IMPLEMENTATION OF YOLOV8-BASED OBJECT DETECTION ALGORITHM IN MINING AREA

• Main Author: Fadillah Rafi, Farhan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85229
• Institution: Institut Teknologi Bandung

The field of occupational safety and health (OSH) plays a pivotal role in the context of industrial operations, particularly within the mining sector, which is characterised by a heightened risk of occupational accidents and health hazards. The objective of OHS supervision is to prevent workplace accidents and to safeguard the health of workers, who are a vital resource for the continued operation of the company. In mining areas, the necessity for close and continuous supervision is paramount in order to guarantee that all activities are conducted in accordance with established safety standards, given the inherently complex and high-risk nature of the work environment. PT X, a mining company in Indonesia, employs the use of the Mining Eyes Analytics system, which is based on the use of CCTV cameras and the YOLOv4 algorithm, for the purpose of detecting any deviations or unsafe conditions that may arise within the mining area. The system is designed to detect the presence of individuals, heavy-duty trucks (HD), and light vehicles (LV). However, this surveillance system still exhibits limitations in terms of detecting small objects, objects in poor lighting conditions, and the occurrence of detection errors. This research employs YOLOv8-based object detection to address these shortcomings. The YOLO models utilised are YOLOv8, a modified version of YOLOv8 at the head to enhance multi-scale detection capability, and BGF-YOLO, which incorporates multi-scale detection optimisation and the application of an attention mechanism in the feature fusion process to improve model focus. The data employed in this study comprises daytime mine area data encompassing a range of mine environmental conditions, including poor lighting and the presence of overlapping objects. The total number of data points utilized is 11,000. The results of the evaluation demonstrate that the BGF-YOLO model exhibits superior performance compared to other models. The BGF-YOLO model demonstrated a precision of 94.9%, a recall of 91.1%, and a mean average precision (mAP@50) of 86.5%. Although the optimised YOLOv8 model exhibits the highest precision at 96.7%, its overall performance remains inferior to that of BGF- YOLO and YOLOv8. This discrepancy can be attributed to the fact that the introduction of the object head enables precise detection but simultaneously diminishes the model's responsiveness to novel data.