Search

Inability of safety managers to timely detect misuse of Personal protective equipment (PPE) causes a number of injuries and financial losses. Considering sizes of industry halls and number of workers, there is an increasing demand for computerized tools that could help companies to enhance the implementation of strictinging workplace safety standards. As a solution, we propose a procedure that: 1) reduces the problem of PPE compliance to the binary classification, and 2) enables compliance of arbitrary type and number of PPE that could be mounted on various body parts. To prove this hypothesis, we studied 18 different PPE types used across various industries for protecting 5 physiological body parts/functions. The HigherHRNet pose estimator was used for defining the PPE regions of interest, while six different image classification architectures were assessed for the compliance/classification of the considered regions. All classifiers were pretrained on the ImageNet data set and fine-tuned using the dedicated data set developed during this study. Top-performing models were MobileNetV2, Dense-Net, and ResNet, while the MobileNetV2 was recommended as the most optimal choice considering its lower computation demands. Compared to previous studies, the proposed approach demonstrated competing performances with unique ability to be easily adopted for performing compliance of various PPE by slight editing of the predefined lists of PPE types and corresponding body parts. Considering the present data/privacy/ computational constraints, the procedure is recommended as suited for the digitalization of PPE compliance in: 1) self-check points, and 2) safety-critical workplaces.

Non-ergonomic execution of repetitive physical tasks represents a major cause of work-related musculoskeletal disorders (WMSD). This study was focused on the pushing and pulling (P&P) of an industrial handcart (which is a generic physical task present across many industries), with the aim to investigate the dependence of P&P execution on the operators’ psychological status and the presence of pain syndromes of the upper limbs and spine. The developed acquisition system integrated two three-axis force sensors (placed on the left and right arm) and six electromyography (EMG) electrodes (placed on the chest, back, and hand flexor muscles). The conducted experiment involved two groups of participants (with and without increased psychological scores and pain syndromes). Ten force parameters (for both left and right side), one EMG parameter (for three different muscles, both left and right side), and two time-domain parameters were extracted from the acquired signals. Data analysis showed intergroup differences in the examined parameters, especially in force integral values and EMG mean absolute values. To the best of our knowledge, this is the first study that evaluated the composite effects of pain syndromes, spine mobility, and psychological status of the participants on the execution of P&P tasks—concluding that they have a significant impact on the P&P task execution and potentially on the risk of WMSD. The future work will be directed towards the development of a personalized risk assessment system by considering more muscle groups, supplementary data derived from operators’ poses (extracted with computer vision algorithms), and cognitive parameters (extracted with EEG sensors).

The compliance of industrial personal protective equipment (PPE) still represents a challenging problem considering size of industrial halls and number of employees that operate within them. Since there is a high variability of PPE types/designs that could be used for protecting various body parts and physiological functions, this study was focused on assessing the use of computer vision algorithms to automate the compliance of head-mounted PPE. As a solution, we propose a pipeline that couples the head ROI estimation with the PPE detection. Compared to alternative approaches, it excludes false positive cases while it largely speeds up data collection and labeling. A comprehensive dataset was created by merging public datasets PictorPPE and Roboflow with author’s collected images, containing twelve different types of PPE was used for the development and assessment of three deep learning architectures (Faster R-CNN, MobileNetV2-SSD and YOLOv5)—which in literature were studied only separately. The obtained results indicated that various deep learning architectures reached different performances for the compliance of various PPE types—while the YOLOv5 slightly outperformed considered alternatives (precision 0.920 ± 0.147, and recall 0.611 ± 0.287). It is concluded that further studies on the topic should invest more effort into assessing various deep learning architectures in order to objectively find the optimal ones for the compliance of a particular PPE type. Considering the present technological and data privacy barriers, the proposed solution may be applicable for the PPE compliance at certain checkpoints where employees can confirm their identity.

Apstrakt: Termin Industrija 4.0 (I4.0) označava transformaciju globalne privrede pod uticajem tehnoloških dostignuća XXI veka. U okviru ove disertacije analizirane su mogućnosti primene tehnoloških pilara I4.0 za rešavanje reprezentativnih problema (model zona) u LEAN industrijskim sistemima. U prvoj studiji predložen je okvir za upravljanje nebezbednim uslovima i nebezbednim postupcima primenom mobilnih i internet tehnologija. Takođe, izvršena je analiza zahteva i predloženo je rešenje prilagođeno za mala i srednja preduzeća. U drugoj studiji slučaja predloženo je rešenje za unapređenje dimenzione kontrole ekstrudovanih gumenih profila. Rešenje je bazirano na upotrebi digitalne kamere i algoritama kompjuterske geometrije i obrade slika – čime je postignuto je poboljšanje performansi operatera od ~6%. U trećoj studiji slučaja predložen je sistem za popis materijala koji se transportuju na paletama u štamparskoj industriji. Digitalizovani popis pokazao se kao neuporedivo brži i pouzdaniji od trenutnog-ručnog popisa skladišta. U četvrtoj studiji slučaja predloženo je rešenje za analizu neergonomskog transporta tereta ručnim kolicima. Upotrebom IoT modula detektuju se povećane vrednosti sile kojom radnik deluje, što omogućava analizu odgovarajućih frejmova snimljenih internet kamerom. Identifikacijom kritičnih momenata, u kojima algoritmi dubokog učenja omogućavaju trodimenzionalni prikaz položaja tela operatera, omogućena je preciznija analiza ergonomskih rizika. Validacija predloženih algoritama i rešenja ukazuje na veliki potencijal za njihovu širu praktičnu primenu. Zaključeno je da postoji veliki broj specifičnih problema u industriji koji zahtevaju namenska rešenja – što otvara velike mogućnosti za dalja istraživanja i razvoj digitalizovanih rešenja baziranih na pilarima I4.0., The term Industry 4.0 (I4.0) refers to the transformation of the global economy under the technological advances of XXI century. The aim of this dissertation was to assess possibilities for applying I4.0 technological pillars for solving various representative problems (model zones) of LEAN industrial systems. In the first case study, a framework for managing unsafe conditions and unsafe acts by using mobile and internet technologies was proposed. Moreover, an analysis of requirements and аdaptation of the proposed solution for use in small and medium enterprises was performed. In the second case study, a solution for the dimensional inspection of extruded rubber profiles was proposed. The solution is based on the use of a digital camera and algorithms for computational geometry and image processing – which improved operators’ performances by ~ 6%. In the third case study, the system for the inventory of materials that are transported on pallets in the printing industry was proposed. The digitized inventory has shown as significantly faster and robust compared to the currentmanual inventory practice. In the fourth case study, a solution for the detection and analysis of non-ergonomic cargo transport by handcart was proposed. By using an IoT module, increased values of the force exerted by the worker were detected, which enabled the analysis of corresponding frames acquired by the IP camera. Identification of critical moments, by using Deep learning algorithms for three-dimensional assessment of the operator's body pose, enables a more precise analysis of ergonomic risks. Validation of developed algorithms and solutions indicated a high potential for wider applications. It was concluded that there are a number of specific problems in the industry that require dedicated solutions - which opens up opportunities for further research and development of digitalized solutions based on the I4.0 pillars.

Workplace safety is a scientific discipline that has been constantly evolving along with industrial development. Nowadays technological progress of tools and materials used in industry, in addition to all the positive impacts, increase the probability of injuries of the operators that use them. Consequently, there are industry standards and recommendations that specify appropriate personal protective equipment (PPE) for certain workplaces. Although every company is able to provide protective equipment for its employees, the major challenge is the compliance and control of their proper use. The aim of this study was to assess the possibility of applying artificial intelligence and deep learning techniques for automated PPE compliance, which could help in taking preventive action with the aim of reducing injuries caused due to non-use or misuse of prescribed PPEs. The obtained results showed that the YOLOv5 algorithm achieved high precision (average 0.857) for the detection of various types of head-mounted personal protective equipment. Accordingly, there is a high potential for future use of such tools in improving workplace safety and PPE compliance. Potential users of the application based on this recognition algorithm would be companies which regulations define the type of PPEs that have to be used at a certain working position.

Koronarna angiografija bazirana na iks-zracima i intravaskularni ultrazvuk (engl. IVUS) su, i pored invazivne i dvodimenzionalne prirode, najčešće korišćene procedure za dijagnostikovanje i tretiranje obolelih koronarnih arterija (KA). Ova disertacija fokusirana je na razvoj algoritama za trodimenzionalnu rekonstrukciju KA iz nekalibrisanih angiografskih projekcija. Predložen je prošireni matematički model formiranja angiografskih slika, koji je zajedno sa robusnim Genetskim algoritmom, korišćen za optimizaciju-kalibraciju parametara koji opisuju orijentaciju uređaja tokom akvizicije. Predložena je metoda za uparivanje nesinhronizovanih frejmova bazirana na parcijalnom preklapanju signala. Njihovom integracijom, rekonstruisan je lumen iz angiografskih projekcija. Takođe, predložena je procedura za diskretizaciju kompleksnih koronarnih stabala bazirana na jednostavnijem nezavisnom procesiranju povezanih koronarnih grana. Fuzijom metoda za rekonstrukciju centralne linije KA sa metodama za segmentaciju IVUS slika rekonstruisan je lumen i zid KA, koji su korišćeni za numeričko modeliranje progresije plaka (restenoze). Metodom konačnih elemenata, progresija plaka je simulirana sprezanjem Navije-Stoksovih jednačina protoka sa difuznom jednačinom transporta. Metodom parcijalnog preklapanja, automatizovano je uparivanje bazalnih i IVUS snimaka snimljenih nakon datog perioda praćenja., Despite their invasiveness and two-dimensional nature, X-ray angiography (XRA) and intravascular ultrasound (IVUS) still represent the gold standard for diagnosis and treatment of coronary artery (CA) disease. Since safety and performance standards have become more stringent, needs for robust computer methods for 3D reconstruction, modeling and visualization of CAs are constantly increasing. This thesis proposed algorithms for three-dimensional modelling of CA from uncalibrated XRA projections. A comprehensive mathematical model of XRA image formation was developed and used in conjunction with a robust genetic algorithm optimizer to determine the calibration parameters across X-ray views. Furthermore, an approach for finding frames correspondences between nongated XRA was developed using a partial-matching approach. By integrating the developed procedures, the vascular lumen was reconstructed from its XRA projections. Moreover, a nonuniform rational B-spline-based method that simplifies the problem of meshing of complex CA trees to the simpler task of independent meshing of connected branches was introduced. By fusing the reconstructed centerline with segmented IVUS images CA reconstructions were obtained and used for further numerical simulation of plaqure progression (in-stent restenosis). The finite element simulation of plaque progression was performed by coupling, the flow equations with the transport equation and applying realistic boundary conditions at the wall. By using a partial-matching approach, the alignment of baseline and follow-up data was performed automatically.