- 學位論文
以卷積神經網絡識別遠距與高俯角的工地行人與測量標記
Recognition of Long-distance and Large Depression-angle Worker and Survey Markers on Construction Sites with Convolutional Neural Networks
摘要
目前以深度學習進行工地工人或個人安全裝備的研究已經很多。但由於工地的攝影機經常以較大的俯角以及距離監控工地，這方面的研究上較少被定量考慮。因此本次研究的主題為「以卷積神經網絡識別遠距與高俯角的行人」，主要在探討工地行人拍攝的距離與俯角與辨識準確度的定量關係，作為設置工地的攝影機位置與數量的參考。本研究使用Yolov4做為識別工地行人的工具。根據行人高佔照片較小的邊的比例分成大型(1/5以上)、中大型(1/5~1/10)、中型(1/10~1/15)、中小型(1/15~1/20)、小型(1/20以下)。根據行人高寬比來判斷角度(假設行人是站立)高/寬比例越小俯角越大，高/寬比例越大俯角越小(接近水平)，分成水平(3.5以上)、小俯角(2.5~3.5)、中俯角(2~2.5)、大俯角(1.5~2)、很大俯角(1.5以下)。研究結果顯示，以角度來說，水平拍攝偵出率最高達87.38%，隨著俯角增加，偵出率呈非線性下降，到很大俯角剩64.71%；在距離的不同下，大型與中大型還有中型行人的偵出率最高，達80%以上，隨著目標物距離增加，逐漸遞減，至小型行人剩57.14%，故不同的俯角與距離會影響識別的成功率。 此外，利用深度學習實現已知點標記的自動識別和二維定位，研究結果顯示，單標示單分類的識別能力極佳，二維定位誤差為2.65個像素，結果表明深度學習可以精確識別與定位標記。
並列摘要
There have been many studies on the use of deep learning for identifying site workers or personal safety equipment. Construction site cameras often monitor construction sites at large depression angles and distances; however, there are few studies in the quantitative relationship between them. Therefore, the theme of this study is to investigate the quantitative relationship between the distance and depression angle of the accuracy of recognition of the construction site pedestrians. This study uses Yolov4 as a tool to identify pedestrians at construction sites. According to the ratio of pedestrian height to the smaller side of the photo, the images are divided into five levels: very large (1/5 or more), large (1/5~1/10), medium (1/10~1/15), small (1/15~ 1/20), very small size (less than 1/20). Assuming that the pedestrian is standing, according to the pedestrian aspect ratio, the smaller the height/width ratio, the greater the depression angle, and the larger the height/width ratio, the smaller the depression angle. Hence, the images are divided into five levels: horizontal (above 3.5), small depression angle (2.5~ 3.5), medium depression angle (2~2.5), large depression angle (1.5~2), very large depression angle (below 1.5). The results show that the detection rate of horizontal shooting is as high as 87.38%. As the depression angle increases, the detection rate decreases non-linearly, reaching 64.71% at a very large depression angle. The results show that the detection rate of very large, large and medium pedestrians is the highest, reaching more than 80%. As the distance of the target increases, it gradually decreases to 57.14% for very small pedestrians. In addition, the results of automatic recognition and 2D localization of surveying markers show that the recognition ability is excellent and 2D localization average error is 2.65 pixels. The results showed that the deep learning can accurately identify and localize surveying markers.