A Two-Stage Real-Time Gesture Recognition Framework for UAV Control

Unmanned aerial vehicle (UAV) has been widely used in various fields. Traditional UAV controllers require much experience, whereas the control method based on gesture recognition has the advantages of simplicity and flexibility. However, gestures are often simply recognized by current deep learning algorithms, and the static and dynamic properties of gestures are liable to be overlooked, which affects the efficiency of gesture recognition. Hence, a two-stage real-time gesture recognition framework based on the differentiation between static gestures and dynamic gestures is proposed, and gestures toward real-world UAV control are accurately recognized in real time. Besides, a fast correlation-based filter (FCBF) is used to acquire the optimal features. Fifteen gestures, including three static gestures and 12 dynamic gestures, are defined to evaluate the performance of our framework. A practical data glove is meticulously designed with multiple inertial measurement units (IMUs) to obtain the gesture data. Experimental results show that the two-stage framework with FCBF achieves an accuracy of 98.27% under cross-subject cross-validation, outperforming other methods. This work proves the feasibility of optimizing the gesture recognition method by studying the static and dynamic properties of gestures, expecting to facilitate the development of human-computer interaction.