Detecting Personal Protective Equipment using Convolutional Neural Network (Research Case: Company X)

Muhammad, Andra Fahreza (2024) Detecting Personal Protective Equipment using Convolutional Neural Network (Research Case: Company X). Diploma thesis, Universitas Andalas.

	Text (Cover dan Abstrak) Cover-Abstrak.pdf - Published Version Download (197kB)
	Text (Bab 1 Pendahuluan) Bab 1.pdf - Published Version Download (155kB)
	Text (Bab 6 Penutup) Bab 6.pdf - Published Version Download (36kB)
	Text (Daftar Pustaka) Daftar Pustaka.pdf - Published Version Download (173kB)
	Text (Skripsi Full Text) Full.pdf - Published Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Telecommunications industry is one of the fastest-growing sectors in the world, including Indonesia with its market value reach as high as US$14,22 billion as of 2023. Significant growth of the industry is due to increasingly eased bureaucracy, increasing competition and investment on R&D. To accommodate the immense growth, companies started providing solutions which requires maintaining infrastructure with skilled technicians, this however, comes with risk of accidents. In Indonesia alone, there has been 265.334 cases recorded in 2022, increased significantly around 13.26% from previous year, this leads to productivity loss, extra cost, and mental/physical pain towards employees. Company X as one of the major telecommunication enterprises in the region also experienced cases of workplace accidents caused by the lack of compliance in enforcing personal protective equipment (PPE). Thus, an urgency and need to conduct research and developing a solution to PPE compliance is present, through various possible alternatives. This includes the utilization computer vision to detect personal protective equipment on technicians, out of several frameworks/algorithms to develop object detection model, CNN, specifically YOLO method is used due to its balance between speed and accuracy while also being lightweight and scalable. PPE detection model development is initiated by collecting 4.611 images available for public use depicting PPE usage from different sources into a dataset and mapping them into 7 classes: helmet, goggles, mask, vest, gloves, boots, and person. Next, dataset is pre-processed by transformation (re-sizing, re-orienting, and re-mapping), followed by augmentation (applying distortion effects to images). This outputs a total of 8.046 images that will then be used for model training-validation-testing process lasting for 20 epochs using NVIDIA T4 GPU computing engine provided by Google Colab. This process outputs model weight with mAP50 score of 0,881 and mAP50-95 of 0,618, with inference time of 4,6ms per image. Furthermore, integrating and retrofitting detection model to existing working procedure allows the automation of job report reviewing process which enables positive feedback loop in the system, this includes increased safety and supervision for technicians, decreased workload for supervisors by introducing automation to handle repetitive tasks which reduces the possibility of error and negligence and increasing supervisors’ work efficiency by 9,38%, other than that, model utilization enables reinforced learning for the model through active feedback from users, which in turn increases its performance.

Actions (login required)

View Item