A lightweight personnel wearing equipment intelligent recognition method under a small sample condition

Abstract

The present application relates to a kind of small sample conditions under lightweight personnel wearing equipment intelligent identification method, comprising: the video frame image of industrial production environment is collected;Video frame image is carried out human posture estimation, judge whether there is a person, if there is a person, then extract human key point, and carry out smoothing processing, form human skeleton information, if there is no person, then video frame image is re-collected;The wearing equipment region of human body is carried out target detection, and with the cutting of set key wearing area is fused, associated to corresponding human target, obtain wearing equipment region image;Wearing equipment region image is carried out wearing feature extraction, and constructs wearing feature library;The wearing feature of wearing equipment region image to be identified is acquired, wearing feature library is searched and matched, while adopting hysteresis comparison strategy carries out multi-frame integration optimization, obtains the wearing identification conclusion of wearing equipment.Compared with prior art, the present application has the advantages of real-time lightweight, result accurate and the like.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a lightweight intelligent recognition method for wearable equipment under small sample conditions. Background Technology

[0002] In modern industrial production, workplace safety remains a core issue of great concern to both enterprises and society. As industrial production environments become increasingly complex and equipment scales up, the potential safety risks during production processes increase significantly. Therefore, the proper wearing of safety protective equipment by personnel is crucial for reducing the accident rate. Traditional monitoring methods primarily rely on manual inspections, which suffer from high labor intensity, high error rates, poor real-time performance, and high monitoring costs. In recent years, deep learning-based intelligent monitoring technology has been increasingly applied to workplace safety management, automatically identifying personnel's protective gear and improving the automation level of safety supervision.

[0003] However, existing intelligent recognition methods still have many problems: First, existing methods rely on large-scale labeled data. Current deep learning methods typically require a large amount of manually labeled training data, but in industrial environments, data collection is difficult, making it hard to build large-scale datasets. Second, existing methods struggle to adapt quickly to different industrial scenarios. In different production environments, the types, colors, and shapes of equipment worn by workers vary significantly. Traditional models often need to be retrained when migrating between scenarios, which is time-consuming and labor-intensive. Finally, existing methods consume a lot of computing resources, mainly relying on high-performance GPU servers. However, in industrial environments, computing resources are limited, making efficient deployment on edge devices difficult, affecting the real-time performance and availability of detection. Summary of the Invention

[0004] The purpose of this invention is to provide a lightweight intelligent identification method for wearable equipment for stable operation in complex industrial scenarios under small sample conditions.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for intelligent recognition of lightweight wearable equipment under small sample conditions includes the following steps:

[0007] Capture video frame images of the industrial production environment;

[0008] The Modified-YOLOv8-Pose network is used to estimate human pose in the video frame images to determine whether there is a person. If there is a person, the human key points are extracted and smoothed to form human skeleton information. If there is no person, the video frame images are re-acquired.

[0009] Based on the human skeleton information, target detection is performed on the area of ​​the human body where the equipment is worn, and the data is fused with the set key wearing area cropped and associated with the corresponding human target to obtain an image of the area where the equipment is worn.

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

[0011] The wear features of the wearable equipment region image to be identified are obtained, and the wearable feature library is searched and matched. At the same time, a hysteresis comparison strategy is used to integrate and optimize multiple frames to obtain the wear recognition result of the wearable equipment.

[0012] Furthermore, the Modified-YOLOv8-Pose network is improved by introducing an HGBlock module, a C-BRAM module, and a pose loss function into the YOLOv8-Pose network. The Modified-YOLOv8-Pose network includes a Backbone section, a Neck section, and a Head section. The Backbone section 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 an SPPF module connected in sequence. The Neck section includes a first upsampling module, a first connection module, a third C2f module, a second upsampling 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 to the SPPF module. The first connection module is also connected to the first HG Block module. The third C2f module is connected to the first C2f module, the second connection module is connected to the first C2f module, and the fourth connection module is connected to the SPPF module. The head section includes three pose modules, each of which is connected to the 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 HGBlock module all include multiple consecutive convolutional layers to fuse multi-scale information and support residual connections.

[0014] The C-BRAM module includes a lightweight channel attention submodule and a spatial attention two-layer routing submodule. The processing procedure of the lightweight channel attention submodule includes:

[0015] The feature map obtained by the C-BRAM module is processed by the C2F module in front, and the global information of each channel is extracted to obtain the channel description vector.

[0016] Based on the channel description vector, the initial inter-channel weights are calculated, and the final inter-channel weights are obtained by Sigmoid normalization.

[0017] The processing procedure of the spatial attention two-layer routing submodule includes:

[0018] The feature map after channel attention weighting is divided into regions, and Query, Key, and Value values ​​are generated by linear projection, and region-level routes are constructed.

[0019] Based on the region-level routing, region affinity is calculated for coarse-grained filtering, and spatial weights are obtained through fine-grained attention calculation at the token level.

[0020] Furthermore, the expression for the attitude loss function is:

[0021]

[0022] In the formula, L JD Let L be the loss value. distance The loss is based on Euclidean distance, where λ is an adjustment coefficient used to balance the weights of distance loss and direction loss, and L... direction For joint-oriented loss, N is the total number of keypoints in a single human body, and d 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 key point is located. i Let δ be the fixed standard deviation of the i-th keypoint, where δ = {0, 1}. A value of 0 indicates that the keypoint is not included in the loss function calculation, and a value of 1 indicates that the keypoint is included in the loss function calculation. i v j Let E be the visibility flag for the i-th and j-th key points, E be the edge set for constructing topological connections, and μ be the value of μ. ij Let s be the direction vector of the joint formed by key points i and j. ij Let μ be the cosine similarity between the direction vector and the true direction vector. ij (1-s ij ) represents the joint dynamic weights, and mask() is a function used to characterize the visibility of keypoint pairs if and only if v i >0 and v j When the value is greater than 0, mask > 0.

[0023] Furthermore, a CHP filter is used for smoothing. 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 target angle Calculate the length scaling factor r of joint (i,j) ij and rotation matrix R ij , respectively represented as:

[0034]

[0035] In the formula, Δθ ij For the current angle θ ij From the perspective of the target deviation value

[0036] Based on the length scaling factor r ij and rotation matrix R ij and the confidence level p of key point prediction i p j Scale and rotate the joint (i,j) to update the coordinates of the human body keypoints i and j. And calculate the distance moved between the current coordinates and the updated coordinates. CHP filter adjustment coefficient Where k is the set scaling factor, γ min γ max These are the minimum and maximum values ​​of the adjustment coefficients for the CHP filter;

[0037] Based on the CHP filter adjustment coefficient γ i For the f min Dynamically update the rate coefficient β:

[0038]

[0039] β i =γ i ·β i

[0040] In the formula, the subscript i represents the i-th key point.

[0041] Furthermore, the SG-RT-PicoDet network is used for target detection to obtain the wearable equipment detection box. The SG-RT-PicoDet network is improved by adding reparameterizable branches to the backbone of the standard PicoDet network. These reparameterizable branches are connected to the ends of the Bottleneck modules in Stages 2 to 4 of the backbone network and run in parallel with the original depthwise separable convolutional kernels.

[0042] Each reparameterizable branch includes a 1×1 convolution and a 3×3 depthwise separable convolution. Each reparameterizable branch is processed to obtain a single equivalent convolution kernel, and then multi-branch fusion is performed to obtain the fused equivalent convolution kernel and bias, which are expressed as follows:

[0043]

[0044] In the formula, W eq b is the equivalent convolution kernel after fusion. eq The bias after fusion is N, and W is W. i For each reparameterizable branch, the convolution kernel γ i σ is the scaling factor. i μ is the standard deviation. i β is the mean. i This is the offset.

[0045] Furthermore, the SG-RT-PicoDet network incorporates a TAD-ATSS label allocation strategy during training, wherein the processing steps of the TAD-ATSS label allocation strategy include:

[0046] The sample allocation threshold is dynamically adjusted, and the intersection-union ratio (IU) of the preset bounding box and the target bounding box, the predicted bounding box, and the target bounding box is calculated. The joint IU is then derived, whereby the joint IU is expressed as:

[0047]

[0048] In the formula, C IoU (i,j) represents the joint intersection-union ratio, α and β are weight balancing coefficients, and α + β = 1. In the initial stage of model training, the accuracy of the model's prediction results is low, so the value of β is small to reduce the influence of the predicted bounding box on the label assignment. As training progresses, the values ​​are continuously adjusted to achieve dynamic balance. IoU (i,j) represents the intersection-union ratio of the preset bounding box and the target bounding box, P IoU (i,j) represents the intersection-union ratio (I / J) of the predicted bounding box and the target bounding box, a i For the preset anchor frame, g j For the target bounding box, p i For prediction;

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

[0050]

[0051] In the formula, Dynamic threshold for the target bounding box;

[0052] Collaborative optimization of classification and regression tasks, based on joint intersection-over-union ratio (CIU) and classification confidence s.i Calculate the task consistency score The task consistency score is used as the weight for calculating the classification and regression losses, and candidate boxes of positive samples are finally selected as target boxes. Here, ζ is the weight coefficient for the classification task, and η is the weight coefficient for the regression task. The expressions for the classification and regression losses are:

[0053]

[0054] In the formula, L cls For classification loss, N pos For positive samples assigned as target boxes, BCE is the binary cross-entropy. For the normalized task consistency score, N neg These are the negative samples assigned as the target boxes, γ is the focus parameter of the focus loss, and s j L represents the classification confidence of the candidate bounding boxes. reg For regression loss, GIoU is the regression loss function, g i The target bounding box.

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

[0056] The processing steps of the improved LCNet lightweight feature extraction network include:

[0057] Input the image of the wearable equipment area, extract features, and obtain a fixed-length feature vector as the wearable feature.

[0058] Furthermore, the improved LCNet lightweight feature extraction network combines metric loss and classification loss as loss functions, specifically expressed as:

[0059]

[0060] In the formula, L CE The loss function is used for classification, where N is the number of samples and C is the number of classes. For true category labels using one-hot encoding, This indicates that sample i belongs to category c; otherwise, it does not. It is the probability that sample i belongs to class c, as predicted by the network. To measure the loss, triplet Includes anchor point x a Positive sample x p and negative sample x nα is a predefined angle threshold, [z] + =max(0,z) is used to ensure that the loss is non-negative.

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

[0062] Graph construction phase:

[0063] The wearable feature library is hierarchically clustered according to feature categories to construct a multi-level graph structure, with the center vector of each category as the top-level entry node.

[0064] Graph retrieval stage:

[0065] For retrieval vector First, calculate the distance to all category center vectors, and obtain the category center vector with the most similarity based on the distance metric. Use this most similar category center vector as the initial entry node. Let be a set of d-dimensional real numbers, where d is the dimension of the wearable features, and the initial entry node is represented as... It is category c * The set of nodes below, Let μ be the set of category centers. c Let v be the class center vector. The vector traversed in the middle;

[0066] Based on the determined initial entry node, from the highest level L max Starting from the entry node, the nearest neighbor set is searched layer by layer downwards, and the wearing features most similar to the wearing features of the wearable equipment region image to be identified are selected. The specific steps include:

[0067] Each layer uses a greedy search strategy 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 search vector q, then stop the search and proceed to the next level. Let v be the set of neighboring nodes. current d represents the currently retrieved node, and d represents the distance between nodes;

[0068] In the lowest-level search, an expanded search strategy is used to generate a set of candidate nodes. A class consistency constraint is added to the distance constraint, and the candidate nodes are scored. The scoring results are as follows:

[0069]

[0070] In the formula, scores i The candidate nodes are scored, where λ is the weighting coefficient. This indicates the relationship between the current node and the initial entry node v. entry The categories are consistent; a value of 0 indicates inconsistency.

[0071] Based on the scoring results, the k nodes with the lowest scores are selected from the candidate node set as the final matching result, i.e., the most similar wearing features.

[0072] Furthermore, the step of using a hysteresis comparison strategy for multi-frame integration optimization includes:

[0073] The percentage of video frames that are deemed to be properly worn is calculated based on the most recent N video frames and compared with set high and low thresholds. If the percentage of video frames exceeds the high threshold, the wear is deemed to be properly worn; if the percentage of video frames is below the low threshold, the wear is deemed to be improper.

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

[0075] (1) This invention uses a Modified-YOLOv8-Pose network for human pose estimation, extracts the coordinates of key human points, and combines smoothing processing for stability optimization, thereby improving detection accuracy and robustness. Based on human skeleton information, an SG-RT-PicoDet network is used for accurate localization of the wearable equipment area, and the localization is optimized by combining a TAD-ATSS label allocation strategy, thereby improving the efficient feature extraction capability of the target detection algorithm and enabling wearable equipment recognition to operate stably even in complex industrial scenarios (such as dense crowds, uneven lighting, target occlusion, etc.).

[0076] (2) In constructing the wearable feature library based on few-sample learning, this invention uses the lightweight LCNet network for feature extraction and constructs an efficient graph-structured wearable equipment feature library. Combined with the C-HNSW graph retrieval algorithm for target region feature matching, this invention enables efficient modeling and rapid transfer under limited sample data conditions, solving the problem of traditional methods' dependence on large-scale labeled data. This method supports wearable recognition under open-category conditions and is applicable to different work scenarios and various wearable equipment categories.

[0077] (3) This invention uses the SG-RT-PicoDet network to locate the wearable equipment area and combines it with the TAD-ATSS tag allocation strategy for area positioning optimization. This detection process can automatically adapt to different types of wearable equipment, reducing false detection problems caused by differences in equipment form.

[0078] (4) The Modified-YOLOv8-Pose network of this invention introduces the HG Block module to improve the network's ability to detect small targets (such as key points of hands and feet), and combines the C-BRAM attention mechanism to enhance the feature expression between channels through the lightweight channel attention module. The global spatial dynamic modeling is realized through the spatial attention dual-layer routing module, thereby enhancing the network's ability to extract features from key areas and reducing complex background interference in industrial scenarios.

[0079] (5) In the detection stage, this invention performs target region feature retrieval and temporal fusion. Feature vectors are extracted for the target region, and efficient retrieval and matching are performed using a C-HNSW graph retrieval algorithm optimized based on category centers, ensuring high-precision matching results even with small sample sizes. To improve the stability of detection, this method employs a hysteresis comparison strategy to perform temporal series fusion optimization on multi-frame detection results to obtain the current recognition result.

[0080] (6) This invention supports rapid adaptation to different industrial scenarios. Only a small number of samples of the new target are needed to update the wearable feature library to complete the model transfer. It is applicable to a variety of industries (such as manufacturing, chemical industry, construction, etc.) and has broad application prospects and market value. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0082] Figure 2 This is a diagram of the improved Modified-YOLOv8-Pose network structure of the present invention;

[0083] Figure 3 This is a diagram of the SG-RT-PicoDet network structure of the present invention;

[0084] Figure 4 This is a diagram of the improved LCNet lightweight feature extraction network structure of the present invention;

[0085] Figure 5 This is a hierarchical diagram of the software system of the present invention. Detailed Implementation

[0086] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0087] This embodiment provides a lightweight intelligent identification method for wearable equipment under small sample conditions. This method achieves efficient identification of wearable equipment, rapid adaptation to different industrial environments, and meets real-time operation requirements under low computing resource conditions, reducing the occurrence of dangerous accidents and improving the safety and intelligence level of industrial production, even with limited sample data. Specifically, as... Figure 1 As shown, the method includes the following steps:

[0088] S1. Acquire video frame images from the industrial production environment through a network camera and transmit them to a backend server or edge computing device.

[0089] This embodiment uses the RTSP protocol for video stream transmission and performs image sampling at a fixed frame rate to ensure the real-time performance and stability of subsequent algorithms.

[0090] S2. The proposed Modified-YOLOv8-Pose algorithm is used to estimate the human pose in the image, obtain the coordinates of human key points, and combine it with the designed CHP filter to optimize the stability of key points, thereby improving the detection accuracy and robustness.

[0091] In this step, the designed Modified-YOLOv8-Pose algorithm is used for human pose estimation. Based on the human pose estimation results, it is determined whether a human target exists in the current frame. If no worker is detected, the process proceeds directly to the next frame; if a worker is detected, the process continues to the next step of keypoint screening and region detection.

[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 μ is the cosine similarity between the orientation vector and the true orientation vector, thus measuring the degree of orientation matching between human joints. ij (1-s ij ) represents the joint dynamic weights, and mask() is a function used to characterize the visibility of keypoint pairs if and only if v i >0 and v j When the value is greater than 0, mask > 0.

[0095] The improved Modified-YOLOv8-Pose network structure diagram is as follows: Figure 2 As shown, the network structure includes a Backbone section, a Neck section, and a Head section. The Backbone section 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 an SPPF module connected in sequence. The Neck section includes a first upsampling module, a first connection module, a third C2f module, a second upsampling 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 to the SPPF module. The first connection module is also connected to the first HG Block module, the third C2f module is also connected to the third connection module, the second connection module is also connected to the first C2f module, and the fourth connection module is also connected to the SPPF module. The Head section includes three Pose modules, each of which is connected to the corresponding second HG Block module, third HG Block module, and fourth HG Block module.

[0096] The first, second, third, and fourth HG Block modules each include multiple consecutive convolutional layers to fuse multi-scale information and support residual connections. The HG Block module consists of six consecutive 3×3 convolutional layers, fusing multi-scale information and employing compression and activation mechanisms to reduce computational overhead, enhance feature representation capabilities, and support residual connections to mitigate the gradient vanishing problem in deep networks.

[0097] The C-BRAM module includes a lightweight channel attention submodule and a spatial attention two-layer routing submodule. The processing of the lightweight channel attention submodule includes:

[0098] The feature map obtained by the C-BRAM module is processed by the C2F module in front, and the global information of each channel is extracted to obtain the channel description vector.

[0099] Based on the channel description vector, the initial inter-channel weights are calculated, and the final inter-channel weights are obtained by Sigmoid normalization, thereby enhancing the feature representation between channels.

[0100] The processing steps of the spatial attention-based two-layer routing submodule include:

[0101] The feature map after channel attention weighting is divided into regions, and Query, Key, and Value values ​​are generated by linear projection, and region-level routes are constructed.

[0102] Based on regional routing, regional affinity is calculated for coarse-grained filtering, and spatial weights are obtained through token-level fine-grained attention calculation to ensure the network's feature representation capability.

[0103] This step also involves designing a CHP filter to optimize keypoint coordinates and improve the stability of the detection results. The CHP filter, based on human joint topology information (as shown in Table 1 below), combines joint length constraints and angle constraints to optimize the stability of keypoint coordinates during video stream inference.

[0104] Table 1 Topological structure of key points in the human body

[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 α of the filter is adaptively adjusted to maintain a balance between filter strength and delay. The relationship between the original value and the smoothed value satisfies... in, X represents the predicted coordinates of the keypoints at times t and t-1. t Here are the original coordinates of the human body key points at time t, and the smoothing coefficient. The rate coefficient β is used to determine the degree of influence of the signal change rate on the cutoff frequency, τ is the time constant of the filter, and T e f is the sampling period, i.e., the inference frequency of the attitude estimation algorithm. c f is the cutoff frequency. min The minimum cutoff frequency is used to control the basic smoothing speed. The input signal is the rate of change of the coordinates of the key points.

[0107] When establishing joint length and angle constraints, for a pair of human keypoints i and j with physical connections, the joint vector v of joint (i,j) is first calculated. ij =(x j -x i ,y j-y i Next, calculate its length. Angle with joint (p,q) 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.

[0108] Next, combining the target value from the previous frame and the smoothing coefficient α, the target length for the t-th frame is calculated. and target angle Updated, after the update Thus, the length scaling factor of the key point (i,j) can be calculated. Rotation matrix Where Δθ ij For the current angle θ ij From the perspective of the target deviation value

[0109] Finally, based on r ij R ij and the confidence level p of key point prediction i ,p j The joints are scaled and rotated to update the coordinates of keypoints i and j. Get the distance moved between the current value and the updated value. and filter adjustment coefficient Based on this coefficient, the parameters of the CHP filtering algorithm are dynamically adjusted: β i =γ i ·β i Where k is the scaling factor, which is set to 1.5 in this paper, γ min γ max These are the minimum and maximum values ​​of the adjustment coefficients for the CHP filter.

[0110] S3. Based on human skeleton information, the proposed SG-RT-PicoDet algorithm is used for wearable equipment area localization, and the designed TAD-ATSS tag allocation strategy is combined to optimize target detection, thereby improving detection accuracy and the accuracy of the detection area.

[0111] like Figure 3As shown, this step improves upon the standard PicoDet network by adding a reparameterizable branch to the standard PicoDet backbone to maintain inference speed while enhancing the network's ability to extract complex features; designing a TAD-ATSS label allocation strategy to optimize the allocation of positive samples in object detection tasks and improve detection accuracy in complex environments; and incorporating a candidate box optimization method based on skeleton heatmaps to guide the model to focus on the area surrounding the human body and improve the localization accuracy of the wearable equipment area. The SG-RT-PicoDet algorithm detects the main target in the input image and outputs a filtered high-confidence wearable equipment detection box, which serves as the input for subsequent feature extraction and recognition modules.

[0112] The reparameterizable branches add 1×1 convolutions and 3×3 depthwise separable convolutions at the ends of the Bottleneck modules in Stages 2 through 4 of the network, running in parallel with the existing 5×5 depthwise separable convolution kernels to improve the network's feature extraction capabilities. During model inference, branch fusion is performed, expanding smaller convolution kernels to the same size as the main branch using zero-padding. For each branch's convolution kernel W... i The corresponding batch normalization parameters are the mean μ i Standard deviation σ i Scaling factor γ i and offset β i Calculate a single equivalent convolution kernel that fuses multiple branches. and bias This helps maintain the speed of reasoning.

[0113] The TAD-ATSS label allocation strategy is based on the task alignment learning concept, dynamically and adaptively allocating positive samples for model training. First, it dynamically adjusts the sample allocation threshold, calculates the intersection-union ratio (IU) of the preset bounding box and the target bounding box, and the predicted bounding box and the target bounding box, and then derives the joint IU. Calculate the dynamic threshold of the target box based on its mean and standard deviation.

[0114] Next, the classification and regression tasks are collaboratively optimized based on the joint intersection-union ratio and the classification confidence s calculated during model inference. i Calculate the task consistency score The task consistency score is then used as the weight for calculating the classification and regression losses. Finally, the task consistency score t... i Normalization As classification loss L cls and regression loss L reg The calculated weights further guide the model to better align with classification and regression tasks. Where BCE is the binary cross-entropy, γ is the focus loss focusing parameter, and N pos These are the positive samples assigned as target boxes, N neg These are negative samples assigned as target boxes, and GIoU is the regression loss function.

[0115] S4. Based on the LCNet lightweight classification network, feature extraction is performed on a small number of samples of wearable equipment to build a wearable feature library.

[0116] like Figure 4 As shown, in this step, the LCNet lightweight classification network uses a depthwise separable convolutional structure, and the network architecture is optimized for adaptability. This includes removing intermediate ReLU activation layers, adding bias terms to the convolutional layers to improve the fitting ability to small sample distributions, and adjusting the convolutional stride to increase the feature extraction capability in fine-grained space. In the feature encoding stage, a BN Neck module is introduced to normalize the output feature vector, and the consistency of small sample feature distributions is improved by simultaneously optimizing the classification loss and metric loss. The entire network execution flow is as follows: after the input image extracts features through convolutional layers and depthwise convolutional modules, a fixed-length feature vector is output, which serves as the basis for constructing the wearable feature library and subsequent graph retrieval matching.

[0117] To extract feature vectors with good intra-class compactness and inter-class separability, this embodiment of the invention combines metric loss and classification loss as loss functions. For a classification task containing C classes and N samples, the classification loss is defined as... For sample i, These are true category labels using one-hot encoding. This indicates that sample i belongs to category c; otherwise, it does not. This is the probability predicted by the model that sample i belongs to class c. It measures the loss. Among them, triplet Includes anchor point x a Positive sample x p and negative sample x n α is a predefined 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. entryThe categories are consistent, while a value of 0 indicates inconsistency. The k nodes with the lowest scores are selected as the final matching results. This process significantly improves the retrieval efficiency and accuracy of graph structures under small sample conditions, supports O(log N) level approximate nearest neighbor query capability, and is suitable for real-time requirements in wearable recognition scenarios.

[0122] The above-mentioned hysteresis comparison strategy optimizes the detection results in the following way: it calculates the proportion of qualified images based on the most recent N frames and compares them with the set high and low thresholds; when the proportion of qualified frames exceeds the high threshold, the wearable equipment is judged to be qualified; when the proportion of qualified frames is lower than the low threshold, the wearable equipment is judged to be unqualified; this strategy integrates temporal information to improve the stability of the detection results and reduce misjudgments caused by false detections of single frames.

[0123] S6. Deploy the trained model to the domestic AX650N chip for edge computing to achieve real-time detection, violation alarm, monitoring screen preview and statistical visualization functions.

[0124] In this embodiment, the trained model is deployed on the high-performance, low-power AX650N chip, and model pruning and quantization techniques are used to reduce computing resource consumption, enabling the system to run efficiently on edge devices and meet the real-time requirements of industrial sites.

[0125] Ultimately, the system displays the detection results through a real-time visual interface and supports violation alarms, monitoring screen previews, and statistical visualization functions. For personnel not wearing equipment correctly, the system can automatically issue warnings and notify management personnel for handling via voice reminders and warning light alarms.

[0126] The system is like Figure 5 As shown, it includes:

[0127] 1) Perception Layer: Responsible for the acquisition and preliminary processing of real-time monitoring video streams. It utilizes the RTSP protocol for data transmission, providing high-quality video frame images to the business and application layers.

[0128] 2) Data Layer: Used to store and manage structured and unstructured data generated during system operation. Structured data includes camera information, exception logs, etc.; unstructured data includes abnormal images. The data layer receives data to support the visualization of data in the application layer.

[0129] 3) Business Layer: Integrates service scheduling and recognition algorithm modules, and provides corresponding API interfaces. Based on the detection results of the recognition algorithm, the business layer generates information on unqualified wearables and transmits it to the data layer through standardized interfaces to support real-time display and user interaction on the front end.

[0130] 4) Application layer: Provides users with an intuitive interactive interface, supporting functions such as real-time monitoring display, abnormal behavior statistics and trend analysis, and viewing and managing network camera information.

[0131] Tests on real-world test datasets show that the wearable equipment recognition accuracy reaches 98.6%, and it can maintain stable performance even in complex scenarios (changing lighting, dense crowds, and diverse equipment); the inference frame rate reaches 30 FPS, meeting the requirements of real-time monitoring; and only a small amount of sample data (5-10 images per type of equipment) is needed to complete model transfer, which meets the needs of rapid deployment in practical applications.

[0132] This system enables efficient edge computing deployment. After optimizing the trained model, it is deployed to the domestically produced AX650N chip. Model pruning and quantization techniques are employed to reduce computational resource consumption, allowing the system to operate efficiently in embedded devices or low-computing-power environments, meeting the real-time requirements of industrial sites. Simultaneously, the system supports monitoring screen previews, violation alarms, and statistical visualization functions, improving the level of intelligence in industrial production.

[0133] This system enables accurate, rapid, and convenient real-time automated detection of equipment worn by factory workers, reducing manual inspection costs, avoiding the instability of manual inspection, and improving the efficiency of production safety management. Through real-time alarms and visual feedback, it helps managers quickly identify non-compliant wearing behaviors, reducing the incidence of safety accidents and ensuring the safety of industrial production.

[0134] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0136] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0137] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0138] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0139] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0140] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent recognition of lightweight wearable equipment under small sample conditions, characterized in that, Includes the following steps: Capture video frame images of the industrial production environment; A Modified-YOLOv8-Pose network is used to estimate human pose in the video frame images to determine if a person is present. If a person is present, key points are extracted and smoothed to form human skeleton information. If no one is present, the video frame images are re-acquired. The Modified-YOLOv8-Pose network is improved by introducing HGBlock, C-BRAM, and pose loss functions into the YOLOv8-Pose network. The Modified-YOLOv8-Pose network includes a Backbone, a Neck, and a Head section. The Backbone section 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 an SPPF module connected in sequence. The Neck section includes a first upsampling module, a first connection module, a third C2f module, a second upsampling module, a second connection module, a second HG Block module, a fifth CBS module, a third connection module, and a third HG Block module connected in sequence to the SPPF module. The head section includes a block module, a sixth CBS module, a fourth connection module, and a fourth HG block module. The first connection module is also connected to the first HG block module. The third C2f module is also connected to the third connection module. The second connection module is also connected to the first C2f module. The fourth connection module is also connected to the SPPF module. The head section includes three pose modules, each of which is connected to the corresponding second HG block module, third HG block module, and fourth HG block module. Based on the human skeleton information, target detection is performed on the area of ​​the human body where the equipment is worn, and the data is fused with the set key wearing area cropped and associated with the corresponding human target to obtain an image of the area where the equipment is worn. An improved LCNet lightweight feature extraction network is used to extract wear features from the wearable equipment region image and construct a wearable feature library. The improved LCNet lightweight feature extraction network is improved by removing the intermediate ReLU activation layer in the standard LCNet lightweight feature extraction network, adding a bias term in the convolutional layer, adjusting the convolution stride, and introducing a BN Neck module in the feature encoding stage. The wear features of the wearable equipment region image to be identified are obtained, and the wearable feature library is searched and matched. At the same time, a hysteresis comparison strategy is used to integrate and optimize multiple frames to obtain the wear recognition result of the wearable equipment.

2. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, The first HG Block module, the second HG Block module, the third HG Block module, and the fourth HG Block module all include multiple consecutive convolutional layers to fuse multi-scale information and support residual connections; The C-BRAM module includes a lightweight channel attention submodule and a spatial attention two-layer routing submodule. The processing procedure of the lightweight channel attention submodule includes: Based on the feature map obtained by the C2F module before the C-BRAM module, global information of each channel is extracted to obtain the channel description vector; Based on the channel description vector, the initial inter-channel weights are calculated, and the final inter-channel weights are obtained by Sigmoid normalization. The processing procedure of the spatial attention two-layer routing submodule includes: The feature map after channel attention weighting is divided into regions, and Query, Key, and Value values ​​are generated by linear projection, and region-level routes are constructed. Based on the regional routing, regional affinity is calculated for coarse-grained filtering, and spatial weights are obtained through fine-grained attention calculation at the token level.

3. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 2, characterized in that, The expression for the attitude loss function is: In the formula, This is the loss value. For loss based on Euclidean distance, The adjustment coefficient is used to balance the weights of distance loss and orientation loss. For joint-oriented losses, The total number of key points for a single human body. For the first i The Euclidean distance between the predicted and actual locations of each key point The area of ​​the human body where the key point is located. For the first i A fixed standard deviation at key points ={0,1}, where 0 indicates that keypoints are not included in the loss function calculation, and 1 indicates that keypoints are included in the loss function calculation. , For the first i, j Visibility indicators of key points To construct the edge set for topological connectivity, Key point i , j The direction vector of the joint, Let be the cosine similarity between the direction vector and the true direction vector. For joint dynamic weights, For a function used to characterize the visibility of keypoint pairs, if and only if and hour, .

4. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, A CHP filter is used for smoothing. The relationship between the smoothed human key points and the original human key points satisfies: In the formula, , for t and t- Predicted coordinates of key points at time 1. For smoothing coefficients, for t The original coordinates of key human body points at any given moment. The time constant of the filter, The sampling period is the inference frequency of the attitude estimation algorithm. The cutoff frequency, The minimum cutoff frequency is used to control the basic smoothing speed. The rate coefficient is used to determine the degree to which the rate of signal change affects the cutoff frequency. The input signal is the rate of change of the key point coordinates. in, and rate coefficient To perform dynamic updates, the specific steps include: For a pair of key points of the human body with a physical connection i and j Calculate joints joint vectors ,in Key points of the human body i coordinates Key points of the human body j The coordinates; Based on the joint joint vectors Calculate joints length and with joints The included angle The calculation expressions are as follows: Combining the target length, target angle, and smoothing coefficient of the previous video frame image , for the t Target length of video frame image and target angle The update is performed, and the update expressions are as follows: In the formula, For the first t- Target length of 1 video frame image For the first t- Target angle of 1 video frame image; Based on the updated target length and target angle Calculate joints Length scaling factor and rotation matrix , respectively represented as: In the formula, From the current perspective From the perspective of the target deviation value ; Based on the length scaling factor and rotation matrix and the confidence level of key point prediction , For joints Perform length scaling and rotation to adjust the human body key points. i and j The coordinates are updated to , And calculate the distance moved between the current coordinates and the updated coordinates. CHP filter adjustment coefficient ,in, The scaling factor is set. , The minimum and maximum values ​​of the adjustment coefficients for the CHP filter; Based on the CHP filter adjustment coefficient Regarding the above and rate coefficient Dynamic updates: In the formula, the subscript i Let i be the i-th key point.

5. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, Target detection is performed using the SG-RT-PicoDet network to obtain bounding boxes for wearable equipment. The SG-RT-PicoDet network is improved by adding reparameterizable branches to the backbone of the standard PicoDet network. These reparameterizable branches are connected to the ends of the Bottleneck modules in Stages 2 to 4 of the backbone network and run in parallel with the original depthwise separable convolutional kernels. Each reparameterizable branch includes a 1×1 convolution and a 3×3 depthwise separable convolution. Each reparameterizable branch is processed to obtain a single equivalent convolution kernel, and then multi-branch fusion is performed to obtain the fused equivalent convolution kernel and bias, which are expressed as follows: In the formula, The equivalent convolution kernel after fusion. For the bias after fusion, for, For each reparameterizable branch, the convolutional kernel, Scaling factor Standard deviation, The mean, This is the offset.

6. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 5, characterized in that, The SG-RT-PicoDet network incorporates the TAD-ATSS label allocation strategy during training. The processing steps of this TAD-ATSS label allocation strategy include: The sample allocation threshold is dynamically adjusted, and the intersection-union ratio (IU) of the preset bounding box and the target bounding box, the predicted bounding box, and the target bounding box is calculated. The joint IU is then derived, whereby the joint IU is expressed as: In the formula, For joint intersection and union comparison, , This is the weighting balance coefficient, and In the initial stage of model training, the accuracy of the model's prediction results is low, therefore The value is set relatively small to reduce the impact of the predicted bounding box on label assignment. It is continuously adjusted as training progresses to achieve dynamic equilibrium. This is the intersection-union ratio between the preset bounding box and the target bounding box. The intersection-union ratio (IUU) of the predicted bounding box and the target bounding box. For the preset anchor frame, For the target bounding box, For prediction; The target box dynamic threshold is calculated based on the mean and standard deviation of the joint intersection-union ratio, wherein the target box dynamic threshold is expressed as: In the formula, Dynamic threshold for the target bounding box; Collaborative optimization of classification and regression tasks, based on joint intersection-over-union ratio and classification confidence. Calculate the task consistency score The task consistency score is used as the weight for calculating the classification and regression loss, and finally, candidate boxes of positive samples assigned as target boxes are selected. These are the weighting coefficients for the classification task. The weights for the regression task are given, and the expressions for the classification and regression losses are: In the formula, For classifying losses, For positive samples assigned as target boxes, BCE is the binary cross-entropy. The normalized task consistency score, These are the negative samples assigned as the target boxes. It is the focus loss focusing parameter. The classification confidence score for the candidate bounding boxes. To regress the loss, It is a regression loss function. The target bounding box.

7. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, The processing steps of the improved LCNet lightweight feature extraction network include: Input the image of the wearable equipment area, extract features, and obtain a fixed-length feature vector as the wearable feature.

8. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, The improved LCNet lightweight feature extraction network combines metric loss and classification loss as loss functions, specifically expressed as follows: In the formula, For classifying losses, For the sample size, For the number of categories, For true category labels using one-hot encoding, This indicates that the sample i Category c Otherwise, it does not belong to [the category of] ... These are samples predicted by the network. i Category c The probability, To measure the loss, triplet , including anchor points Positive samples and negative samples , For a predefined angle threshold, Used to ensure that losses are non-negative.

9. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, The C-HNSW graph retrieval algorithm is used for retrieval and matching. The specific steps include: Graph construction phase: The wearable feature library is hierarchically clustered according to feature categories to construct a multi-level graph structure, with the center vector of each category as the top-level entry node. Graph retrieval stage: For retrieval vector First, calculate the distance to all category center vectors, and obtain the category center vector with the most similarity by distance metric. Use the most similar category center vector as the initial entry node, where... Let be a set of d-dimensional real numbers, where d is the dimension of the wearable features, and the initial entry node is represented as... , It is a category The set of nodes below, , For the set of category center points, For the class center vector, for The vector traversed in the middle; Based on the determined initial entry node, from the highest level Starting from the entry node, the nearest neighbor set is searched layer by layer downwards, and the wearing features most similar to the wearing features of the wearable equipment region image to be identified are selected. The specific steps include: Each layer uses a greedy search strategy 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... If the distance is reached, the search stops and proceeds to the next level. For the set of neighboring nodes, This is the currently retrieved node. The distance between nodes; In the lowest-level search, an expanded search strategy is used to generate a set of candidate nodes. A class consistency constraint is added to the distance constraint, and the candidate nodes are scored. The scoring results are as follows: In the formula, Scoring the candidate nodes These are the weighting coefficients. Indicates the relationship between the current node and the initial entry node. The categories are consistent; a value of 0 indicates inconsistency. Based on the scoring results, the node with the lowest score is selected from the candidate node set. The nodes are used as the final matching result, i.e., the most similar wearing features.

10. The intelligent recognition method for lightweight wearable equipment under small sample conditions according to claim 1, characterized in that, The steps for multi-frame integration optimization using a hysteresis comparison strategy include: The percentage of video frames that are deemed to be properly worn is calculated based on the most recent N video frames and compared with set high and low thresholds. If the percentage of video frames exceeds the high threshold, the wear is deemed to be properly worn; if the percentage of video frames is below the low threshold, the wear is deemed to be improper.

Images