Çalışma ortamında kas-iskelet sistemi bozuklukları işle ilgili en yaygın hastalıklar arasındadır. Genel olarak, Kas İskelet Sistemi Rahatsızlıkları (KİSR) tek bir neden ve sonuca bağlı olmamaktadır. Tersine, fiziksel, biyomekanik, psiko-sosyal ve bireysel risk faktörleri ile ortak bir şekilde meydana gelmektedir. KİSR genellikle hızlıca etki göstermemektedir. Ancak uzun vadede KİSR bağlantılı hastalıklar birkaç belirgin şikâyet ile ortaya çıkmaktadır. Sırt ve boyun ağrıları en sık görülen belirtiler arasındadır. Kas yaralanmaları, yorucu veya tekrarlayan faaliyetlerden kaynaklanabilmektedir. Fiziksel olarak zorlu çalışma koşullarının yanı sıra psikososyal stres de vücuttaki gerilimin nedeni olabilmektedir. İnsanların yaşadıkları tüm sağlık problemlerine ek olarak, iş kazaları ve işle ilgili hastalıklar, şirketler ve bir bütün olarak ekonomi için hem doğrudan hem de dolaylı maliyetlere neden olmaktadır. Uygunsuz duruş, tekrarlayan hareketler ve tek yönde ağırlık baskısı gibi etkenlerle oluşan rahatsızlıklar , çalışanlar için sağlıklı yaşam yıllarının kaybı gibi ise uzun vadeli ve dolaylı maliyetleri oluşturmaktadır.Uzun vadede KİSR, çalışanlarda kalıcı rahatsızlıklar , felç veya kısmı ve tam işgörmezlik riskini beraberinde getirmektedir. Çalışmanın amacı, iş yerlerinde ergonomik risk değerlendirmelerini (ERD) web tabanlı, fotoğraf ve videolar üzerinden çalışan ERD analizi yaparak KİSR rahatsızlıklarına erken aşamada önlem alınmasını ve doğru analiz yapılmasını sağlamaktır. Çalışmada, bir web platformu üzerinde bilgisayarlı görü destekli ergonomik risk değerlendirme yazılımı yapılandırılmış ve kullanıcı arayüzü ile çalışma sahasında kullanılması sağlanmıştır. Web platformu, kullanıcı tarafından yüklenen fotoğraflar ile Hızlı Tüm Vücut Değerlendirmesi (Rapid Entire Body Assessment-REBA), Hızlı Tüm Vücut Değerlendirmesi (Rapid Upper Limb Assessment-RULA) ve Ovako Çalışma Duruş Analiz Sistemi (Ovako Work Posture Analysis System-OWAS) metotları için aynı anda sonuç hesaplayıp çıktı üretebilmektedir. Çalışmada makine öğrenmesi ve bilgisayarlı görü tabanları kullanılmıştır. Poz tahmini aşaması için MediaPipe kütüphanesinde poz tahmini teknolojisi ile vücut açılarının analizleri gerçekleştirilmiştir. Yapay zekâ modeli olarak Önerilen platformun validasyonu amacıyla, poz tahmini algoritmalarında kullanılan Anahtar Nokta Benzerliği (Object Keypoint Similarity-OKS) testi uygulanmıştır. Test, 32 vücut anahtar noktasının her birine uygulanmış ve genel ortalamada %92 doğruluk oranı elde edilmiştir. Diğer test sürecinde ise ERD metotlarında kullanılmak üzere ölçülen vücut eklem açılarının doğruluğu hesaplanmıştır. Program tarafından ölçülen 32 vücut eklemi açısının her biri gerçek olarak baz alınan açılarla karşılaştırılmış ve ortalamada 7,7°'lik Kök Ortalama Karesel Hata (Root Mean Sqaured Error-RMSE) xxii değeri elde edilmiştir. Elde edilen RMSE değeri ve OKS sonucu güncel literatür ile kıyaslandığında değerlerin tutarlı olduğu belirlenmiştir. Çalışmada ergonomik risk değelerlendirme uygulamalarının testi yapılmıştır. RMSE değeri 0,52 ve sonucu %95 olarak ölçülmüştür. Tutarlılık seviyesinin yüksek olduğu belirlenmiştir. Bu çalışmada, nesne algılama ve sınıflandırma problemlerinin çözümünde etkili bir yöntem olarak göze çarpan Bölgesel Evrişimli Sinir Ağı ( Region-based Convolutional Neural Network-RCNN ) algoritması benimsenmiştir. RCNN, görüntü içerisindeki önceden belirlenmiş bölgelerin (region proposals) özel bir evrişim sinir ağı mimarisi aracılığıyla dikkate alındığı ve bu bölgelerin sonrasında ayrı ayrı işlendiği inovatif bir yaklaşımdır. Algoritmanın başarımı, nesnelerin tespiti ve doğru sınıflandırılmasında gösterdiği yüksek hassasiyet ve doğrulukla kanıtlanmıştır. RCNN yöntemi kapsamlı bir şekilde değerlendirilerek, çeşitli görsel veri setleri üzerindeki performansı ayrıntılı bir şekilde incelenmiştir. Elde edilen sonuçlar, RCNN'in nesne algılama alanında önemli bir çözüm olarak değerlendirilmesini destekleyici niteliktedir.
Musculoskeletal disorders (MSD) are among the most common work-related illnesses. Most work-related MSRs develop over a long period of time. In general, MSD does not develop due to a single cause and result. Instead, it occurs in association with various risk factors. These are physical, biomechanical, psycho-social and individual factors. Physical and biomechanical risk factors can be listed as bending and turning the trunk, especially during the carrying of loads, repetitive or difficult movements, awkward and static postures, and sitting or standing in the same position for a long time. MSD usually does not occur quickly or show no effect. However, in the long term, MSD-related diseases emerge with few obvious complaints. Back and neck pain are among the most common symptoms. Muscle injuries can result from strenuous or repetitive activities. Joint diseases can occur as a result of wear and tear. In addition to physically demanding working conditions, psychosocial stress can also be the cause of tension in the body. Often the psychological link to the complaints is not clear. Health-impairing factors have an indirect effect on the musculoskeletal system through stress. Work-related stress often occurs in combination. In addition to all the health problems people experience, occupational accidents and work-related diseases cause both direct and indirect costs for companies, those affected, and the economy as a whole. Direct costs account for 10% of the total cost burden of work-related diseases. It accrues to those who pay for the health system in the form of treatments and compensation payments immediately after diagnosis. Determining the total costs of work-related accidents and illnesses takes a holistic approach and includes both long-term and indirect costs, such as loss of healthy life years for employees. Employees are exposed to long-term damage to their health as a result of work accidents or diseases that reduce their lifetime quality of life. Taking timely and necessary measures for MSD provides a significant benefit in the fight against permanent ailments. One of these measures is to determine the ergonomic risks in the working environment and to take precautions accordingly. By identifying and assessing ergonomic hazards, employers can take steps to reduce the risk of MSD and improve overall employee well-being. There are many ergonomic risk assessment (ERA) methods available, from simple checklists to more complex assessments with detailed measurements and detailed reporting analysis. Regardless of the method used, the aim is to identify potential hazards in all of them and take action to reduce them. However, the common problems of these methods are that when applied manually, they take a long time, they are challenging, and they can create inconsistency because they contain relative situations that can vary from expert to expert. In this study, the Ergonomic risk assessment process was structured on the web platform and used by an expert in the work area with the user interface. The web platform can calculate results and produce output,Rapid Entire Body Assessment (REBA), Rapid Upper Limb Assessment ( RULA), Ovako Work Posture Analysis System (OWAS) REBA, RULA and OWAS methods with the photos uploaded by the user at the same time. REBA evaluates the ergonomic risk on the whole body extremity, RULA on the upper body subject, OWAS evaluates the general working moment. Machine learning and computer vision bases were used in the study. On the server side, analysis of body angles was performed with pose estimation technology in the MediaPipe library. In the study, Pose Estimation, which is one of the artificial neural network techniques, was chosen. Body posture estimation is the process of determining the position and orientation of objects or individuals in an image or video. It is a critical aspect of computer vision and is used in a wide variety of applications, including augmented reality, robotics, human-computer interaction and sports analysis.Model-based body posture estimation is a technique for determining the position and orientation of objects or people in an image or video by comparing the observed 2D image with a predefined 3D model of the object or person. This approach is based on computer vision and geometry and is widely used in a variety of applications such as augmented reality, robotics and human-computer interaction.The first step is to identify key points in the 2D image, such as joints in the human body or the corners of an object. These key points serve as landmarks to align the 2D image and 3D model.The next step is to align the 3D model with the 2D image. This is usually done with a model in which the 3D coordinates of the key points and their corresponding 2D image coordinates are used to calculate the pose of the object or person in the image. The body coordinates obtained in the system algorithm are passed through a series of calculations. These obtained data are transferred to the common detection result pool as in Figure 1. After the pool, ERA methods proceed by taking the necessary angles or questions from here. In this way, after a single calculation operation, a common flow is performed for all three methods. Time and resource savings are achieved. ERA methods generally require a 3D view of the person. For this reason, the system considers the results obtained from the side and front visual inputs separately. In Figure 1, the cosine and right triangle theorems have been applied to the results obtained with the side view of the staff. Body angles were obtained entirely from the side view. The choice of which method will be found with which angle measurement has been made by testing with real life measurements. For example, the cosine theorem is used for leg angle. MediaPipe offers a number of pre-trained models for specific tasks, including face recognition, hand tracking, and pose estimation. This can offer time and labor savings compared to zero training models. The model is compatible with a variety of platforms including mobile devices, web browsers and desktop computers. This simplifies the deployment of models in different environments and devices and allows simultaneous access from many platforms. It allows users to create pipelines that combine different pre-processing, inference, and post-processing components, making it easy to experiment with different architectures and configurations. MediaPipe is optimized for on-device rendering; This means that models can be run directly on a user's device without the need for cloud-based processing. In other words, user can use Mediapipe without internet connection and powerful hardware.As the artificial intelligence model, Tensorflow's pose estimation model was used and it was ensured that they could work asynchronously on the server side. Max-Planck-Institut für Informatik (MPII) Multi-person human posture dataset was used for the test of the study. The speed, accuracy and stability tests of the system were carried out with 1586 photographs of the workers at the work sites. The accuracy of the analysis was tested with a statistical tool called Object Key Point Similarity. The human coordinates given in the MPII data set in the 1586 photographs given as input to the platform were accepted as true and compared with the analysis results obtained from the web platform used in the study. Object Key Point Similarity test was used for comparison. For the validation of the proposed platform, the Object Keypoint Similarity (OKS) test, which is used in pose estimation algorithms, was applied. The test was applied to each of the 32 body key points, with an overall average accuracy of 92%. In the other test process, the accuracy of the measured body joint angles was calculated to be used in ERA methods. Each of the 13 body joint angles was compared with the actual baseline angles and an average Root Mean Square Error (RMSE) of 7.7° was obtained. When the RMSE value and OKS result obtained were compared with the current literature, it was determined that the values were consistent. In the study, ergonomic risk assessment applications were tested. The RMSE value was 0.52 and the result was 95%. It was determined that the level of consistency was high. In this study, the Region-based Convolutional Neural Network (RCNN) algorithm, which has proven to be an effective approach for solving object detection and classification problems, was employed. RCNN adopts an innovative methodology where pre-determined regions of interest (region proposals) within an image are selectively considered and subsequently processed separately through a specialized convolutional neural network architecture. The algorithm's performance has been validated by demonstrating high precision and accuracy in object detection and proper classification. This paper comprehensively evaluates the RCNN method, thoroughly examining its performance on various visual datasets. The obtained results underscore RCNN's significance as a prominent solution in the field of object detection.
Except where otherwise noted, this item's license is described as http://creativecommons.org/licenses/by/4.0/
