| 摘要: | 人體姿態識別在近年來廣泛的應用在醫療、運動、嬰兒老人監控以及犯罪監視等 等，而在目前人體姿態識別上較多人使用的為 Openpose，因為它有著簡易的特性， 不需要高階攝像頭，以普通 2D 的 RGB 圖像即可達成關節點估計，不僅僅可以偵測身 體支點，同時也可以偵測手部以及臉部關節點。本研究將此運用於車手和非車手的 ATM 提領影片中，再透過和專業警員的訪談，得出車手在 ATM 提領時的動作特徵， 分別為插卡、看手機、左顧右盼、取錢和清點等等，然而將影片所得出之座標點轉換 為特徵值，以特徵值的方式去描述 ATM 提領的動作特徵，又因所述動作特徵在影片裡 佔比偏低，資料有不平衡情況，所以本研究透過 n gram 以及 undersampling 的方式， 利用深度學習模型雙向 LSTM(Long Short-term memory 長短期記憶模型)進行判 斷，以追求較高的精確率和召回率。;In recent years, human activity recognition(HAR) has been widely used in health-care, sports, baby and elderly monitoring and crime surveillance, etc. At present, Openpose is used by most people in HAR, because it has simple characteristics and does not A high-level camera is needed, and the joint point estimation can be achieved with ordinary 2D RGB images. It can not only detect the body joint, but also the hand and face joint points. Our study applies Openpose to the ATM withdrawal videos of moneymule and non-moneymule, and then through interviews with professional police officers, we can get the characteristics of the driver′s actions when withdrawing from the ATM. It including inserting a card, looking at the phone, looking around, and withdrawing money,and counting, etc. However, we convert the coordinate points obtained from the film into feature, and describe the action characteristics of ATM withdrawal in the form of feature. Also, because the action features occupy a low proportion in the video, the data is imbalance, so we use the deep learning model, bi-LSTM (Long Short-term memory model) to make evaluate in the way of n gram and undersampling, in order to pursue a higher precision and recall rate. |
