工地安全之跌倒偵測－以深度學習為工具

Fall Detection of Construction Site Safety: Using Deep Learning as a Tool

鄭群嚴(Chun-Yen Cheng)

指導教授 ： 詹瀅潔

國立臺灣大學/工學院/土木工程學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202003075

摘要

跌倒是職業災害的主要類型之一，在營造相關的產業中，跌倒佔了所有職業災害人次的15%，且營造業為職業災害高風險產業，對於跌倒需要保持更高的關注。在營造業內，由於環境因素，跌倒造成的傷害通常會較高，傷者也因此無法即時自行尋求救援，最後導致傷者無法及時接受適當的治療，可能留下後遺症甚至是死亡。在營造業實施有效的跌倒偵測系統，將能夠掌握珍貴的治療時間，降低跌倒發生造成的傷害。
近年來由人工智慧衍生出的機器學習與深度學習技術蓬勃發展，深度學習的技術能夠讓電腦自行從資料中學習，產生出對應的最佳成果，而影像辨識即為深度學習的一項應用。本研究透過深度學習技術對於人體進行影像辨識，以各種姿勢之人體影像以及營造業內工作姿勢作為訓練用資料，採用卷積神經網路建立跌倒辨識偵測模型。跌倒辨識偵測模型首先會透過物件偵測與動態偵測演算法識別追蹤影片中的人員，並進一步進行跌倒姿勢的辨識偵測，最後在偵測到跌倒時發出警報。本研究建立之跌倒辨識偵測模型，以影像辨識為基礎，在實驗結果中顯示，此模型對於跌倒姿勢的辨識具有81.30%的準確率，能夠有效的偵測跌倒，間接降低跌倒造成的傷害。

關鍵字

跌倒 ； 職業安全衛生管理 ； 深度學習 ； 卷積神經網路

並列摘要

Occupational slips, trips and falls on the same level (STFL) is one of the main types of occupational disasters. STFL accounted for 15% of all occupational disasters in the construction industry. The construction industry is a high-risk industry for occupational disasters and needs to pay more attention to STFL. Due to environmental factors, injuries caused by STFL in the construction industry are usually higher, so the injured cannot immediately seek rescue on their own. Eventually, the injured person cannot receive treatment in time, which may cause sequelae or even death. Implementing an effective STFL detection system in the construction industry will increase treatment time and reduce injuries caused by STFL.
In recent years, machine learning and deep learning technologies have flourished. Deep learning is that the computer learns from the data by itself and produces the corresponding best results, and image recognition is an application of deep learning. In this study, deep learning is used to recognize the human body image. The human body images in various postures and the working postures in the construction industry are used as training data. The convolutional neural network is used to establish a STFL recognition detection model. The STFL recognition detection model first recognizes the person in the tracking video through object detection and motion detection algorithms, then performs STFL pose recognition detection, and finally alerts when a STFL is detected. The STFL recognition detection model established in this study is based on image recognition. The experimental results show that this model has 81.30% accuracy for STFL posture recognition, which can effectively detect STFL and reduce injuries caused by STFL indirectly.

並列關鍵字

STFL(Occupational slips, trips and falls on the same level) ； Occupational safety and health management ； Deep learning ； Convolutional neural network

參考文獻

[1] “傻傻分不清楚？人工智慧AI 機器學習 ML 深度學習DL,” 22 07 2017. [線上]. Available: https://kknews.cc/zh-tw/tech/zx5ygop.html.

[2] 職業安全衛生法, 2019.

[3] 營造安全衛生設施標準, 2014.

[4] Brandon, “卷積神經網路的運作原理,” 18 08 2016. [線上]. Available: https://brohrer.mcknote.com/zh-Hant/how_machine_learning_works/how_convolutional_neural_networks_work.html.

[5] “python 新手的救星--Anaconda介紹與安裝,” 16 06 2016. [線上]. Available: https://darren1231.pixnet.net/blog/post/328443678-python-%E6%96%B0%E6%89%8B%E7%9A%84%E6%95%91%E6%98%9F--anaconda%E4%BB%8B%E7%B4%B9%E8%88%87%E5%AE%89%E8%A3%9D.