- 學位論文
針對工業安全於人臉及全身姿態辨識之異常事件檢測系統
An Abnormal Event Detection System with Human Face and Full-Body Posture Recognition for Industrial Safety
摘要
本論文的主要目標是開發一個適用於工廠場景的異常事件之檢測系統，並在偵測到異常事件時向監控端發送警報。首先，本系統使用Lightweight Openpose演算法捕獲人體骨架關鍵點，然後利用提出的多層感知器神經網絡來識別各種人體姿勢，包括跌倒、蹲、跪、站和坐。透過該系統，本文提出的輕量級架構獲得了與常見的卷積神經網絡相當的識別準確率和較低的計算需求，並進行了不同標準的評估測試。隨後，提出的系統還整合了人臉識別功能，使系統不僅能夠檢測異常的人體姿勢，還能夠檢測非法人員進入工廠場地。檢測到異常事件時，即時向監控室發送警報。實驗結果證實了提出的系統在姿勢識別方面能夠取得良好的準確率，証完其輕量級架構在即時影像辨識中的可行性，以達到進行遠端的異常事件檢測。
並列摘要
The primary objective of this study is to develop an abnormal event detection system suitable for factory scenarios, capable of sending alerts to the monitoring end upon detecting abnormal events. First, the Lightweight OpenPose algorithm is employed to capture human skeletal key points, followed by the use of a proposed multilayer perceptron (MLP) neural network to recognize various human postures, including falling, squatting, kneeling, standing, and sitting. Through this system, the proposed lightweight architecture achieves recognition accuracy comparable to common convolutional neural networks (CNNs) with lower computational requirements. Various evaluation tests were conducted under different criteria. Furthermore, the proposed system integrates facial recognition functionality, enabling it not only detect abnormal human postures but also monitor the unauthorized person in the factory. When an abnormal event is detected, an alert is promptly sent to the backend. Experimental results confirm that the proposed system achieves satisfactory accuracy in posture recognition. Its lightweight architecture proves feasible for real-time image recognition and remote abnormal event detection.