應用於以感測器為基礎的連續手勢辨認之長短程記憶體演算法則及FPGA硬體實現之研究

This is a two-year project. The goal of this project is to develop Long Short Term Memory (LSTM) algorithms and architectures for sensor-based continuous hand gesture recognition. In this work, the sensors such as accelerometers and gyroscopes commonly deployed in smart phones and smart bracelets are adopted. This may extend the accessibility of the proposed algorithms and architectures to large varieties of devices for gesture recognition applications. After the first year of this project, we have completed the development and the evaluation of the LSTM algorithm. The hardware architecture for the LSTM algorithm by Field Programmable Gate Array (FPGA) has been implemented at the second year of the project. Furthermore, an evaluation system for the real-time continuous hand gesture recognition has been built. Experimental results reveal that the proposed hardware architecture has utilized only limited hardware resources. In addition, for the systems requiring the classification of 5 classes of gestures, the average hit rate for each class of gestures is above 97%.