Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74964
Title: | 基於深度學習的電腦視覺技術於即時工地人員裝備違規辨識 Deep Learning Based Computer Vision Techniques for Real-time Identification of Construction Site Personal Equipment Violations |
Authors: | Rong-Chun Chien 簡榮均 |
Advisor: | 謝尚賢(Shang-Hsien Hsieh) |
Keyword: | 工地安全,個人防護裝備,電腦視覺,深度學習,卷積神經網路, Construction sites safety,PPE,Computer vision,Deep learning,CNN, |
Publication Year : | 2019 |
Degree: | 碩士 |
Abstract: | 在工地,確實的配戴個人防護裝備(Personal Protective Equipment)是有效避免意外發生的方法之一。然而工地人員為了作業方便、舒適度等原因,常有未依規定穿著適當裝備的情況發生。因此,如何確保個人防護裝備的配戴一直都是工地安全的重點項目之一。
近年來隨著顯示卡的性能提升及深度學習的崛起,影像辨識領域有著相當大的進展。若以工地監視器的畫面進行影像辨識,比起傳統上以感測器為主的監控方式,能以更低的成本達成即時的個人防護裝備偵測。 本研究提出一個新的二階段偵測方法用以偵測未戴安全帽、未穿反光背心和赤裸上身三項PPE違規。首先,基於RetinaNet的物件偵測模型會先辨識出畫面中人員的位置,再利用基於InceptionNet的分類模型對人物影像進行裝備違規的辨識。相較於直接使用物件偵測方法偵測PPE,二階段的辨識方法可以避免在進行物件偵測時PPE因為縮圖而導致特徵遺失的問題。 本研究搜集了3015張工地影像,利用遷移學習完成模型的訓練,並測試其在不同人物解析度下的表現。結果顯示本研究的方法能在10 fps的運算速度下,有效地偵測出工地的PPE違規,在人物解析度為120像素以上時,Precision和Recall都能在0.9以上,而在人物解析度為60像素時也能有0.8的Precision。 Being well-equipped with Personal Protective Equipment (PPE) plays an essential role in construction sites to protect individuals from accidents. However, due to inconvenience and discomfort, it is common to see workers not wearing them on site. Therefore, ensuring the wearing of PPE remains important subject to facilitate construction sites safety. In recent years, because of the increasing performance of graphic cards and the rise of deep learning, Convolutional neural network (CNN) based computer vision techniques is receiving increased attention. Monitoring PPE use via computer vision method has been considered to be effective rather than sensor-based method. In this paper, a two-stage method is proposed to automatically detect 3 kinds of PPE violations, including non-hardhat-use, un-equipped with safety vest and bare to the waist. Combining object detection model and classification model, our two-stage method can avoid feature loss of small-scale PPE to achieve better accuracy. First, the object detection model based on RetinaNet is adopted to detect the presence of worker in the image. Then, by using InceptionNet as classification model, these worker images are input to identify the violations of PPE use. This study collected 3015 site images to preform transfer learning. The results show that our method can effectively detect PPE violations at the frame rate of 10 fps. When the image resolution of the worker is 120 pixels or more, both precision and recall can be above 0.9; while the resolution is only 60 pixels, it could also achieve the precision of 0.8. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74964 |
DOI: | 10.6342/NTU201900514 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 土木工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-108-1.pdf Restricted Access | 2.67 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.