| Summary: | 碩士 === 國立虎尾科技大學 === 電機工程研究所 === 103 === In recent years, the safe identification system used in intelligent environment has been attractive by people and more and more similarly systems were proposed. This paper presented a user identification based on posture and combined the skeleton data which gets from KINECT. It contains two types of features, including non-learning features and learning features of the learning methods. Based on human skeleton joints, there are three user features proposed by the author. The methods in sequence are “Adjacency Joint Distance”, “Confirm Skeleton Angle” and the last one is to combine of the above, and two learning features, “Gravity of Offset” (GLO), “Transfer Matrix of Offset” (TMLO). All of them were used in user identification system as the features.
The paper are also using Support Vector Machine (SVM), Gaussian Mixture Model (GMM) and Principal Component Analysis (PCA) to develop user legality confirmed in SVM and develop user identity recognition in GMM and PCA. And propose three types of user recognition user recognition model, GMM-PCA、PCA learning-SVM、PCA learning-GMM, which trying to modify the original method of single model.
Three types of non-learning features are separately training in SVM、GMM、PCA. And prefer to select features in better recognition rates. SVM and PCA we select “Adjacency Joint Distance” and GMM select combine features. Non-learning features was trained in GMM-PCA, total of score normalization GMM-PCA was decided to recognition result.
The identification of action may change through the time and the habits of user, so it will affect the efficiency of each recognition process. To improve the situation, we add the learning method of machine and developed two learning algorithms. The paper is using Adjacency Joint Distance to train in PCA, and according to PCA learning methods to propose two types of learning offsets features, PCA-GLO and PCA-TMLO are training in SVM and GMM. PCA-GLO is learning 16 times training in SVM, the recognition rates was 94.3%. PCA-GLO is learning 16 times training in GMM, the recognition rates was 99.8%. And PCA-TMLO is learning 10 times training in SVM, the recognition rates was 98.9%. The experiment result, PCA-TMLO training in SVM was better than single SVM by more learning times, and learning times was better than PCA-GLO training in SVM. PCA-GLO training in GMM which recognition rates was better than single GMM. The experiment result, it proved the learning effect which the features was extracted in PCA learning methods.
|
|---|
