| Summary: | 碩士 === 國立清華大學 === 電機工程學系所 === 105 === Gesture recognition is a wide topic in computer science and language technology with the goal of interpreting human gestures via mathematical algorithms.
In this thesis, we we have recorded signals of ten kinds of hand movements into the computer using a wearable inertial measurement unit (IMU) wireless device with six axes data (including the accelerometer and gyroscope). The sensor is worn on the wrist, and the raw data are transmitted to the computer via Bluetooth Low Energy (BLE) to verify the captured data, a recognition system with machine learning classification process is built. Our movements can be divided into two categories, with the first being single gestures, which includes ten basic movements, and the second the continuous combinational gestures, which is com-
posed of the previous ten basic movements through different combinations.
In order to achieve higher recognition accuracy, we used machine learning process in the system and two analyses, principal component analysis (PCA) and linear discriminant analysis (LDA), to extract well distinguished features. The main advantage of PCA and LDA is reducing dimensions of data while preserving as much of the class discriminatory information as possible. In addition, later processing time can be decreased due to reduced dimensions of data. The experiment is then proceeded with support vector machine (SVM) and dynamic time warping (DTW). With SVM technique, we can recognize movement with higher accuracy and less computation time. High dimension data are also supported. Even non-linear relations can be modeled with more precise classication due to SVM kernels. Dynamic time warping increases recognition accuracy by categorizing movements through the measurement
of the resemblance among several temporal sequences which may alter in speed. In our experiment, we can get the accuracy of recognition at 100% for 10 classes with 40 subjects in
single gesture under the case of user-dependent, and for the user-independent case, the recognition rate is 90%. And in continuous combinational gesture for the user-independent case, we can get the accuracy of recognition at 86.99% in fixed combinational gesture, and 60% in arbitrary combinational gesture.
We have also overcame one of restrictions of the support vector machine, instead of running the algorithm off-line after all the data are measured, the algorithm can be held during the process of measurement, which greatly shortened the predict time from 2.118 seconds to 0.195 seconds, enhancing the efficiency of the application.
|
|---|
