Method implemented in computer and system for real-time verification of skin coverage of users by personal protective equipment

Description

/ 21 METHOD IMPLEMENTED IN COMPUTER AND SYSTEM FOR REAL-TIME VERIFICATION OF SKIN COVERAGE OF USERS BY PERSONAL PROTECTIVE EQUIPMENT Field of invention

[001] The present invention falls within the field of computer vision applied to occupational safety, dealing with a computer-implemented method that, through 2D pose estimation and skin chromaticity analysis (human) in RGB images, determines in real time the percentage of skin coverage of anatomical segments, allowing inferences about compliance in the use of Personal Protective Equipment (PPE) in industrial environments, hospitals, professional kitchens, construction sites and other places where dermal exposure represents a risk. Fundamentals of the invention

[002] US patent 9,695,981 B2, filed April 20, 2012, describes a system for verifying PPE compliance in work areas, using image capture, background removal, segmentation of regions of interest, feature extraction, and classification via convolutional neural networks to detect gloves, helmets, and other items in real time. A relevant limitation of the present invention is that this approach requires multiple CNN (Convolutional Neural Network) inferences per frame and large labeled databases, and depends on dedicated hardware, making it vulnerable to occlusions and lighting variations. The technical problem addressed is the ineffectiveness of technologies such as RFID in distinguishing between possession and actual use of equipment, jeopardizing safety and productivity.The technical solution applies a distributed computer vision pipeline across CPU and GPU modules, among others, with geometric analysis of the PPE's positioning in relation to the user's estimated skeleton, granting or denying access to devices; the results indicate low-latency detection compatible with up to 5 fps on dedicated hardware, and accuracy exceeding 90% in identification and registration. Petition 870260032128, dated 07 / 04 / 2026, page 3 / 27 2 / 21 compliance, and automatic generation of access logs. Despite the gains, it maintains a high computational cost, while the proposed invention reduces the dependence on additional networks and data annotation, operating on edge devices < 5W with unique pose inference and skin mask.

[003] The article “Identification System of Personal Protective Equipment Using Convolutional Neural Network (CNN) Method” by Pradana et al., presented at the 2019 International Symposium on Electronics and Smart Devices, describes a CNN-based PPE verification system. The proposal arises from the high rate of workplace accidents in Indonesia (157,313 cases in 2018), resulting from a lack of discipline in the use of helmets, gloves, goggles, masks, and ear protectors in industrial environments. The main technical solution combines image capture by RGB camera and inductive proximity sensor for footwear, followed by grayscale preprocessing, two convolutional layers (5*5*32 and 3*3*64) with maxpooling, a “flatten” resulting in a vector of 91392 *1 and a hidden layer of 256 neurons. Classification occurs in 12 PPE categories in real time via the Tkinter interface, eliminating the need for multiple specialized networks.In real-world tests, accuracy reached 85.83% for training set samples, 80% for external samples, and 73.34% in female users, highlighting a greater failure in detecting eyeglasses and ear protectors. A relevant limitation: this CNN method requires an extensive labeled set for each type of PPE, multiple inferences per frame, and dedicated hardware to maintain performance (mAP > 0.9 at 2-5 fps), being vulnerable to occlusions and lighting variations, which contrasts with the low-cost efficiency and single inference per pose + skin mask of the present invention.

[004] The bibliographic reference from Delhi et al. (2020), “Detection of Personal Protective Equipment (PPE) Compliance on Construction Site Using Computer Vision Based Deep Learning Techniques”, published on September 24, 2020 in Frontiers in Built Environment, describes a transfer learning-based CNN framework with YOLOv3 for near real-time detection of helmets and safety vests on construction sites. The authors collected 2509 images via video from industrial sites and web scraping, performing 90:8:2 partitioning and data augmentation for robustness. The model ran on an Intel i7-4790 CPU (8 GB RAM), processing at 1 to 2 fps in Petition 870260032128, dated 07 / 04 / 2026, page 4 / 27 3 / 21 video stream and presenting average accuracy and recall of 96% in the validation set (F1 = 0.96) and F1 of 0.97 in the test, validating the effectiveness of the modified YOLOv3 for four classes: SAFE, NOTSAFE, NoHardhat, and NoJacket. Despite the good results, this approach relies on bounding-box detections and requires multiple inferences for different types of PPE, without quantitatively measuring the fraction of uncovered skin or providing a continuous coverage index. A relevant limitation of the present invention is that it does not quantify the fraction of exposed skin per anatomical segment, nor does it use a single inference to measure skin coverage with a continuous metric.

[005] The article by Ruoxin Xiong and Pingbo Tang, “Pose guided anchoring for detecting proper use of personal protective equipment,” was received on September 23, 2020. This work addresses the problem of detecting compliance with Personal Protective Equipment (PPE) in construction environments, where conventional methods based on bounding boxes and multiple classifiers require vast labeled sets, show low robustness to partial occlusions and scale variations, and have high computational cost. The main technical solution consists of employing 2D pose estimation (via OpenPose with MobileNet backbone) to generate skeletal anchors that delineate attention regions in body parts (head and upper torso), on which specialized shallow CNN classifiers are applied to recognize the correct use of helmets and vests.In experiments with the CPPE (Construction Personal Protective Equipment) dataset, the method achieved an F1-score of 0.97 in helmet detection and 0.95 in vest detection, with 97.01% accuracy and 96.89% recall for helmets, and 95.68% accuracy and 93.56% recall for vests. A relevant limitation of the present invention is that it does not quantify the fraction of exposed skin, and the method relies on training and inference from multiple CNNs with annotated data, whereas the present invention performs a single pose inference combined with a deterministic skin mask, eliminating the need for additional networks and labeled data.

[006] The article Fast Personal Protective Equipment Detection for Real Construction Sites Using Deep Learning Approaches, submitted on April 16, 2021 (Wang et al., Sensors 21,3478, 2021) presents PPE detectors based on onestage architectures (YOLO v3, v4 and v5) trained on a new CHV (Color Helmet and Vest) set of Petition 870260032128, dated 07 / 04 / 2026, p. 5 / 27 4 / 21 1330 images collected from real construction sites, with annotation of people, reflective vests, and helmets in four colors. The work addresses the shortcomings of previous solutions, which are limited to a few types of PPE, depend on large labeled datasets and dedicated GPUs, suffer from occlusions and lighting variations, and do not achieve real-time performance on modest hardware. The proposed solution adjusts network depth, number of detection layers, and input size, coupled with a training pipeline on Darknet and PyTorch. The models achieve mAP of up to 86.55% with YOLO v5x and speeds of up to 52 FPS with YOLO v5s on GPU, surpassing the results of YOLO v3 and Pelee-RPA on Pictor-v3 and GDUT-HWD datasets by approximately 10-11%, without loss of accuracy in images with blurred faces.A relevant limitation of the present invention is that it relies on multiple CNN inferences and voluminous annotated data for each color and type of PPE, requires a dedicated GPU, and does not offer continuous skin coverage metrics per single pose inference.

[007] IL298700A, an automated barcode-based system for PPE usage control, requires multiple spaced IP cameras and a central controller; solution filed on June 4, 2021. The technical problem addressed is the continuous and accurate monitoring of the use of personal protective equipment in industrial environments, aiming to reduce accidents caused by lack of PPE. The main solution combines image analysis to detect human form and read barcodes associated with each PPE item, applying a probabilistic prevalence algorithm to at least ten frames captured at intervals of 5 to 60 seconds, executing safety action when the probability of correct use is below a predefined threshold. This computer vision approach relies on human form and barcode detectors, without requiring wearable sensors or GPUs, operating in adverse conditions and dispensing with additional neural networks.As a result, the system supports real-time monitoring of multiple workers with low computational cost. However, IL298700A requires physical labeling of each PPE and does not quantify the fraction of exposed skin area or assess body coverage continuously, limitations that the present invention overcomes by employing pose inference and skin masking to directly measure the visible skin area in each body segment. Petition 870260032128, dated 07 / 04 / 2026, page 6 / 27 5 / 21

[008] Patent CN 113361397 B, filed on June 4, 2021 by Chongqing University, describes a deep learning-based method for detecting facial mask usage conditions in RGB images. The CSPDarkNet-X backbone extracts global features, and the Neck module (SPPNet / PANet / FPN) performs multiscale fusion, culminating in a Prediction network for localization and classification into three categories (no mask, incorrect use, and correct use). The work addresses the limitations of traditional methods, which suffer from low robustness in complex scenarios, high latency, and dependence on sparse and noisy datasets. It proposes data augmentation, CSPDarkNet-Y modules, and Hard-Swish to optimize computational efficiency and system generalization. In experiments with a dataset of 10,000 images (7,000 / 1,000 / 2,000 for training / validation / testing), the model achieved robust PR curves, high mAP, and an average inference time of 0.022 s per image (approx.45 fps), demonstrating viability for real-time applications in surveillance and access control environments. A relevant limitation of the present invention: despite the performance, the approach still requires multiple network modules trained on datasets labeled for each mask usage category and does not continuously quantify the fraction of exposed skin nor operate with a single pose inference combined with algorithmic skin masking, differing from the proposed solution.

[009] In reference CN 114022845 A, filed under number 202111318710.5 on November 9, 2021, a method for real-time detection of insulating gloves on electricians from video captured by a common RGB camera is described. The technical problem addressed is the accurate identification of gloves on the hands of workers in an electrical environment with variations in lighting and ample movement space. The proposed solution combines pedestrian detection by SSD (Single Shot MultiBox Detector) with Inception layers and pose estimation via OpenPose to extract wrist and elbow coordinates, delineate the hand region, convert to HSV space, and calculate the ratio of pixels whose gray intensity meets a predefined glove threshold, generating a binary recognition criterion.As a result, the method achieves high processing efficiency and high detection accuracy without the need for large labeled sets, resulting in improved efficiency and accuracy in glove recognition. Petition 870260032128, dated 07 / 04 / 2026, page 7 / 27 6 / 21 A relevant limitation of the present invention is that this method requires multiple inferences (SSD detector and pose estimation) and relies on fixed thresholds in the HSV space, thus not providing a continuous, weighted metric of skin coverage for quantitative assessment of various PPEs from a single pose inference.

[010] WO 2021 / 250539 A1, application PCT / IB2021 / 054968, filed on June 8, 2021, describes an automated Personal Protective Equipment (PPE) confirmation system based on computer vision. The document addresses the need to replace manual “four-eyes” inspection in environments such as construction sites, mining operations, and healthcare facilities, ensuring that gloves, helmets, goggles, and other PPE are correctly used before granting access. The proposed infrastructure integrates RGB cameras, a memory module containing a predefined list of PPE, and an object recognition model (e.g., YOLOv3 / Bounding-Box CNN) to detect each item in multiple images, comparing the detections with the list and triggering permission or blocking commands according to the result.As an advantage, the method reduces manual intervention and allows for real-time inspection, scaling up PPE verification without the need for physical sensors or special labels, although it does not provide quantitative metrics of accuracy or response time. A relevant limitation of the present invention: the solution relies on multiple neural network inferences trained with specific labeled sets and does not quantify the fraction of exposed skin nor employ unique pose inference, compromising energy efficiency and simplicity in edge devices.

[011] The work of Protik et al. (Real-time Personal Protective Equipment Detection Using YOLOv4 and TensorFlow, 2021) presents a method for real-time identification of masks, face shields, and gloves in video streams to mitigate the transmission of COVID-19. From a dataset of 1392 images annotated into four classes and scaled via Roboflow, the authors fine-tune the YOLOv4 detector in Google Colab and convert the weights to TensorFlow, enabling live counting and record storage. The system achieves a mAP of 79% after 8000 iterations, APs per class of 86.92% (mask), 73.08% (no mask), 84.56% (face shield), and 71.53% (gloves), and operates at 27 fps on a GTX 1660 Ti GPU, with 78% accuracy, 80% recall, and F1 of Petition 870260032128, dated 07 / 04 / 2026, page 8 / 27 7 / 21 However, this approach relies on multiple labeled classes, GPU acceleration, and recurrent deep network inferences on each frame, remaining vulnerable to occlusions and lighting variations, whereas the present invention uses a single pose inference and deterministic skin mask sampling to infer coverage fraction without additional networks or large annotated datasets.

[012] This paper presents the work of Collo et al., “A COVID-19 Safety Monitoring System: Personal Protective Equipment (PPE) Detection using Deep Learning”, submitted in 2022 to DASA (International Conference on Decision Aid Sciences and Applications). The study focuses on identifying complete or incomplete use of medical protective equipment (masks, gloves, goggles, and gowns) in still and video images in a hospital setting during the COVID-19 pandemic. The technical solution is based on the YOLOv3 object detection algorithm, which processes each frame using a single CNN and generates bounding boxes for each PPE, using a set of 400 manually annotated images and training on a Google Colab platform with GPU. mAP values ​​of up to 96.59% were obtained in model evaluation and an average accuracy of 95% in the detection of complete use, while the detection of incomplete use varied between 40% and 82% depending on lighting conditions.Despite demonstrating high performance, the approach relies on large labeled databases, requires multiple inferences per frame and dedicated CNN hardware, and is sensitive to partial occlusions and variations in color and lighting—limitations that the present invention overcomes by using only one pose inference and a deterministic skin mask, eliminating the need for additional training and operating efficiently on a low-power CPU.

[013] Document CN 114758363 A, filed on June 16, 2022, presents a method and system based on a pre-trained neural network for detecting insulating gloves worn by electrical workers. The solution captures images of the hand region using a mobile RGB camera mounted on a circular rail, extracting luminance features to compare them with a glove model and generate gray detection information above a threshold; subsequently, it evaluates color features to identify the use of insulating gloves and, finally, constructs contours by overlaying them with a model of Petition 870260032128, dated 07 / 04 / 2026, page 9 / 27 8 / 21 gloves to verify the correct positioning of the equipment, generating presence, isolation, and compliance results. This approach, however, depends on multiple overlapping thresholds and additional neural network training and inference, implying higher computational cost and the need for a labeled database, in contrast to the present invention which performs PPE verification by single pose inference and mathematical skin model without additional networks.

[014] US patent application publication 2024 / 0185608 A1, filed December 1, 2022, describes a computational system for automatic safety compliance verification on scaffolding, in which image capture devices obtain frames of workers on the structure and a multi-stage image recognition pipeline—with a YOLO architecture trained to detect people, scaffolding tags, and PPE, a ResNet-50 model for tag type classification and region of interest analysis, and a ResNet-3D network for pose estimation—determines the presence and correct positioning of helmets, gloves, and safety harnesses, updating a compliance record. The system applies if / then rules based on scaffolding height to validate PPE and generates non-compliance notifications in near real-time, facilitating integration with networked safety infrastructures.As illustrated in the examples and Table I, PPE detection rates exceeding 90% are reported with adequate processing for continuous monitoring on modest hardware. However, this solution requires multiple deep neural network inferences—each PPE type demands specific labeled datasets and multiple model runs per frame—and does not provide a continuous coverage metric, a limitation that the present invention overcomes by employing a single pose inference combined with deterministic skin masking to directly calculate the percentage fraction of exposed skin in all PPE segments through a unified, computationally inexpensive workflow.

[015] CN 116846059 A, filed on March 7, 2023, presents a risk detection system for electrical power distribution networks using edge computing for drone inspection and fixed monitoring. The document addresses the challenge of processing large volumes of video data collected by cameras in real time. Petition 870260032128, dated 07 / 04 / 2026, page 10 / 27 9 / 21 embedded systems, under varying lighting conditions and with limited computational and energy resources. The technical solution consists of an integrated edge device comprising a lightweight hardware module (image, power, communication), a processing module that extracts feature vectors via CNN, and an algorithmic module employing lightweight networks such as YOLOv5 with SE and U-Net modules, as well as pruning, quantization, and knowledge distillation techniques, communicating with a smart station and a cloud platform. In comparative tests, the system showed an increase in mAP and frames per second, enabling reliable detection of small objects and anomalies in real time (Table 2: mAP vs. FPS).However, this approach requires multiple inferences from trained networks for each type of PPE and large labeled databases, implying high computational cost and maintenance of several models, limitations that the present invention overcomes by using only one pose inference and a deterministic skin mask without the need for additional training.

[016] The reference WANG, Honggang, “Detection of Personal Protective Equipment (PPE) using an Anchor-Free Convolutional Neural Network”, International Journal of Advanced Computer Science and Applications, Vol. 15, No. 2, 2024 presents a method for detecting PPE in industrial environments focused on detecting items such as helmets, gloves, aprons, and safety footwear. The technical problem addressed consists of reconciling high accuracy, non-destructive monitoring, and real-time operation capability without resorting to multiple inferences per frame. The main technical solution is based on an anchor-free object detector based on the YOLOv8 architecture, trained on a set of 3,897 original images tripled to 11,691 using data augmentation techniques (rotation, translation, color adjustment, etc.). The model employs deep convolutions to recognize various classes of PPE and was evaluated in a 70% / 20% / 10% split (training / validation / test).In experiments, it achieved an accuracy of 0.95, a recall of 0.84, and a mAP0.5 of 0.75, demonstrating high identification capacity and a low false positive rate. However, this approach depends on large labeled datasets, requires GPU resources to maintain real-time performance, and does not provide continuous metrics of the exposed skin fraction, limiting its applicability in edge devices and the quantitative assessment of skin coverage, as proposed in [reference needed]. Petition 870260032128, dated 07 / 04 / 2026, page 11 / 27 10 / 21 present invention.

[017] The publication “Detection of personal protective equipment using YOLOv8 architecture on object detection benchmark datasets: a comparative study”, received on May 18, 2023 and accepted on March 15, 2024, by Barlybayev et al. presents a solution for automatic detection of PPE — helmets, vests, and goggles — on construction sites and industrial environments using YOLOv8-based computer vision. The work addresses the challenge of replacing manual compliance verification with a real-time, lightweight, and computationally inexpensive system applicable to edge devices. The technical proposal consists of comparing five versions of YOLOv8 (nano, small, medium, large, and x-large) trained on the CHV and SHEL5K datasets, evaluating classic object detection metrics (precision, recall, mAP50) and inference times.The best performances were achieved with YOLOv8x on the CHV (94.5% accuracy, 86.9% recall, 92.9% mAP50) and with YOLOv8l on the SHEL5K (92.0% accuracy, 82.4% recall, 89.1% mAP50). Despite their robustness, such methods based on bounding-box and deep networks require extensive sets labeled by PPE type and remain vulnerable to partial occlusions and adverse lighting conditions, whereas the present invention eliminates additional training by combining single pose inference with a deterministic skin mask, generating a continuous skin coverage metric without the need for specific labels.

[018] Despite the progress made by PPE detection systems based on “bounding-box” CNNs, pose-guided frameworks with additional classifiers, and solutions that physically mark the equipment, none of the cited works manages to simultaneously: 1. Eliminate reliance on large databases labeled by category and color of PPE; 2. Perform the entire verification with only one network inference (that is, without running people detector + object detector + classifier); 3. Produce a continuous metric of the fraction of exposed skin on each limb — an essential requirement for evaluating sleeves, gloves, boots, and shin guards; Petition 870260032128, dated 07 / 04 / 2026, pages 12 / 27 / 21 4. Operate on edge hardware (< 5W, Cortex-A53 CPU or similar), eliminating the need for a GPU, controlled lighting, or additional labels; 5. Maintain robustness under partial occlusions and variable industrial lighting without manual calibration of thresholds per scenario.

[019] Current methods address, at most, two of these points together; for example, YOLOv5s delivers speed but requires thousands of labels, while CN 114022845 A reduces labeling but requires two networks and manually adjusted HSV thresholds for gloves only.

[020] The invention overcomes the gaps in the prior art because it performs all verification in a single 2D pose inference – obtained by MediaPipe Lite or any equivalent estimator that offers greater accuracy or speed – combined with a purely algorithmic skin mask, dispensing with the cascade of two or three networks (person detector, object detector, classifier) ​​typical of YOLO + Pose or SSD + OpenPose approaches. Furthermore, it uses dense sampling weighted by bone visibility to produce a continuous C% skin coverage index, where previous methods only performed binary checks at a few points or were restricted to the head and torso.As a result, it eliminates the need for labeled datasets for each color and type of PPE, requiring only the skeleton generated by the pose engine and simple mathematical rules; it operates with a common RGB camera, without the labels or special inks required by barcode-based solutions; and it provides a quantitative coverage value per segment, allowing for gradual alarms (e.g., when C < 80%) instead of the mere "has / does not have" of bounding-box detectors. In practice, these characteristics translate into low processing power (<5W on a Cortex-A53 CPU, such as single-board computers from the Raspberry Pi family, among others), deployment without retraining, immediate scalability to new PPE, robustness to partial occlusions thanks to the visibility factor, and automatic generation of percentage reports compatible with the requirements of NR-06 / NR-12, all without being tied to a specific library implementation. Petition 870260032128, dated 07 / 04 / 2026, page 13 / 27 12 / 21

[021] In summary, the invention solves the remaining technical problem - automatic, accurate and low-cost verification of effective skin coverage by various PPEs - by a path not foreseen by the state of the art, combining only a pose inference with light mathematical processing, without additional networks, without training labels and with continuous skin exposure metrics applicable to any part of the body.

[022] It is important to emphasize that the solution now proposed is not a trivial derivation for a technician versed in computer vision applied to occupational safety. The scope of the state of the art reported here — involving more than ten initiatives from universities, industrial groups, and patent holders that tested bounding-box-based approaches with deep CNNs, pose-guided anchors with additional classifiers, printed barcodes, and local chromatic thresholds — demonstrates the sector's ongoing efforts to overcome the limitations of extensive labeling, computational cost, and low robustness to occlusions. Even after such efforts, no document combines a single 2D pose inference with a deterministic skin mask and the dense, visibility-weighted calculation of the fraction of skin exposed per body segment, capable of generating a quantitative coverage index (C%) in real time and on edge hardware < 5W.Thus, from a technical point of view, the inventive route followed – which eliminates additional netting, supervised training, and physical markings on PPE – was not available to the specialist, reinforcing the inventive nature of the present proposal. Brief description of the invention

[023] The present invention proposes a computer-implemented method and a low-power embedded system designed to verify, in real time, compliance in the use of Personal Protective Equipment (PPE) from RGB images. The preferred operating flow comprises: (i) acquisition of a frame by RGB camera; (ii) unique 2D pose estimation of the user using a lightweight pose-estimation model (e.g., MediaPipe Lite, but not limited to this model) that produces 33 anatomical landmarks; (iii) deterministic generation of a skin mask by chromaticity, using linear combination of RGB channels followed by thresholds to classify each pixel as Petition 870260032128, dated 07 / 04 / 2026, page 14 / 27 13 / 21 skin or non-skin; (iv) dense sampling weighted by bone visibility (connections between landmarks) to compute, in each body segment, the fraction C% of pixels covered by PPE; and (v) issuing a decision / alarm when C falls below a configurable threshold.

[024] The method eliminates the need for multiple deep networks and large labeled databases, as it derives all skin cover judgment from a single pose inference combined with simple mathematical rules. In the preferred but not limited embodiment, for example, Python routines perform skin mask calculation, glove and boot verification on specific hand and foot landmarks, and serialization of results into JSON for integration with industrial interfaces.

[025] It is understood, however, that the form described above is preferred and not limiting: the skin detection algorithm may employ other color spaces (YCbCr, HSV), pose estimation may be obtained by any network that provides compatible landmarks (BlazePose, OpenPose Lite, or future more precise techniques), and the camera may be replaced by multispectral, thermal, or depth sensors. Such variations, as well as adjustments to threshold, resolution, or processing device (CPU, GPU, NPU), are considered to be included in the spirit of the invention and do not preclude the claimed protection. Brief description of the drawings

[026] The invention will be described below with reference to typical embodiments thereof and also with reference to the attached drawings, in which: - Figure 1 presents the flowchart of the PPE Verification System, highlighting the RGB camera (200), the Processing Unit (300) with its functions of 2D pose estimation (301), skin mask generation (303), weighted dense sampling (305) and skin coverage index calculation (307), in addition to the program memory and logs (400) and the alarm interface (500). Figure 2 illustrates the flow of information between the main modules: capture Petition 870260032128, dated 07 / 04 / 2026, page 15 / 27 14 / 21 by RGB camera (200), 2D pose inference (301), skin mask generation (303), weighted dense sampling (305), calculation of C-index % coverage (307) and presentation of decision / alert on the interface (500). - Figure 3 shows the result of applying the skin mask to a real image (600), highlighting uncovered regions at references 601 (face), 602 (left hand) and 603 (left leg), coverage detected at points 604 (left foot) and 605 (right hand) and occlusion of the right forearm (606). - Figure 4 presents a human silhouette (501) with circular markings on exemplary body segments (501, 503, 505, 507) and text boxes (502, 504, 506, 508) demonstrating the annotation of position, coverage index and visibility, in addition to illustrating the occlusion effect by a rectangular object (509). Description of the invention

[027] The present invention relates to a computer-implemented method and a system (100) designed to verify, in (near) real-time, compliance in the use of Personal Protective Equipment (PPE) by humans in an industrial, hospital or similar environment. The system uses only an RGB image capture (200), a single 2D pose inference (301) and light mathematical operations to calculate, for each body segment, the percentage fraction C% of skin covered by PPE, automatically generating an alert or recording on the interface (500) when C falls below a configurable threshold. The entire operational flow is shown, preferably but not limited to, in the flowchart of Figure 1 (boxes, arrows indicating flows and reference codes as described in the elements of the figure).

[028] Acquisition and pre-processing—A common RGB camera (200) generates a frame Ie RhxWx3(1) which is forwarded, via flow (210), to the Processing Unit (300). The unit may be embedded in the camera, in a local microcontroller or in a Petition 870260032128, dated 07 / 04 / 2026, page 16 / 27 15 / 21 cloud instance, without affecting the spirit of the invention; the data link (210) can be wired or wireless (Wi-Fi, 5G, industrial Ethernet).

[029] 2D pose inference—Module (301) runs a lightweight 2D pose estimator (e.g., MediaPipe Pose Landmarker Lite), producing {pk}k=0, pk =(xk,yk), vkG[0, 1](2) of normalized landmarks (human body reference points, such as joints between human body limbs, e.g., wrists, elbows, knees, etc.) and a visibility for each point. Future techniques with greater accuracy or speed (BlazePose, OpenPose Lite, HRNet Tiny, etc.) may be adopted without departing from the intended scope.

[030] Deterministic Skin Mask Generation—Module (303) produces a binary mask {1, if Δ(x, y) g [', u] (3) 0, otherwise, where Δ = I - Imax, I = I(x,y) [0.30, 0.59, 0.14]T and Imax = max{G,B} in each pixel. In the preferred embodiment the thresholds are ' = 0.03 and u = 0.12, extracted empirically with grid-search on, for example, 2048 manually annotated skin and non-skin samples using, for example, the CPPE (Construction Personal Protective Equipment) dataset. The same logic works on other chromatic bases (YCbCr, HSV) or on multispectral sensors; such variations are considered covered by the spirit of the present invention.

[031] Weighted dense sampling—Between each pair of landmarks that define a bone (e.g., left arm: (11, 13); right forearm: (14, 16); right thigh: (24, 26) etc.), the module (305) discretizes the segment into n = H|pi -pj ||2] points equally Petition 870260032128, dated 07 / 04 / 2026, page 17 / 27 16 / 21 spaced (xs,ys) and calculates / 1ncij = iooçi - ^^m (-^) s=1 v ij = min {vi,vj}, (4) where cij is the percentage of skin covered and Vij is the minimum visibility of the bone in question. Only segments with Vij > 0.9 are considered reliable (this is a value that can be calibrated).

[032] Coverage index calculation—The module (307) aggregates the cij values ​​of all reliable segments to produce, by member,Cmember— Σ(i,j)GSmember vijcij (ij)gSmember ^ij(5) where 5member is the set of bones belonging to the analyzed member. The final result is a vector C — [C_arm_E,C_forearm_E, ,C_leg_D](6)

[033] Decision / Alert—The interface (500) receives C through the flow (310) and executes business logic (e.g., trigger visual / audible alarm or log entry) when the member coverage index falls below the threshold τ, where τ is defined by the security manager. Figure 2 shows the chaining of the internal modules (301-307) until the decision / alert trigger (500); Figure 3 presents a real example of a skin mask (600) with uncovered regions (601, 602, 603, and 606) while also showing covered regions (604 and 605, left shoe and right glove, respectively); and Figure 4 illustrates the information overlay on the graphical interface (circles 501, 503, 505, 507 and boxes 502, 504, 506, 508).

[034] Extension by PPE color analysis—In addition to detecting whether the skin is covered, the system can assess whether the covering cord meets specific requirements (e.g., blue gloves and white apron). For each member with Cmember> τmin, the average is calculated. Petition 870260032128, dated 07 / 04 / 2026, page 18 / 27 17 / 21 color h in HSV in the region not classified as skin: ih = íõí Σ hsv(i())· {(xy) । m(x,y) = o}· (7)1Ω1(x,y) e Ω A preferred range [hmin, hmax] is defined by PPE color. The module compares h to the corresponding range; if outside the range, the alert indicates "covered with inappropriate color". That is, it only evaluates colors that are within the space delimited by body parts.

[035] Implementation and industrial requirements—A reference code, written in Python 3.10, demonstrated execution on a Cortex-A53 CPU with less than 5 W of power, processing between 5 and 10 frames per second at a resolution of 640 by 480. Storage (400) holds both the program and the C and alarm registers, enabling later auditing; flows (410) and (420) in Figure 1 illustrate firmware loading and sending reports to the operator panel.

[036] Variations within the spirit of the invention—The skin detection algorithm may employ other linear mixture models, color transformations, or look-up table classification; pose estimation may be 3D; the C% metric may incorporate temporal weighting (moving window) or infer tissue quality via Fourier texture. Similarly, the decision may integrate lightweight machine learning (decision tree) trained in C to automatically calibrate τ by environment. All these variations maintain the single inventive concept: using a single pose inference combined with a deterministic skin mask to obtain a quantitative and segmented skin coverage index in real time. The state-of-the-art material demonstrated here shows efforts towards achieving such a solution, but these efforts are still far from being as effective as the invention presented here.

[037] Technical advantages—(i) Eliminates the need for thousands of labels per color / type of PPE; (ii) reduces energy consumption because it performs only a lightweight network and arithmetic operations; (iii) provides continuous metrics, allowing for gradual alarms and statistical reports; (iv) is robust to occlusions by incorporating vij visibility; (v) operates on low-cost hardware or in the cloud. Petition 870260032128, dated 07 / 04 / 2026, page 19 / 27 18 / 21

[038] In this concluding section, all parts of the System (100) of Figure 1 are listed, including the RGB camera (200), the image stream (210) to the Processing Unit (300) and its internal modules for 2D pose estimation (301), skin mask generation (303), weighted dense sampling (305) and calculation of the C % coverage index (307), as well as the program memory and logs (400), the program loading streams (410) and results presentation (420) and the alarm and recording interface (500).

[039] The detailed operation of this pipeline is illustrated in Figure 2, where the frame captured by the RGB camera (200) goes through the link (210) to the 2D pose inference (301), has its landmarks transmitted by the flow (302) to the mask generation (303), whose result is sent by the flow (304) to the weighted dense sampling (305), whose coefficients pass through the flow (306) to the calculation of the coverage index (307) and, finally, the value of C % is forwarded via (310) to the decision / alert process on the interface (500).

[040] The practical result of the algorithm is exemplified in Figure 3, where a real image was converted into a masked image (600), revealing the absence of PPE on the face (601), left hand (602), left leg (603) and right forearm (606), while detecting adequate coverage on the left foot (604) and right hand (605) (avoiding exhaustion to cite all other detected coverings that can be observed in Figure 3), proving the accuracy in identifying covered and uncovered regions. It can be verified that other parts of the image were classified as skin or non-skin, however, such parts are outside the scope of the human body and do not affect the accuracy of the model, since the entire process is limited to the space detected by the human pose.

[041] To communicate these results intuitively, Figure 4 displays a human silhouette (501) with representative circles on the limbs — exemplified by markers (501), (503), (505) and (507) — complemented by text boxes (502), (504), (506) and (508) that inform the position, coverage index and visibility of each segment, in addition to illustrating the occlusion effect caused by the rectangle (509), whose impact on the visibility of the right leg demonstrates the robustness of the method in real-world scenarios. Petition 870260032128, dated 07 / 04 / 2026, page 20 / 27 19 / 21 partial vision blockage.

[042] Thus, it is concluded that all the numbered elements in Figures 1 to 4 are interconnected to ensure the complete operation of the method and the system, from the acquisition and sequential processing of each step to the clear presentation of the results to the user and security manager, consolidating the proposal described. Examples of embodiments of the invention

[043] This section describes feasible embodiments of the proposed method and system, encompassing both edge (fully embedded) configurations and hybrid architectures with cloud processing. In all cases, the MediaPipe Pose Landmarker Lite (Lightweight network based on MobileNetV2 with 33 landmarks that achieves an average PDJ of 87.0% (Percentage of Detected Joints) in tracking mode for 256 x 256 input, as evaluated in 14 geographic sub-regions) is used, preferably but not exclusively, for single pose estimation, combined with a deterministic skin classifier derived from luminance differences (Linear transformation I = 0.30 R + 0.59 G + 0.14 B and error Δ = I - max {G, B}; pixels with 0.03 < Δ < 0.12 are labeled as skin, reducing the dimensionality from 3 to 1 and allowing real-time classification).

[044] Minimum embedded configuration. • Hardware: Raspberry Pi 3 Model B+ (ARM Cortex-A53, 1 GB RAM) coupled with an 8 MP CSI2 camera; overall power consumption <5W. • Software: Linux Bullseye 32-bit, Python 3.11, tflite_runtime, mediapipe-lite. Pose inference « 44 FPS via XNNPack; full pipeline (pose + skin mask + C% calculation) renders 5-10 FPS at 640x480 px. • Operation: Each frame generates a C vector of skin coverage per limb; local decision triggers a red LED, 85dB buzzer, or blocking relay whenever Cmember < τ (typically τ = 80). JSON logs (timestamp, C, state) are published via MQTT TLS 1.2 to the AWS IoT Core broker. Petition 870260032128, dated 07 / 04 / 2026, pp. 21 / 27 20 / 21

[045] Hybrid edge-cloud architecture. • Capture: RTSP IP cameras (H.264, 15 FPS, 720p) stream to an EC2 t4g.micro instance (Graviton2) running a Docker container with the same pipeline. • Scalability: Auto-scaling group increases to g4dn.xiarge when > 30 cameras; metrics are aggregated by AWS Kinesis Data Firehose in S3 + Glue Catalog; dashboards in Amazon QuickSight show C% time curve by sector, non-conformity heat map, and weekly PDF reports. • Interfaces: REST API (API Gateway + Lambda) allows real-time querying; MQTT port maintains backward compatibility with Raspberry Pi field nodes.

[046] Optional mobile application. An Android 13 smartphone (min. Snapdragon 778G) runs the lite model locally (» 49 FPS GPU), uses the same skin mask and sends only alarms and C to the serverless backend. This enables PPE auditing on inspection visits without fixed infrastructure.

[047] Table 1 - Accuracy metrics and validation test - To corroborate effectiveness, a test was performed on a subset of 1000 images of the CPPE (Construction Personal Protective Equipment). Non-conformity detection was defined as Cmember < 80%, where: Item PPE Accuracy (%) Recall (%) Ft Helmet (head) 96.8 ±0.5 95.9 0.963 Glove (right hand) 94.1 ± 0.7 92.4 0.933 Glove (left hand) 93.6 ±0.8 91.8 0.926 Boot (right / left foot) 95.2 ±0.6 93.7 0.944 PDJ (Percentage of Detected Joints) - also referred to as PDJ@0.2 or PCK@0.2 - considers a landmark correctly located if its distance (x,y) normalized by the torso diameter is < 20%. The lite model achieved a PDJ of 87.0%, while the full model reached 91.8% and the heavy model 94.2%. Petition 870260032128, dated 07 / 04 / 2026, pp. 22 / 27 / 21

[048] Dynamic threshold: in environments with variable illumination (plus or minus twenty-five percent lux), the τ value is adjusted in real time by the median of the ten percent brightest pixels in the V channel in HSV; this dynamic threshold stabilizes the skin mask, maintaining variation of less than one point two percent in overall accuracy.

[049] Summary of technical advantages - The concrete examples illustrate that the system: 1. Performs all verification in a single network inference + elementary algebra, without additional CNNs. 2. Operates on hardware < 5W or, optionally, integrates with AWS back-end with real-time ingestion. 3. Achieves accuracy rates > 94% in most PPEs with variation < 1% between edge and cloud. 4. It tolerates partial occlusions thanks to the visibility factor vknos the sampled bones. 5. Generates C as a continuous metric (versus binary approaches), allowing for gradual alarms and historical reports.

[050] These implementation methods result in immediate applicability in production lines, hospitals, industrial kitchens, industry in general and construction sites, providing autonomous monitoring of PPE compliance and transparent integration with safety management systems.

[051] Those skilled in the art will appreciate the knowledge presented here and will be able to reproduce the invention in the forms presented and in other variants, covered within the scope of the appended claims. Petition 870260032128, dated 07 / 04 / 2026, pp. 23 / 27

Claims

1 / 3 CLAIMS 1. COMPUTER-IMPLEMENTED METHOD FOR REAL-TIME VERIFICATION OF HUMAN USER SKIN COVERAGE BY PERSONAL PROTECTIVE EQUIPMENT (PPE), characterized by comprising, for each image frame captured by an RGB camera (200) and transmitted by (210), the steps of: (i) executing in the processing unit (300) a single two-dimensional pose estimation neural network inference (301), with between 1.0 x 105 and 3.0 x 106 parameters, preferred value 2.0 x 106, and execution time between 5 ms and 50 ms, preferred value 25 ms, in an ARMclass processor at 1 GHz, producing from 10 to 700 landmarks with two-dimensional coordinates and respective visibilities, preferred value 33;(ii) generate binary skin mask (303) by calculating a chromaticity value obtained by linear combination of RGB channels with coefficients a, b and c, with coefficient a between 0.25 and 0.35, preferred value 0.30, coefficient b between 0.50 and 0.65, preferred value 0.59, and coefficient c between 0.10 and 0.20, preferred value 0.14, followed by subtracting the larger value between the G and B channels, and labeling as skin the pixels whose chromaticity value is between a lower threshold between 0.02 and 0.05, preferred value 0.03, and an upper threshold between 0.10 and 0.15, preferred value 0.12;(iii) discretize each defined body segment between landmarks whose minimum visibility is between 0.5 and 1.0, preferred value 0.9, into a number n of points between 10 and 200, preferably corresponding to the ceiling of the Euclidean distance between the ends of the segment, and calculate for each segment a percentage skin coverage index corresponding to one hundred times one minus the average of the mask values ​​at the sampled points; (iv) compose a coverage vector (307) of visibility-weighted indices and compare each element to a configurable threshold situated between 50 and 98, preferred value 80; and (v) emit a conformity or non-conformity signal via an alarm interface (500) depending on whether all or at least one element of the coverage vector exceeds or falls below the aforementioned configurable threshold, respectively; 2. Method according to claim 1, characterized by, in step (ii), implementing complementary skin detection (303) in YCbCr or HSV space. Petition 870260032128, dated 07 / 04 / 2026, p. 24 / 27 2 / 3 3. A method, according to any one of claims 1 or 2, characterized by dynamically adjusting said configurable limit as a function of the median of the ten percent of pixels with the highest luminance in the V channel of the HSV space of the current frame.

4. A method, according to any one of claims 1 to 3, characterized by evaluating, for each segment whose coverage exceeds the aforementioned configurable threshold, whether the average color of the region not classified as skin belongs to a range previously associated with the type of PPE, recording color non-conformity when outside this range.

5. Method, according to any one of claims 1 to 4, characterized in that it is executed on an edge processing unit (300) with power consumption not exceeding 5 W and a 64-bit processor.

6. A method, according to any one of claims 1 to 5, characterized by publishing, via real-time messaging protocol (420), records containing a timestamp and cover vector encoded in an interoperable structured format.

7. Method, according to any one of claims 1 to 6, characterized by producing a mask (600), in which covered (604, 605) and uncovered (601, 602, 603 and 606) regions are used in the calculation of the coverage vector (307) 8. Method, according to any one of claims 1 to 7, characterized by producing annotations relating to regions of the human body where PPEs are applied (501, 503, 505 and 507), with information presented about such regions (502, 504, 506, 508), including the identification of occlusions (509) Petition 870260032128, dated 07 / 04 / 2026, page 25 / 27 3 / 3 9. PERSONAL PROTECTIVE EQUIPMENT CONFORMITY VERIFICATION SYSTEM (100), characterized by comprising: (a) at least one RGB camera (200) facing the inspection zone; (b) a processing unit (300) programmed to execute the method according to any of claims 1 to 6; (c) non-volatile memory (400) for storing the program and operational records; and (d) an alarm interface (500) capable of emitting a visual or audible signal or a lockout command when non-conformity is detected.

10. System according to claim 9, characterized in that the processing unit is a single-board computer with a 64-bit multi-core processor.

11. System according to any one of claims 9 or 10, characterized in that the alarm interface (500) comprises a dry contact relay with a switching capacity of 250 VAC / 5 A.

12. System, according to any one of claims 9 to 11, characterized by including a wireless communication module compatible with IEEE 802.11 b / g / n or LTE-Cat 4 to transmit the records referred to in claim 6 to a remote server.

13. COMPUTER-READABLE NON-TRANSIENTIAL MEMORY STORAGE, characterized by containing instructions that, when executed by a processing unit, cause the system to execute the method according to any of claims 1 to 8.

14. Memory support, according to claim 13, characterized in that the instructions are encoded in intermediate bytecode interpretable by an embedded virtual machine (400) Petition 870260032128, dated 07 / 04 / 2026, pp. 26 / 27

Images