000			03358nam a22002657a 4500
003			KOHA
005			20250109090220.0
008			240927d2024 cy o|||| |||| 00| 0 eng d
040			_aCY-NiCIU _beng _cCY-NiCIU _erda
041			_aeng
090			_aYL 3467 _bP48 2024
100	1		_aPeter, Erikpara Gbesimi
245	1	0	_aPOTHOLE DETECTION SYSTEM USING MACHINE LEARNING ON MOBILE DEVICE / _cERIKPARA GBESIMI PETER ; SUPERVISOR, PROF. DR. ERBUĞ ÇELEBİ
264			_c2024
300			_a83 sheets ; _c30 cm _e+1 CD ROM
336			_2rdacontent _atext _btxt
337			_2rdamedia _aunmediated _bn
338			_2rdacarrier _avolume _bnc
502			_aThesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Computer Engineering
520			_aAs vehicular traffic continues to increase globally, the condition of road surfaces has become a significant concern for both infrastructure integrity and public safety. Traditional road maintenance methods, often reliant on manual inspections, are labor-intensive and insufficiently comprehensive. This research addresses the urgent need for an effective solution to monitor and assess road damage through advanced technological methods. By focusing on the problem of detecting road defects like potholes and cracks, the study aims to develop a more efficient and accurate system for road condition assessment. To tackle this issue, the research employs a machine learning-based approach, specifically leveraging the YOLOv8n (You Only Look Once version 8n) object detection model for automatic road defect detection. The methodology includes collecting high-resolution images of road surfaces, preprocessing these images for optimal input, and training a Convolutional Neural Network (CNN) to identify and categorize various types of road damage. Additionally, the study explores the implementation of this model on mobile devices to assess its feasibility for real-time, on-site road condition evaluations. The experimental results reveal that the YOLOv8n model achieves a precision of 0.625 and a recall of 0.582, with an mAP50 score of 0.598 and an mAP50-95 score of 0.321. Performance evaluations indicate that the model can process images with an average inference time of 3.2 milliseconds and a total image processing time of approximately 7.4 milliseconds. Furthermore, the research explores the impact of varying the number of threads used in the mobile app, demonstrating that increasing the number of threads from 2 to 4 significantly reduces inference time, thus enhancing the efficiency of the detection system. In conclusion, this study successfully demonstrates the potential of machine learning techniques for road damage detection and the feasibility of implementing these methods on mobile devices for real-time applications. The findings highlight the effectiveness of YOLOv8n in detecting road defects and provide insights into optimizing system performance through threading strategies. Future work could explore further improvements in model accuracy, scalability for diverse road iii conditions, and integration with automated maintenance systems to better support road management efforts.
650		0	_aComputer Engineering _vDissertations, Academic
700	1		_aÇelebi, Erbuğ _esupervısor
942			_2ddc _cTS
999			_c292961 _d292961