基於卷積神經網路及影像強化之工地安全帽偵測系統及應用

Construction site safety helmet detection system and application based on convolutional neural network and image enhancement

廖昱唐(Yu-Tang Liao)

指導教授 ： 陽毅平

國立臺灣大學/工學院/機械工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201200

摘要

在現今科技發達的社會裡，越來越多高樓大廈林立，建築工人的需求越來越多。然而，每年都有許多工地的意外，都跟沒有使用安全帽或安全護具有關。筆者實際對工地作業人員進行問卷調查，發現未配戴工地安全帽的人高達一半。由此可知，除了加強工地作業人員的教育訓練外，有一套系統輔助提醒工地作業人員配戴安全帽是刻不容緩的。故本文提出一套工地安全帽偵測系統，並架設攝影機在工地入口，用於偵測工人是否配戴安全帽進入工地。
由於工地攝影機的架設方向無法事先得知，而攝影機架設朝向東西向或向陽時，將會影響影像辨識模型對於類別預測的準確度。本研究結合了影像強化中的限制對比度自適應直方圖均衡化及直方圖均衡化來強化攝影機的影像，並將強化後的影像及原影像輸入給Yolov4-tiny模型做自定義類別的預測。實際測驗可以得知，透過限制對比度自適應直方圖均衡化及直方圖均衡化的影像給模型預測，比原影像直接給模型預測更準確。
除了影像強化與Yolov4-tiny之外，本文加入了LED警示燈及蜂鳴器來提醒未配戴工地安全帽之工作人員之外，還有透過SQLite資料庫管理系統來記錄進入工地人員及是否配戴安全帽，最後，透過Django架設了網站供擁有權限之管理者可以隨時查看。

關鍵字

工地安全帽 ； 影像辨識 ； 限制對比度自適應直方圖均衡化 ； Yolov4-tiny ； 直方圖均衡化 ； SQLite資料庫管理系統 ； Django

並列摘要

With the advance in technology, buildings and infrastructure increased, as also the demands of construction workers. However, some of the accidents arise from workers who don’t wear safety helmets and gear. The author conducted a questionnaire analysis for construction workers regarding the safety on-site and found out interesting insight - half of the workers don’t wear a helmet on the construction site. Therefore, the author advises we should increase the training with the construction worker, and also a system that reminds them to wear a helmet is needed. In the thesis, the author creates an automatic detection system for workers’ helmets, and also set up cameras in the entry of construction sites, which are used for the detection of workers who wear or don’t wear them.
Since the direction of the camera on the construction site cannot be known in advance, the accuracy of the image recognition model for category prediction will be affected. This study combines the contrast limited adaptive histogram equalization (CLAHE) and histogram equalization (HE) to enhance the image of the camera, which are input to the Yolov4-tiny model for prediction of custom categories. The actual test shows that the model prediction by contrast limited adaptive histogram equalization and histogram equalization is more accurate than the original image directly to the model prediction.
Besides the model mentioned above, this article adds LED warning lights and buzzers to remind workers who don’t wear safety helmets on-site and import the SQLite database management system to record workers who enter the site and whether they wear helmets or not. Finally, the author creates a website using Django for managers to inspect anytime.

並列關鍵字

Yolov4-tiny ； histogram equalization ； contrast limited adaptive histogram equalization ； Construction site safety helmet detection ； Django ； SQLite

參考文獻

[1]勞動檢查統計年報。摘自：https://www.osha.gov.tw/1106/1164/1165/1168/

[2]職業安全衛生法。摘自：https://law.moj.gov.tw/LawClass/LawAll.aspx?pcode=n0060001

[3]營造安全衛生設施標準。摘自：https://law.moj.gov.tw/LawClass/LawAll.aspx?pcode=N0060014

[4] M.-W. Park and I. Brilakis, “Construction worker detection in video frames for initializing vision trackers,” Automation in Construction, vol. 28, pp. 15–25, Dec. 2012.

[5] H. Son, H. Choi, H. Seong, and C. Kim, “Detection of construction workers under varying poses and changing background in image sequences via very deep residual networks,” Automation in Construction, vol. 99, pp. 27–38, Mar. 2019.