土石方收容處理場上物件特徵與擬製指向性辨識之研究

NCUIR > Engineering College > Graduate Institute of construction engineering > Electronic Thesis & Dissertation > Item 987654321/95952

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/95952

Title:	土石方收容處理場上物件特徵與擬製指向性辨識之研究
Authors:	沈廉;Shen, Lian
Contributors:	土木系營建管理碩士班
Keywords:	土方管理;無人飛行載具(UAV);物件偵測;YOLOv8演算法;指向性辨識;Earthwork Management;Unmanned Aerial Vehicles(UAV);Object Detection;YOLOv8 algorithm;Directional Recognition
Date:	2024-07-12
Issue Date:	2024-10-09 17:26:14 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著社會經濟蓬勃成長與都市化建設發展，重大公共工程與民間營造工程每年產生大量之營建副產物，剩餘土石方循環利用與有效管理逐漸受到主管機關之重視，有效之監控與管理能夠減少違規傾倒與環境破壞的發生。一般仰賴人力於各個土資場出入口進行營建剩餘土石方監控，輔以政府提供之電子系統供查核。無人飛行載具可於空中一次性獲取大量影像資料，在測量與監控領域已行之有年，搭配演算法等電腦深度學習，有助於辨識及監控土石方資源堆置處理場內物件進出情形。本研究使用無人飛行載具獲取土資場相關機具之影像資料，搭配YOLOv8演算法，進行模型訓練與物件偵測，透過模型中的指向性辨識，進行擬製指向性辨識，分析場景中物件之移動方向，以追蹤與監控場中可疑之物件。研究中使用162張影像進行模型訓練與測試，平均偵測率為0.806，符合預期辨識精度，透過擬製指向性辨識可追蹤物件移動方向。此研究成果結合科技與深度學習，將有助於相關業者提高場中之監控效率，同時降低人力管理與時間之負擔。 ;With the vigorous growth of the economy and urbanization, significant public works and private construction projects generate a large amount of construction by-products every year. The recycling and effective management of surplus earthwork materials have gradually attracted attention from regulatory authorities. Effective monitoring and management can reduce illegal dumping and environmental damage. Traditionally, manpower is relied upon to monitor the entrances and exits of various soil resource storage sites, supplemented by government-provided electronic systems for inspection. Unmanned aerial vehicles can obtain large amounts of image data from the air in one go, a practice long established in the fields of surveying and monitoring. Coupled with algorithms such as computer deep learning, they can assist in identifying and monitoring the inflow and outflow of objects in soil resource storage sites. This study utilizes UAVs to acquire image data of relevant equipment in soil yards and employs the YOLOv8 algorithm for model training and object detection. Through directional recognition within the model, simulated directional recognition is conducted to analyze the movement direction of objects in the scene, enabling the tracking and monitoring of suspicious objects in the area. A total of 162 images were used for model training and testing, with an average detection rate of 0.806, meeting the expected recognition accuracy. Through simulated directional recognition, object movement directions can be tracked. This research combines technology and deep learning, which will help relevant stakeholders improve monitoring efficiency on-site while reducing the burden of manpower and time management.
Appears in Collections:	[Graduate Institute of construction engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	5	View/Open

社群 sharing

Loading...