การตรวจจับการล้มโดยใช้การวิเคราะห์องค์ประกอบหลักและเทคนิคการจัดกลุ่ม

Abstract

This thesis proposes the techniques, methodologies and processes to detect the fall with Kinect camera. The proposed method depends on the shape of human actions. The process is divided into 2 phases, i.e., the recognition phase and the fall detection phase. In the recognition phase, we use 4 fuzzy clustering algorithms, i.e., Fuzzy C-Means Clustering (FCM), Gustafson Kessel Clustering (GK), Gath and Geva Clustering (GG), and Possibilistic C-Means Clustering (PCM) to cluster the action groups. The utilized features are generated from the Hu moment invariants combining with Principal Component Analysis (PCA). Then K-Nearest Neighbor (KNN) and Fuzzy K-Nearest Neighbor (FKNN) with the ellipse approximation of shape orientation and the fuzzy rules are used in assigning an action to each frame. Then the action of several consecutive frames is used to determine whether there is the fall occurring or not. From the experiment, the best recognition method is from GG clustering with 80 prototypes per clusters with 7 Principal Component (7PCs) and 3 FKNN. This method yields 93.80% accuracy. For the fall detection phase with the same scheme, the best result of user dependent and user independent are 92.05% and 92.65% respectively. In addition, there are 3 cameras set up in the experiments. We found that the result of 1 from 3 cameras and 2 from 3 cameras that can detect the fall are around 93.05%-99.03% and 94.32%-98.66% for user dependent and user independent data sets, respectively.