically identify the wearing situation of personnel and improve the automation level of safety supervision.

[0003] However, the existing intelligent identification method still has many problems: first, the existing method relies on large-scale labeled data, and current deep learning methods usually require a large amount of manually labeled training data, but in industrial environments, data collection is difficult, and it is difficult to build a large-scale data set. Second, the existing method is difficult to quickly adapt to different industrial scenarios, and the types, colors, and shapes of equipment worn by workers vary greatly in different production environments. Traditional models often need to be retrained when the scene is transferred, which is time-consuming and labor-intensive. Finally, the existing method consumes a lot of computing resources and mainly relies on high-performance GPU servers, but in industrial environments, computing resources are limited and it is difficult to efficiently deploy on edge devices, affecting the real-time performance and availability of detection. SUMMARY

[0004] The purpose of the present application is to provide a lightweight personnel wearing equipment intelligent identification method under small sample conditions that can stably operate in complex industrial scenarios.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A lightweight personnel wearing equipment intelligent identification method under small sample conditions, comprising the following steps:

[0007] Collecting video frame images of industrial production environments;

[0008] Using a Modified-YOLOv8-Pose network to perform human pose estimation on the video frame images, determining whether there is a person, if there is a person, extracting human key points and performing smoothing processing to form human skeleton information, if there is no person, re-collecting video frame images;

[0009] Based on the human skeleton information, the wearing equipment region of the human body is detected, and the set key wearing region is cropped and fused, and is associated to the corresponding human body target to obtain a wearing equipment region image;

[0010] An improved LCNet lightweight feature extraction network is used to extract wearing features of the wearing equipment region image, and a wearing feature library is constructed.

[0011] The wearing features of the wearing equipment region image to be recognized are obtained, the wearing feature library is searched and matched, and a hysteresis comparison strategy is used for multi-frame integration optimization to obtain a wearing equipment recognition result.

[0012] Further, the Modified-YOLOv8-Pose network is improved by introducing an HGBlock module, a C-BRAM module and a pose loss function in the YOLOv8-Pose network, the Modified-YOLOv8-Pose network includes a Backbone part, a Neck part and a Head part, the Backbone part includes a first CBS module, a second CBS module, a first C2f module, a third CBS module, a first HG Block module, a fourth CBS module, a second C2f module, a C-BRAM module and a SPPF module connected in sequence, the Neck part includes a first up-sampling module, a first connection module, a third C2f module, a second up-sampling module, a second connection module, a second HG Block module, a fifth CBS module, a third connection module, a third HG Block module, a sixth CBS module, a fourth connection module and a fourth HG Block module connected in sequence with the SPPF module, the first connection module is further connected with the first HG Block module, the third C2f module is further connected with the third connection module, the second connection module is further connected with the first C2f module, the fourth connection module is further connected with the SPPF module, and the Head part includes three Pose modules, each Pose module is connected with a corresponding second HG Block module, third HG Block module and fourth HG Block module,

[0013] The first HG Block module, the second HG Block module, the third HG Block module and the fourth HG Block module each include a plurality of continuous convolution layers for fusing multi-scale information and supporting residual connection.

[0014] The C-BRAM module includes a lightweight channel attention submodule and a spatial attention double-layer routing submodule, and the processing process of the lightweight channel attention submodule includes:

[0015] The C-BRAM module-based pre-C2F module processes the feature map to extract global information of each channel and obtain a channel description vector;

[0016] Based on the channel description vector, an initial inter-channel weight is calculated, and a final inter-channel weight is obtained by Sigmoid normalization.

[0017] The processing process of the spatial attention double-layer routing sub-module includes:

[0018] The channel attention weighted feature map is divided into regions, and Query, Key and Value values are generated by linear projection, and a regional-level routing is constructed.

[0019] Based on the regional-level routing, a regional affinity is calculated, a coarse-grained screening is performed, and a spatial weight is obtained through token-level fine-grained attention calculation.

[0020] Further, the expression of the posture loss function is:

[0021]

[0022] In the formula, L JD is a loss value, L distance is a loss based on Euclidean distance, λ is an adjustment coefficient for balancing the weight of distance loss and direction loss, L direction is a loss based on joint direction, N is the total number of key points of a single human body, d i is the Euclidean distance between the predicted and true positions of the i-th key point, A j is the area of the region of the human body where the key point is located, σ i is the fixed standard deviation of the i-th key point, δ={0,1}, which is 0 when the key point is not included in the loss function calculation, and is 1 when the key point is included in the loss function calculation, v i , v j are the visibility flags of the i-th and j-th key points, E is an edge set for constructing a topological connection, μ ij is the direction vector of the joint composed of key points i and j, s ij is the cosine similarity of the direction vector and the true direction vector, μ ij (1-s ij ) is the joint dynamic weight, and mask() is a function for representing the visibility of key points, which is greater than 0 only when v i >0 and v j >0.

[0023] Further, a CHP filter is used for smoothing processing, and the relationship between the smoothed human key points and the original human key points satisfies:

[0024]

[0025] In the formula, Here are the predicted coordinates of key points at times t and t-1, where α is the smoothing coefficient and X is the value of the key point. t Let T be the original coordinates of the human body key points at time t, τ be the time constant of the filter, and T be the time constant of the filter. e The sampling period, i.e., the inference frequency of the attitude estimation algorithm, is f. c f is the cutoff frequency. min β is the minimum cutoff frequency, used to control the basic smoothing speed, and β is the rate coefficient, used to determine the degree of influence of the signal change rate on the cutoff frequency. The input signal is the rate of change of the key point coordinates.

[0026] Among them, f min The rate coefficient β is dynamically updated, and the specific steps include:

[0027] For a pair of human keypoints i and j that are physically connected, calculate the joint vector v of joint (i,j). ij =(x j -x i ,y j -y i ), where (x i ,y i Let (x) be the coordinates of key point i on the human body, and (x) be the coordinates of key point i on the j ,y j () represents the coordinates of key point j on the human body;

[0028] Based on the joint vector v of the joint (i,j) ij Calculate the length L of joint (i,j) ij and the angle θ with joint (p,q) ij The calculation expressions are as follows:

[0029]

[0030] Based on the target length, target angle, and smoothing coefficient α of the previous video frame, the target length of the t-th video frame is calculated. and target angle The update is performed, and the update expressions are as follows:

[0031]

[0032] In the formula, Let be the target length of the (t-1)th video frame. Let be the target angle of the (t-1)th video frame image;

[0033] based on the updated target length and the target angle calculate the length scaling factor r of the joint (i, j) ij and the rotation matrix R ij , respectively, are expressed as:

[0034]

[0035] where Δθ ij is the deviation value of the current angle θ ij and the target angle

[0036] based on the length scaling factor r ij and the rotation matrix R ij and the key point prediction confidence p i , p j , the joint (i, j) is scaled in length and rotated to update the coordinates of the human key points i and j to and calculate the moving distance of the current coordinates and the updated coordinates CHP filter adjustment coefficient where k is a set scaling factor, γ min , γ max are the minimum and maximum values of the CHP filter adjustment coefficient;

[0037] based on the CHP filter adjustment coefficient γ i , the f min and the rate coefficient β are dynamically updated:

[0038]

[0039] β i = γ i · β i

[0040] where subscript i is the i-th key point.

[0041] Further, a SG-RT-PicoDet network is used for target detection to obtain a detection frame of the worn equipment, the SG-RT-PicoDet network is improved by adding a re-parameterizable branch to the backbone network of a standard PicoDet network, the re-parameterizable branch is connected to the end of the Bottleneck module of each stage of Stage2 to Stage4 of the backbone network, and is parallel to the original depth separable convolution kernel,

[0042] ​Each reparameterizable branch includes a 1x1 convolution, a 3x3 depthwise separable convolution, each reparameterizable branch is processed to obtain a single equivalent convolution kernel, and multi-branch fusion is performed to obtain a fused equivalent convolution kernel and a bias, which are represented as follows:

[0043]

[0044] In the formula, W eq is the fused equivalent convolution kernel, b eq is the fused bias, N is, W i is the convolution kernel of each reparameterizable branch, γ i is a scaling factor, σ i is a standard deviation, μ i is a mean, and β i is a bias.

[0045] Further, the SG-RT-PicoDet network introduces a TAD-ATSS label assignment strategy during training, and the processing process of the TAD-ATSS label assignment strategy includes:

[0046] The sample assignment threshold is dynamically adjusted, the intersection over union of the preset box and the target box, the predicted box and the target box is calculated, and the joint intersection over union is further obtained, wherein the joint intersection over union is represented as:

[0047]

[0048] In the formula, C IoU (i,j) is the joint intersection over union, α and β are weight balance coefficients, and ɑ+β=1. In the initial stage of model training, the prediction accuracy of the model is low, so β is small to reduce the influence of the prediction box on label assignment. With the continuous adjustment of training, dynamic balance is achieved, A IoU (i,j) is the intersection over union of the preset box and the target box, P IoU (i,j) is the intersection over union of the predicted box and the target box, a i is the preset anchor box, g j is the target box, p i is the prediction;

[0049] The target box dynamic threshold is calculated based on the mean and standard deviation of the joint intersection over union, and the target box dynamic threshold is represented as:

[0050]

[0051] In the formula, is the target box dynamic threshold;

[0052] The classification and regression tasks are cooperatively optimized based on the joint intersection over union and the classification confidence si , the task consistency score is calculated and the task consistency score is taken as a calculation weight of the classification and regression loss to finally screen out the candidate box of the positive sample assigned as the target box, wherein ζ is a weight coefficient of the classification task, η is a weight coefficient of the regression task, and an expression of the classification and regression loss is:

[0053]

[0054] In the formula, L cls is a classification loss, N pos is a positive sample assigned as the target box, BCE is a binary cross entropy, is a normalized task consistency score, N neg is a negative sample assigned as the target box, γ is a focal loss focusing parameter, s j is a classification confidence of the candidate box, L reg is a regression loss, GIoU is a regression loss function, g i is the target box.

[0055] Further, the improved LCNet lightweight feature extraction network is improved by removing the intermediate layer ReLU activation layer in the standard LCNet lightweight feature extraction network, adding a bias term in the convolution layer, adjusting the convolution stride, and introducing a BN Neck module in the feature encoding stage,

[0056] The processing process of the improved LCNet lightweight feature extraction network includes:

[0057] The wearing equipment area image is input, features are extracted, and a fixed-length feature vector is obtained as a wearing feature.

[0058] Further, the improved LCNet lightweight feature extraction network combines the metric loss and the classification loss as a loss function, and the specific expression is:

[0059]

[0060] In the formula, L CE is a classification loss, N is a sample quantity, and C is a class quantity, is a real class label encoded by one-hot, indicates that the sample i belongs to the class c, otherwise not, is a probability that the sample i predicted by the network belongs to the class c, is a metric loss, a triple contains an anchor x a , a positive sample x p , and a negative sample x n, a is a predefined angle threshold, [z] + = max(0, z) is used to ensure the loss is non-negative.

[0061] Further, the C-HNSW graph retrieval algorithm is used for retrieval matching, and the specific steps include:

[0062] The graph construction stage:

[0063] The wearing feature library is hierarchically clustered according to the feature categories, a multi-level graph structure is constructed, and the category center vectors are used as top-level entry nodes.

[0064] The graph retrieval stage:

[0065] For the retrieval vector First, the distance from all category center vectors is calculated to obtain the most similar category center vector in terms of distance measurement, and the most similar category center vector is used as the initial entry node, wherein is a d-dimensional real set, d is the dimension of the wearing feature, and the initial entry node is represented as is the node set under category c * , is the category center point set, μ c is the category center vector, and v is the vector traversed in

[0066] According to the determined initial entry node, starting from the entry node of the highest layer L max , the nearest neighbor set is searched layer by layer downward, and the wearing feature most similar to the wearing feature of the wearing equipment region image to be identified is selected, and the specific steps include:

[0067] Each layer is based on a greedy search strategy, and the nearest node is selected as the next node If the distance of all nodes in the neighbor set of the current node is greater than the distance from the current node to the retrieval vector q, the search is stopped, and the next layer is entered, wherein, is the neighbor node set, v current is the current retrieved node, and d is the distance between nodes.

[0068] When searching the bottom layer, an extended search strategy is used to generate a candidate node set, and a category consistency constraint is added on the basis of the distance constraint, and the candidate nodes are scored, and the scoring result is:

[0069]

[0070] In the formula, scores i is the score of the candidate node, λ is a weight coefficient, ​represents the current node and the initial entry node v entry The category is consistent, and the value 0 represents inconsistency.

[0071] According to the scoring results, the k nodes with the lowest scores are selected from the candidate node set as the final matching results, that is, the most similar wearing features.

[0072] Further, the step of adopting the hysteresis comparison strategy for multi-frame integration optimization comprises:

[0073] According to the recent N video frame images, the proportion of qualified wearing video frame images is calculated, and compared with the set high and low thresholds. If the proportion of video frame images exceeds the high threshold, it is determined that the wearing is qualified, and if the proportion of video frame images is lower than the low threshold, it is determined that the wearing is unqualified.

[0074] Compared with the prior art, the present application has the following beneficial effects:

[0075] (1) The present application estimates human posture by Modified-YOLOv8-Pose network, extracts human key point coordinates, and optimizes stability by smoothing processing, improves detection accuracy and robustness. Based on human skeleton information, SG-RT-PicoDet network is used for accurate positioning of wearing equipment area, and TAD-ATSS label allocation strategy is used for positioning optimization, which improves the efficient feature extraction ability of target detection algorithm, so that the wearing equipment recognition can still run stably in complex industrial scenes (such as personnel intensive, uneven illumination, target occlusion, etc.).

[0076] (2) The present application constructs a wearing feature library based on small sample learning, uses LCNet light network for feature extraction, and constructs an efficient graph structure wearing equipment feature library, and combines C-HNSW graph retrieval algorithm for target region feature matching, so that the present application can complete efficient modeling and rapid migration under the condition of small sample data, solving the problem of dependence on large-scale labeled data in traditional methods. This method supports wearing recognition under the condition of open categories, and can be applied to different operation scenes and various wearing equipment categories.

[0077] (3) The present application uses SG-RT-PicoDet network for positioning of wearing equipment area, and combines TAD-ATSS label allocation strategy for regional positioning optimization. This detection process can automatically adapt to different types of wearing equipment, reducing the false detection problem caused by equipment form difference.

[0078] (4) The HG Block module is introduced in the Modified-YOLOv8-Pose network of the application, which improves the detection ability of the network for small targets (such as hand and foot key points), and combines the C-BRAM attention mechanism, a light channel attention module to enhance the feature expression between channels, and a spatial attention double-layer routing module to realize global spatial dynamic modeling, thereby enhancing the feature extraction ability of the network for key areas and reducing the complex background interference in the industrial scene.

[0079] (5) In the detection stage, the application performs target region feature retrieval and time sequence fusion. The feature vector is extracted for the target region, and the C-HNSW graph retrieval algorithm based on class center optimization is used for efficient retrieval matching, so that high-precision matching results can be obtained under small sample conditions. In order to improve the stability of detection, the application adopts a hysteresis comparison strategy to perform time sequence fusion optimization on multiple frame detection results to obtain the current recognition result.

[0080] (6) The application supports rapid adaptation to different industrial scenes, and model migration can be completed by updating the wearing feature library with a small amount of samples of new targets, which is suitable for various industries (such as manufacturing, chemical industry, construction, etc.), and has wide application prospects and market value. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 is a method flowchart of the application;

[0082] Figure 2 is a modified Modified-YOLOv8-Pose network structure diagram of the application;

[0083] Figure 3 is an SG-RT-PicoDet network structure diagram of the application;

[0084] Figure 4 is an improved LCNet light feature extraction network structure diagram of the application;

[0085] Figure 5 is a software system hierarchy diagram of the application. DETAILED DESCRIPTION

[0086] The application will be described in detail below in combination with the drawings and specific embodiments. The embodiments are implemented on the premise of the technical solutions of the application, and detailed implementation methods and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0087] The embodiment provides a lightweight personnel wearing equipment intelligent recognition method under a small sample condition. The method realizes efficient recognition of wearing equipment, rapid adaptation to different industrial environments under the condition of limited sample data, meets real-time operation requirements in a low-computing-resource environment, reduces the occurrence of dangerous accidents, and improves the safety and intelligent level of industrial production. Specifically, as shown in Figure 1 the method includes the following steps:

[0088] S1. Obtain video frame images in an industrial production environment through a network camera and transmit the video frame images to a back-end server or an edge computing device.

[0089] The embodiment transmits video streams through an RTSP protocol and samples images at a fixed frame rate to ensure the real-time performance and stability of subsequent algorithms.

[0090] S2. Estimate the human body posture in the image by using the proposed Modified-YOLOv8-Pose algorithm, obtain the human body key point coordinates, and combine the designed CHP filter to optimize the stability of the key points, thereby improving the detection accuracy and robustness.

[0091] In this step, the Modified-YOLOv8-Pose algorithm is used to estimate the human body posture, and it is judged whether there is a human body target in the current picture according to the human body posture estimation result. If no worker is detected, the next frame is directly processed; if a worker is detected, the key point screening and region detection in the next step are continued.

[0092] During the human posture estimation process, the coordinates of 17 key points are extracted for each human body, including the positions of the head, neck, shoulders, elbows, wrists, hands, hips, knees, ankles, feet, spine, chest, waist, eyes, ears, nose, and buttocks. In some deep modules of the YOLOv8-Pose backbone network, and in parts that affect the detection of small and medium-sized targets, a computationally efficient HG Block (Hourglass Block) feature enhancement module is introduced to improve the model's ability to detect small targets (such as key points of hands and feet). A C-BRAM attention module (Cross-Boundary Attention Module) combining two-layer routing spatial attention and lightweight channel attention is designed. This lightweight channel attention module enhances the feature representation between channels, and the two-layer routing module of spatial attention achieves global spatial dynamic modeling, thereby enhancing the network's ability to extract features from key regions, reducing interference from complex backgrounds in industrial scenarios, improving the detection accuracy of key regions, and reducing the instability of pose recognition caused by inter-frame jitter during video stream inference. Furthermore, a pose loss function JD Loss for joint directions is designed, combining the Euclidean distance loss between predicted points and ground truth values ​​with the direction loss between joints (the line connecting two key points) to strengthen the modeling of human topology. The expression for this pose loss function JD Loss is:

[0093]

[0094] In the formula, L JD Let L be the loss value. distance The loss is based on Euclidean distance, with λ as an adjustment coefficient used to balance the weights of distance and direction losses. Experiments show that setting it to 1 achieves the best balance between accuracy and stability. direction For joint-oriented loss, N is the total number of keypoints in a single human body, which is set to 17 in this paper. i Let A be the Euclidean distance between the predicted and actual locations of the i-th keypoint. j σ represents the area of ​​the human body where the keypoints are located, used to normalize the pose loss of different human bodies and eliminate the influence of different scales. i v is a fixed standard deviation of the i-th keypoint, used to reflect the importance of different keypoints. Keypoints with higher accuracy requirements are assigned smaller values. i It is the visibility flag of the i-th keypoint, v i =0 means the keypoint is invisible; in this case, δ=0, and the keypoint is not included in the loss function calculation. i When μ = 1, δ = 1, and this key point is included in the calculation. E is the edge set for constructing topological connections, and the directed edges it contains are shown in the table below. ij Let s be the direction vector of the joint formed by key points i and j.ij is the cosine similarity of the direction vector and the real direction vector, thereby measuring the matching degree of the direction between the human joints, μ ij (1-s ij is the joint dynamic weight, and mask() is a function for representing the visibility of the key points, and mask>0 only when v i >0 and v j >0.

[0095] The improved Modified-YOLOv8-Pose network structure diagram is as shown in Figure 2 The network structure includes a Backbone part, a Neck part and a Head part, the Backbone part includes a first CBS module, a second CBS module, a first C2f module, a third CBS module, a first HG Block module, a fourth CBS module, a second C2f module, a C-BRAM module and a SPPF module connected in sequence, the Neck part includes a first up-sampling module, a first connection module, a third C2f module, a second up-sampling module, a second connection module, a second HG Block module, a fifth CBS module, a third connection module, a third HG Block module, a sixth CBS module, a fourth connection module and a fourth HG Block module connected in sequence with the SPPF module, the first connection module is further connected with the first HG Block module, the third C2f module is further connected with the third connection module, the second connection module is further connected with the first C2f module, the fourth connection module is further connected with the SPPF module, and the Head part includes three Pose modules, each Pose module is connected with a corresponding second HG Block module, third HG Block module and fourth HG Block module.

[0096] The first HG Block module, the second HG Block module, the third HG Block module and the fourth HG Block module each include a plurality of continuous convolution layers for fusing multi-scale information and supporting residual connection. The HG Block module is composed of 6 continuous 3*3 convolution layers, fuses multi-scale information, adopts a compression and excitation mechanism, reduces the calculation overhead, enhances the feature expression capability, and supports residual connection to reduce the gradient vanishing problem in the deep network.

[0097] The C-BRAM module includes a lightweight channel attention sub-module and a spatial attention double-layer routing sub-module, and the processing process of the lightweight channel attention sub-module includes:

[0098] Based on the feature map processed by the front C2F module of the C-BRAM module, the global information of each channel is extracted to obtain a channel description vector.

[0099] Based on the channel description vector, the initial inter-channel weight is calculated, and the final inter-channel weight is obtained by adopting Sigmoid normalization, so as to enhance the feature expression between channels;

[0100] The processing procedure of the spatial attention double-layer routing sub-module comprises:

[0101] The feature map subjected to the channel attention weighting is regionally divided, Query, Key and Value values are respectively generated through linear projection, and a region-level routing is constructed;

[0102] Based on the region-level routing, the region affinity is calculated, coarse-grained screening is performed, and spatial weight is obtained through token-level fine-grained attention calculation, so as to guarantee the feature expression capability of the network.

[0103] This step also designs a CHP filter to optimize the key point coordinates and improve the stability of the detection results. The CHP filter is based on the human joint topology information (shown in Table 1 below), combined with joint length constraints and angle constraints, to realize the stability optimization of the key point coordinates in the video stream inference process.

[0104] Table 1 Human key point topology structure

[0105]

[0106] The specific design principle of the CHP filter is as follows. The One Euro filter is an adaptive low-pass filter based on the cutoff frequency f c The smoothing coefficient a of the filter is adaptively adjusted, so as to maintain the balance between the filtering strength and the delay. The relationship between the original value and the smoothed value satisfies wherein, is the predicted value of the key point coordinates at t and t-1, X t is the original human key point coordinate at t, the smoothing coefficient The rate coefficient b is used to determine the influence degree of the signal change rate on the cutoff frequency, tau is the time constant of the filter, T e is the sampling period, that is, the inference frequency of the pose estimation algorithm, f c is the cutoff frequency, f min is the minimum cutoff frequency, used to control the basic smoothing speed, is the input signal, that is, the change rate of the key point coordinate value.

[0107] When establishing the joint length constraint and the angle constraint, for a pair of human key points i and j having a physical connection relationship, first, the joint vector v ij of the joint (i, j) is calculated. j x i y j-y i ); secondly, calculate its length angle with joint (p, q) where (x i ,y i ) is the coordinate of human key point i, (x j ,y j ) is the coordinate of human key point j.

[0108] Next, the target value of the last frame and the smoothing coefficient a are combined to update the target length and the target angle of the t-th frame. The updated Thus, the length scaling factor of key point (i, j) is calculated Rotation matrix where Δθ ij is the deviation value of the current angle θ ij and the target angle

[0109] Finally, based on r ij , R ij and the key point prediction confidence p i , p j , the joint is scaled in length and rotated, and then the key point i, j coordinate position is updated to The moving distance of the current value and the updated value is obtained and the filter adjustment coefficient According to this coefficient, the CHP filter algorithm parameters are dynamically adjusted: β i = γ i · β i , where k is the set scaling factor, which is set to 1.5 in this paper, and γ min , γ max are the minimum and maximum values of the CHP filter adjustment coefficient.

[0110] S3, based on human skeleton information, the SG-RT-PicoDet algorithm proposed is used for wearable equipment region positioning, and the designed TAD-ATSS label allocation strategy is combined for target detection optimization, to improve the detection accuracy and the accuracy of the detection region.

[0111] As Figure 3 ​As shown, the SG-RT-PicoDet network improves the standard PicoDet network in this step, including adding a re-parameterizable branch to the standard PicoDet backbone network to enhance the network's ability to extract complex features while maintaining inference speed; designing a TAD-ATSS label assignment strategy to optimize the distribution of positive samples in the target detection task and improve detection accuracy in complex environments; and fusing the candidate box optimization method of the skeleton heat map to guide the model to focus on the surrounding area of the human body and improve the positioning accuracy of the wearable equipment area. The SG-RT-PicoDet algorithm detects the subject target of the input image, and finally outputs the filtered high-confidence wearable equipment detection box as the input of the subsequent feature extraction and recognition module.

[0112] Among them, the re-parameterizable branch, at the end of the Bottleneck module in each stage of Stage2 to Stage4 of the network, respectively adds a 1x1 convolution, a 3x3 depth separable convolution, which is parallel to the original 5x5 depth separable convolution kernel in the network, thereby improving the network feature extraction capability. During model inference, branch fusion is performed, and the convolution kernel with smaller kernel size is expanded to the same kernel size as the main branch through zero padding. For each branch, the convolution kernel W i The corresponding batch normalization parameters are mean μ i , standard deviation σ i , scaling factor γ i and offset β i , calculate the single equivalent convolution kernel and bias of the fused multiple branches, so as to maintain the inference speed.

[0113] Among them, the TAD-ATSS label assignment strategy is based on the task alignment learning idea, and dynamically and adaptively assigns positive samples for model training. First, the sample distribution threshold is dynamically adjusted, the intersection over union of the preset box and the target box, the predicted box and the target box is calculated, and then the joint intersection over union Based on the mean and standard deviation, the target box dynamic threshold is calculated

[0114] Next, the classification and regression tasks are optimized in coordination, based on the joint intersection over union and the classification confidence s i calculated in the model inference process, the task consistency score t is calculated, and the task consistency score is taken as the calculation weight of the classification and regression loss. Finally, the task consistency score t i is normalized to as the calculation weight of the classification loss L cls and the regression loss L reg , further guiding the model to better align the classification and regression tasks. where BCE is binary cross entropy, g is focal loss focusing parameter, N pos is positive sample assigned as target box, N neg is negative sample assigned as target box, GIoU is regression loss function.

[0115] S4, based on LCNet lightweight classification network, a small amount of samples of the wearing equipment are extracted features, and a wearing feature library is constructed.

[0116] As Figure 4 shown, in this step, the LCNet lightweight classification network uses a depth separable convolution structure, and the network architecture is adaptively optimized, including removing the intermediate layer ReLU activation layer, adding a bias term in the convolution layer to improve the fitting ability to small sample distribution, and adjusting the convolution stride to increase the feature extraction ability of fine-grained space; in the feature encoding stage, the BN Neck module is introduced to normalize the output feature vector, and by optimizing the classification loss and the metric loss, the consistency of the small sample feature distribution is improved. The whole network execution process is: after the input image is extracted features by the convolution layer and the depth convolution module, a fixed-length feature vector is output, which is the basis for constructing the wearing feature library and subsequent image retrieval matching.

[0117] In order to extract a feature vector with good intra-class compactness and inter-class separability, the metric loss and the classification loss are combined as the loss function in the embodiment of the application. For a classification task containing C classes and N samples, the classification loss is defined as where, for sample i, is the real class label encoded by one-hot, indicates that the sample i belongs to class c, otherwise it does not belong to; is the probability of the model predicting that the sample i belongs to class c. The metric loss where the triple contains anchor x a , positive sample x p and negative sample x n ; a is a pre-defined angle threshold, [z] + =max(0,z) is used to ensure that the loss is non-negative.

[0118] In Batch Normalization (BN) Neck, a batch normalization layer is added after the global pooling layer, transforming the features into two different spaces, one for optimizing classification loss and the other for optimizing metric learning loss. The feature vector obtained after the global pooling layer retains the original distribution, making it suitable for metric learning that relies on true geometric relationships, while the normalized features are more compact in space, making them suitable for class classification. Through this spatial separation, BN Neck avoids the optimization conflict caused by classification loss and metric learning loss acting simultaneously on the same feature vector.

[0119] S5. Design the C-HNSW graph retrieval algorithm to achieve target region feature retrieval and matching. Use the hysteresis comparison strategy to optimize the time series fusion of multi-frame detection results, reduce single-frame false detections, and improve recognition stability.

[0120] This step first obtains the wearing features of the wearable equipment region image to be identified based on the steps described above, and then uses the C-HNSW graph retrieval algorithm to search and match them with the wearable feature database. This C-HNSW graph retrieval algorithm optimizes the HNSW algorithm based on category centers, constructing an efficient multi-level graph structure to improve the speed and accuracy of small-sample feature matching.

[0121] Specifically, the C-HNSW algorithm optimizes the HNSW algorithm from the perspectives of graph construction and updating, and retrieval, providing a more efficient graph network structure for vector retrieval and improving retrieval speed and accuracy. In the graph construction stage, samples in the wearable feature library are hierarchically clustered according to categories to construct a multi-level graph structure, with the center vector of each category as the top-level entry node. In the graph retrieval stage, the most similar category center node is quickly located using a distance metric. Then, starting from this center node, the nearest neighbor set is traversed layer by layer downwards, ultimately selecting the sample most similar to the feature vector of the target region to be detected. For retrieval vectors... First, calculate the distance to all category centers, and then select the initial entry point v based on the nearest category center. entry , in, It is category c * The set of nodes below, After determining the initial entry point, start from the highest level L. max Starting from the entry point, the search proceeds layer by layer downwards; at each layer, a greedy search strategy is used to select the nearest node as the next node. If the distance to all nodes in the current node's neighbor set is greater than the distance from the current node to the retrieval vector q, then the search stops and proceeds to the next level. During the lowest level search, an extended search strategy is used to generate a candidate set, adding a class consistency constraint on top of the distance constraint, and scoring the candidate nodes. in, This indicates the relationship between the current node and the initial entry point v. entryCategory consistent, value is 0, which means inconsistent. From which the lowest score of k nodes are selected as the final matching result. This process significantly improves the retrieval efficiency and accuracy of the graph structure under small sample conditions, supports O(log N) level approximate nearest neighbor query capability, and is suitable for real-time demand scenarios in wearable identification.

[0122] The above hysteresis comparison strategy optimizes the detection result by the following way: based on the last N frames of image, the proportion of qualified image of wearing is calculated, and compared with the set high and low threshold value; when the qualified frame proportion exceeds the high threshold value, the wearing equipment is judged to be qualified; when the qualified frame proportion is lower than the low threshold value, the wearing equipment is judged to be unqualified; using this strategy to integrate time sequence information, improve the stability of detection result, reduce the misjudgment caused by single frame misdetection.

[0123] S6, deploy the trained model to the domestic AX650N chip for edge computing, realize real-time detection, violation alarm, monitoring picture preview and statistical visualization function.

[0124] This embodiment deploys the trained model to high-performance low-power chip AX650N, and uses model pruning and quantization technology to reduce the consumption of computing resources, so that the system can run efficiently on edge devices and meet the real-time requirements of industrial sites.

[0125] Finally, the system displays the detection result through the real-time visualization interface, and supports violation alarm, monitoring picture preview and statistical visualization function. For personnel who do not wear equipment correctly, the system can automatically issue a warning, notify the management personnel to handle through voice reminder and warning light alarm mode.

[0126] The system as shown in Figure 5 , comprising:

[0127] 1) Perception layer: responsible for the collection and preliminary processing of real-time monitoring video stream in the field. Use RTSP protocol for data transmission, provide high-quality video frame image for business layer and application layer.

[0128] 2) Data layer: used for storing and managing structured and unstructured data generated during system operation. Structured data includes camera information, abnormal log, etc.; unstructured data includes abnormal pictures. The data layer provides data support for the visualization data display of the application layer.

[0129] 3) Business layer: integrate service scheduling, identification algorithm module, and provide corresponding API interface. According to the detection result of the identification algorithm, the business layer generates unqualified wearing information, and transmits it to the data layer through the standardized interface, to provide support for real-time display and user interaction of the front end.

[0130] 4) Application layer: Provide an intuitive interaction interface for users, support real-time monitoring picture display, abnormal behavior statistics and trend analysis, view and manage network camera information, etc.

[0131] The test results on the test data set in the actual scene show that the recognition accuracy of the wearable equipment reaches 98.6%, and stable performance can be maintained in complex scenes (light changes, personnel concentration, and equipment diversity); the inference frame rate reaches 30FPS, meeting the real-time monitoring requirements; only a small amount of sample data (5-10 images per type of equipment) is needed to complete model migration, meeting the needs of rapid deployment in actual applications.

[0132] The system can realize efficient edge computing deployment, deploy the trained model to the domestic AX650N chip after optimization, and use model pruning and quantization technology to reduce computing resource consumption, so that the system can efficiently run in embedded devices or low-power environments, meeting the real-time requirements of industrial sites. At the same time, the system supports monitoring picture preview, violation alarm, and statistical visualization functions, improving the intelligent level of industrial production.

[0133] The system can accurately, quickly and conveniently realize real-time automated detection of factory workers' wearing equipment, reduce manual inspection costs, avoid the instability of manual detection, and improve the efficiency of production safety management. Through real-time alarm and visual feedback, it helps managers quickly find non-compliant wearing behavior, reduces the incidence of safety accidents, and ensures the safety of industrial production.

[0134] The above functions, if realized in the form of software function units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or the parts that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes instructions to make a computer device (which can be a personal computer, server, or network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various program code storage media.

[0135] Those skilled in the art will appreciate that embodiments of the present application can be readily used as software, hardware, or a combination of software and hardware. In a software embodiment, the methods can be tangibly embodied in a machine-readable storage medium having stored thereon instructions that can be used to program a computer to perform any of the methods. The software implementation can be initialized by loading and executing a set of instructions arranged to perform one of the methods into the computer's memory. Alternatively, hard-wired circuitry can be used in place of, or in combination with, software instructions. Thus, the

[0136] The present application is described in reference to the flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing one or more functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0137] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing one or more functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in one or more of the flowchart illustrations and / or block diagrams. Figure 1 means for performing one or more functions specified in one or more of the flowchart illustrations and / or block diagrams.

[0139] While preferred embodiments of the application have been described, modifications and variations can be apparent to those skilled in the art once aware of the general underlying concepts. Accordingly, the appended claims are intended to encompass all modifications and variations as falling within the scope of the application.

[0140] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A lightweight personnel equipment intelligent recognition method under small sample conditions, characterized by, The method comprises the following steps: Collecting video frame images of an industrial production environment; Performing human pose estimation on the video frame images by using a Modified-YOLOv8-Pose network, judging whether there is a person, if there is a person, extracting human key points and performing smoothing processing to form human skeleton information, if there is no person, re-collecting video frame images; Based on the human skeleton information, performing target detection on the wearing equipment region of the human body, and fusing with the set key wearing region cropping to be associated to the corresponding human body target to obtain a wearing equipment region image; Performing wearing feature extraction on the wearing equipment region image by using an improved LCNet lightweight feature extraction network, and constructing a wearing feature library; Obtaining the wearing feature of the wearing equipment region image to be recognized, performing retrieval matching on the wearing feature library, and simultaneously performing multi-frame integration optimization by using a hysteresis comparison strategy to obtain a wearing equipment recognition result.

2. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The Modified-YOLOv8-Pose network is improved by introducing an HG Block module, a C-BRAM module and a pose loss function in the YOLOv8-Pose network, and comprises a Backbone part, a Neck part and a Head part, the Backbone part comprises a first CBS module, a second CBS module, a first C2f module, a third CBS module, a first HG Block module, a fourth CBS module, a second C2f module, a C-BRAM module and a SPPF module connected in sequence, the Neck part comprises a first up-sampling module, a first connection module, a third C2f module, a second up-sampling module, a second connection module, a second HG Block module, a fifth CBS module, a third connection module, a third HG Block module, a sixth CBS module, a fourth connection module and a fourth HG Block module connected in sequence with the SPPF module, the first connection module is further connected with the first HG Block module, the third C2f module is further connected with the third connection module, the second connection module is further connected with the first C2f module, the fourth connection module is further connected with the SPPF module, and the Head part comprises three Pose modules, each Pose module is connected with a corresponding second HG Block module, third HG Block module and fourth HG Block module, Wherein, the first HG Block module, second HG Block module, third HG Block module and fourth HG Block module each comprise a plurality of continuous convolution layers for fusing multi-scale information and supporting residual connection; The C-BRAM module comprises a lightweight channel attention submodule and a spatial attention double-layer routing submodule, and the processing process of the lightweight channel attention submodule comprises: Based on the feature map obtained by the preposed C2F module of the C-BRAM module, extracting the global information of each channel to obtain a channel description vector; Based on the channel description vector, an initial inter-channel weight is calculated, and a final inter-channel weight is obtained by Sigmoid normalization; The processing process of the spatial attention double-layer routing submodule includes: The feature map subjected to the channel attention weighting is divided into regions, Query, Key, and Value values are respectively generated through linear projection, and a regional-level routing is constructed; Based on the regional-level routing, a regional affinity is calculated, coarse-grained screening is performed, and a spatial weight is obtained through token-level fine-grained attention calculation.

3. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 2, characterized in that, The expression of the posture loss function is: wherein L JD is a loss value, L distance is a loss based on Euclidean distance, λ is an adjustment coefficient for balancing the weight of distance loss and direction loss, L direction is a loss based on joint direction, N is the total number of key points of a single human body, d i is the Euclidean distance between the predicted and real positions of the i-th key point, A j is the area of the region of the human body in which the key point is located, σ i is the fixed standard deviation of the i-th key point, δ = {0, 1}, δ = 0 indicates that the key point is not included in the loss function calculation, and δ = 1 indicates that the key point is included in the loss function calculation, v i , v j is the visibility flag of the i-th and j-th key points, E is an edge set for constructing a topological connection, μ ij is the direction vector of the joint composed of the i-th and j-th key points, s ij is the cosine similarity of the direction vector and the real direction vector, μ ij (1-s ij ) is a joint dynamic weight, and mask() is a function for representing the visibility of key points, mask>0 only when v i >0 and v j >0.

4. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The relationship between the smoothed human key points and the original human key points satisfies: In the formula, is the key point coordinate prediction value at time t and t-1, α is the smoothing coefficient, X t is the original human body key point coordinate at time t, τ is the time constant of the filter, T e is the sampling period, that is, the inference frequency of the pose estimation algorithm, f c is the cut-off frequency, f min is the minimum cut-off frequency, used to control the basic smoothing speed, β is the rate coefficient, used to determine the influence degree of the signal change rate on the cut-off frequency, is the input signal, that is, the change rate of the key point coordinate value; where f min and the rate coefficient β is dynamically updated, with the specific steps including: For a pair of human key points i and j with a physical connection relationship, calculate the joint vector v of joint (i, j) ij = (x j -x i , y j -y i ), wherein (x i , y i ) is the coordinate of human key point i, and (x j , y j ) is the coordinate of human key point j; a joint vector v based on the joint (i,j) ij a length L of the joint (i,j) ij and an angle Θ with the joint (p,q) ij where the calculation expressions are respectively: The target length and the target angle of the tth video frame image are updated according to the target length, the target angle and the smoothing coefficient a of the previous video frame image, and the update expressions are respectively: L t = L t-1 + a (L t-1 - L t) and θ t = θ t-1 + a (θ t-1 - θ t) In the formula, is a target length of the t-1th video frame image, is a target angle of the t-1th video frame image; based on the updated target length and target angle calculating a length scaling factor r for joint (i,j) ij and rotation matrix R ij are represented as: In the formula, Δθ ij is a deviation value of a current angle θ ij from a target angle ​ scaling coefficient r based on the length ij and a rotation matrix R ij and a key point prediction confidence p i , p j , the joint (i, j) is length-scaled and rotated to update the coordinates of the human body key points i and j as and calculate the moving distance of the current coordinates and the updated coordinates CHP filter adjustment coefficient wherein k is a set scaling factor, γ min , γ max is the minimum and maximum value of the CHP filter adjustment coefficient; adjusting coefficient γ of the CHP filter based on the CH i , the f min and rate coefficient β are dynamically updated: β i = γ i • β i In the formula, subscript i is the i th key point.

5. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The SG-RT-PicoDet network is used for target detection to obtain a wearable equipment detection frame. The SG-RT-PicoDet network is improved by adding a reparameterizable branch to the backbone network of the standard PicoDet network. The reparameterizable branch is connected to the end of the Bottleneck module at each stage of Stage2 to Stage4 of the backbone network and is parallel to the original depth separable convolution kernel, Each reparameterizable branch includes a 1x1 convolution and a 3x3 depth separable convolution. After processing, each reparameterizable branch obtains a single equivalent convolution kernel, and multi-branch fusion is performed to obtain a fused equivalent convolution kernel and a bias, which are represented as follows: In the formula, W eq is the equivalent convolution kernel after fusion, b eq is the bias after fusion, N is, W i is the convolution kernel of each reparameterizable branch, γ i is the scaling factor, σ i is the standard deviation, μ i is the mean, β i is the offset.

6. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 5, characterized in that, The SG-RT-PicoDet network introduces a TAD-ATSS label assignment strategy during the training process. The processing process of the TAD-ATSS label assignment strategy includes: The sample allocation threshold is dynamically adjusted, the intersection over union of the preset frame and the target frame, the predicted frame and the target frame is calculated, and the joint intersection over union is further obtained. The joint intersection over union is represented as: In the formula, C IoU (i,j) is the joint intersection-over-union ratio, a and β are weight balance coefficients, and a+β=1. In the initial stage of model training, the prediction accuracy of the model is low, so the value of β is small to reduce the influence of the prediction box on the label assignment. With the continuous adjustment during the training, dynamic balance is achieved. IoU (i,j) is the intersection-over-union ratio of the preset box and the target box, P IoU (i,j) is the intersection-over-union ratio of the prediction box and the target box, a i is the preset anchor box, g j is the target box, p i is the prediction; The target frame dynamic threshold is calculated based on the mean and standard deviation of the joint intersection over union. The target frame dynamic threshold is represented as: In the formula, TargetBoxDynamicThreshold is the target box dynamic threshold. Co-optimizing classification and regression tasks based on joint intersection over union and classification confidence s i , computing task consistency score and taking the task consistency score as the calculation weight of the classification and regression loss, finally screening out the candidate box of the positive sample assigned as the target box, wherein, is the weight coefficient of the classification task, η is the weight coefficient of the regression task, and the expression of the classification and regression loss is: wherein L cls is the classification loss, N pos is a positive sample assigned as a target box, BCE is a binary cross entropy, is the normalized task consistency score, N neg is a negative sample assigned as a target box, γ is a focal loss focusing parameter, s j is the classification confidence of the candidate box, L reg is the regression loss, GIoU is a regression loss function, g i is the target box.

7. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The improved LCNet lightweight feature extraction network is improved by removing the intermediate layer ReLU activation layer in the standard LCNet lightweight feature extraction network, adding a bias term in the convolution layer, adjusting the convolution stride, and introducing a BN Neck module in the feature encoding stage, The processing process of the improved LCNet lightweight feature extraction network includes: The wearable equipment region image is input, features are extracted, and a fixed-length feature vector is obtained as a wearable feature.

8. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The improved LCNet lightweight feature extraction network combines the metric loss and the classification loss as a loss function, and the specific expression is: where L CE is the classification loss, N is the number of samples, C is the number of classes, is the true class label with one-hot encoding, denotes that the sample i belongs to class c, otherwise not, is the probability that the network predicts that the sample i belongs to class c, is the metric loss, triplet contains the anchor x a , positive sample x p and negative sample x n , a is a predefined angle threshold, [z] + = max(0, z) is used to ensure the loss non-negative.

9. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The C-HNSW graph retrieval algorithm is used for retrieval matching, and the specific steps include: Graph construction stage: The wearable feature library is hierarchically clustered according to the feature categories, a multi-level graph structure is constructed, and the center vectors of each category are used as top-level entry nodes; Graph retrieval stage: For the retrieval vector First, the distance to all category center vectors is calculated, the most similar category center vector is obtained according to the distance metric, and the most similar category center vector is taken as the initial entry node, wherein is a d-dimensional real set, d is the dimension of the wearing feature, and the initial entry node is represented as is the node set under category c * , is the category center point set, μ c is the category center vector, and v is the vector traversed in . According to the determined initial entry node, starting from the highest layer L max , search the neighbor set layer by layer downwards, select the wearing feature most similar to the wearing feature of the wearing equipment region image to be identified, and the specific steps include: Each layer selects the nearest node as the next node based on a greedy search strategy If the distance between the current node and the retrieval vector q is greater than the distance between all nodes in the neighbor set of the current node, the search is stopped and the next layer is entered, wherein, is the neighbor node set, v current is the current retrieved node, and d is the distance between nodes. In the bottom layer search, an extended search strategy is used to generate a candidate node set, a category consistency constraint is added on the basis of the distance constraint, and the candidate nodes are scored, and the scoring result is: In the formula, scores i The scores of the candidate nodes, λ is a weight coefficient, The current node and the initial entry node v entry The category is consistent, and the value of 0 indicates inconsistency. According to the scoring result, the k nodes with the lowest scores are selected from the candidate node set as the final matching result, that is, the most similar wearing feature.

10. The lightweight personnel equipment intelligent identification method under small sample conditions according to claim 1, characterized in that, The step of adopting the hysteresis comparison strategy for multi-frame integration optimization includes: According to the recent N video frame images, the proportion of qualified video frame images is calculated, and compared with the set high and low thresholds, if the video frame image proportion exceeds the high threshold, it is determined that the wearing is qualified, if the video frame image proportion is lower than the low threshold, it is determined that the wearing is unqualified.

Images