Acceleration Gesture Recognition Based On Random Projection

The natural human-computer interaction technology becomes the hot spot of the current computer technology research. Due to its simplicity and natural properties, gesture as an intuitive way of expression has received the attention of researchers. There is no doubt that treating the gesture as a way of human-computer interaction will greatly enhance the experience of human-computer interaction.The current human-computer interaction research based on hand gestures can be divided into three categories according to different gesture data acquisition method: gesture recognition based on computer vision, gesture recognition based on EMG signal and gesture recognition based on wearable sensors. While gesture recognition method based on computer vision is susceptible to light, shooting angle and background, to overcome the shortage mentioned above, this dissertation employs wearable sensor as data acquisition equipment and then presents a gesture recognition algorithm based on Random Projection, which uses the three dimensional acceleration signal of hand motion collected from eight testers. The experiment results show that the algorithm is effective.The main work of this dissertation includes the following aspects:1) Gesture acceleration data acquisition and processing. In this dissertation the Wiimote is employed to collect gesture acceleration data. The original signal is filtering and smoothing first, and then modified SWAB algorithm is used to automatically extract effective data segment from continuous gesture signal.2) Data clustering. The Dynamic Time Warping algorithm and the Affinity Propagation algorithm are used to create an exemplar for each gesture trace in training set.3) Data dimensional reduction. The Random Projection algorithm is adopted to project the unknown gesture trace and the candidate gesture traces onto a lower dimensional subspace.4) Recognition using l1-minimization. By solving the l1-minimization problem to recognize the unknown gesture trace.5) Verify the effectiveness of the RP-based gesture recognition algorithm. The experiments on 2400 gesture traces achieve an accuracy rate of 98.41% for user-independent recognition, and 96.67% for user-dependent recognition, the results show the effectiveness of the presented algorithm.